Skip to content

Spot it. Save it. Free, local-first Chrome extension for images, videos and prompts. Explore 2,000+ curated cases and connect your library to AI Agents via MCP. 提示词导演

修复 DeepSeek 批量分析结果丢失与串行执行

MigenAIe9786d769487
diff --git a/ai-model-capabilities.js b/ai-model-capabilities.js
index cf82d8d..a94fa3e 100644
--- a/ai-model-capabilities.js
+++ b/ai-model-capabilities.js
@@ -5,6 +5,20 @@ const GOOGLE_IMAGE_GENERATION_SOURCE = Object.freeze({
reviewedAt: "2026-08-12"
});
+const DEEPSEEK_VISION_SOURCE = Object.freeze({
+ authority: "DeepSeek API Docs",
+ document: "DeepSeek-V4-Flash-Vision-Exp Release",
+ url: "https://api-docs.deepseek.com/news/news260821/",
+ reviewedAt: "2026-08-24"
+});
+
+const DEEPSEEK_CONCURRENCY_SOURCE = Object.freeze({
+ authority: "DeepSeek API Docs",
+ document: "Models & Pricing",
+ url: "https://api-docs.deepseek.com/quick_start/pricing/",
+ reviewedAt: "2026-08-24"
+});
+
const STANDARD_ASPECT_RATIOS = Object.freeze([
"1:1", "2:3", "3:2", "3:4", "4:3", "4:5", "5:4", "9:16", "16:9", "21:9"
]);
@@ -30,6 +44,20 @@ function googleImageModel(value) {
}
export const AI_MODEL_CAPABILITIES = Object.freeze([
+ Object.freeze({
+ id: "deepseek-v4-flash-vision-exp",
+ providerId: "deepseek",
+ protocol: "chat_completions",
+ tasks: Object.freeze(["textTags", "skillExtraction", "creativePlanning", "imageAnalysis"]),
+ inputModalities: Object.freeze(["text", "image"]),
+ outputModalities: Object.freeze(["text"]),
+ supportedParameters: Object.freeze(["response_format"]),
+ supportedResolutions: Object.freeze([]),
+ supportedAspectRatios: Object.freeze([]),
+ referenceImages: Object.freeze({ supported: true, maxItems: 600, source: "declared", observedAt: "" }),
+ concurrencyLimit: Object.freeze({ value: 2500, source: DEEPSEEK_CONCURRENCY_SOURCE }),
+ source: DEEPSEEK_VISION_SOURCE
+ }),
googleImageModel({
id: "gemini-3.1-flash-lite-image",
supportedParameters: ["response_format", "aspect_ratio", "image_size"],
diff --git a/ai-provider-module.js b/ai-provider-module.js
index 0a1c3f3..7ec7595 100644
--- a/ai-provider-module.js
+++ b/ai-provider-module.js
@@ -117,12 +117,19 @@ async function discoverIdentityModels(fetchImpl, profile, cache) {
const inputModalities = normalizeModalities(architecture.input_modalities ?? architecture.inputModalities);
const outputModalities = normalizeModalities(architecture.output_modalities ?? architecture.outputModalities);
const hasDeclaredModalities = inputModalities.length > 0 && outputModalities.length > 0;
+ const id = clean(item.id ?? item.name).replace(/^models\//, "");
+ const official = getAiModelCapability(profile.id, id);
return modelDescriptor(item, {
- confidence: hasDeclaredModalities ? "declared" : "manual_unverified",
- tasks: hasDeclaredModalities ? tasksFromModalities(inputModalities, outputModalities) : [],
- inputModalities,
- outputModalities,
- source: "provider_models"
+ confidence: official || hasDeclaredModalities ? "declared" : "manual_unverified",
+ tasks: official?.tasks ?? (hasDeclaredModalities ? tasksFromModalities(inputModalities, outputModalities) : []),
+ inputModalities: official?.inputModalities ?? inputModalities,
+ outputModalities: official?.outputModalities ?? outputModalities,
+ supportedParameters: official?.supportedParameters,
+ supportedResolutions: official?.supportedResolutions,
+ supportedAspectRatios: official?.supportedAspectRatios,
+ referenceImages: official?.referenceImages,
+ concurrencyLimit: official?.concurrencyLimit,
+ source: official ? "provider_models+official_capabilities" : "provider_models"
});
}),
cache: response.cache,
@@ -282,7 +289,8 @@ function modelDescriptor(item = {}, options = {}) {
pricing: cloneObject(item.pricing ?? item.pricing_skus),
supportedResolutions: options.supportedResolutions ?? stringArray(item.supported_resolutions ?? item.supportedResolutions ?? parameterDescriptors?.resolution?.values),
supportedAspectRatios: options.supportedAspectRatios ?? stringArray(item.supported_aspect_ratios ?? item.supportedAspectRatios ?? parameterDescriptors?.aspect_ratio?.values),
- referenceImages: options.referenceImages ? structuredClone(options.referenceImages) : undefined
+ referenceImages: options.referenceImages ? structuredClone(options.referenceImages) : undefined,
+ concurrencyLimit: options.concurrencyLimit ? structuredClone(options.concurrencyLimit) : undefined
};
return descriptor;
}
diff --git a/ai-provider-registry.js b/ai-provider-registry.js
index 7f94d3c..4762aed 100644
--- a/ai-provider-registry.js
+++ b/ai-provider-registry.js
@@ -19,6 +19,9 @@ const ASSIGNMENT_EVIDENCE = new Set(["declared", "protocol_inferred", "manual_un
const MANUAL_ASSIGNMENT_TASKS = new Set([
"textTags", "skillExtraction", "creativePlanning", "imageAnalysis", "videoAnalysis"
]);
+const TEXT_ANALYSIS_TASKS = new Set(["textTags", "skillExtraction", "creativePlanning"]);
+export const DEFAULT_TEXT_ANALYSIS_CONCURRENCY = 20;
+export const DEFAULT_MEDIA_ANALYSIS_CONCURRENCY = 10;
export function normalizeAiProviderRegistry(value = {}) {
const source = value?.providers && typeof value.providers === "object" ? value.providers : {};
@@ -33,7 +36,7 @@ export function normalizeAiProviderRegistry(value = {}) {
discovery: { ...(preset?.discovery ?? {}), ...(saved.discovery ?? {}) }
});
}
- return { version: 4, providers };
+ return { version: 5, providers };
}
export function normalizeAiTaskAssignments(value = {}, registryValue = {}) {
@@ -48,6 +51,7 @@ export function normalizeAiTaskAssignments(value = {}, registryValue = {}) {
? source.evidence
: inferredEvidence === "manual_unverified" ? inferredEvidence : "";
if (evidence) assignment.evidence = evidence;
+ assignment.concurrency = normalizeTaskConcurrency(task.id, source.concurrency, discovered?.concurrencyLimit?.value);
return [task.id, assignment];
}));
}
@@ -237,10 +241,34 @@ function normalizeDiscoveredModels(values) {
supportedAspectRatios: cleanArray(item?.supportedAspectRatios),
referenceImages: normalizeReferenceImages(item?.referenceImages),
contextLength: Number(item?.contextLength) > 0 ? Number(item.contextLength) : null,
- pricing: item?.pricing && typeof item.pricing === "object" ? structuredClone(item.pricing) : null
+ pricing: item?.pricing && typeof item.pricing === "object" ? structuredClone(item.pricing) : null,
+ concurrencyLimit: normalizeConcurrencyLimit(item?.concurrencyLimit)
})).filter((item) => item.id);
}
+function normalizeConcurrencyLimit(value) {
+ const limit = Number(value?.value);
+ if (!Number.isInteger(limit) || limit < 2) return null;
+ const source = value?.source && typeof value.source === "object" ? {
+ authority: clean(value.source.authority),
+ document: clean(value.source.document),
+ url: clean(value.source.url),
+ reviewedAt: clean(value.source.reviewedAt)
+ } : null;
+ return { value: limit, source };
+}
+
+function normalizeTaskConcurrency(taskId, value, documentedLimit) {
+ const fallback = TEXT_ANALYSIS_TASKS.has(taskId)
+ ? DEFAULT_TEXT_ANALYSIS_CONCURRENCY
+ : DEFAULT_MEDIA_ANALYSIS_CONCURRENCY;
+ const requested = Number(value);
+ const normalized = Number.isInteger(requested) && requested >= 2 ? requested : fallback;
+ return Math.min(normalized, Number.isInteger(documentedLimit) && documentedLimit >= 2
+ ? documentedLimit
+ : Number.POSITIVE_INFINITY);
+}
+
function normalizeReferenceImages(value) {
const source = value && typeof value === "object" ? value : {};
const maxItems = source.maxItems === null || source.maxItems === undefined || source.maxItems === ""
diff --git a/analysis-batch.js b/analysis-batch.js
index ccf68c7..0797ea1 100644
--- a/analysis-batch.js
+++ b/analysis-batch.js
@@ -8,9 +8,11 @@ import {
} from "./analysis-revision.js";
import { prepareFacetRebuild, validateAnalysisTagResponse } from "./tag-taxonomy.js";
import { entryMediaAssets, primaryImageAsset } from "./media.js";
+import { ANALYSIS_RETRY_POLICY } from "./analysis-retry-policy.js";
export const ANALYSIS_BATCH_VERSION = 2;
-export const ANALYSIS_BATCH_CONCURRENCY = 3;
+export const ANALYSIS_BATCH_CONCURRENCY = 20;
+export const VISION_BATCH_CONCURRENCY = 10;
export async function previewAnalysisBatch(entries = [], options = {}) {
const eligible = [];
@@ -85,9 +87,14 @@ export async function createAnalysisBatchJob(entries, options = {}) {
createdAt: now,
updatedAt: now,
analysisModel: preview.analysisModel,
+ providerId: String(options.providerId ?? "").trim(),
+ model: preview.analysisModel,
+ outputProtocol: String(options.outputProtocol ?? "json_object").trim(),
+ retryPolicy: structuredClone(ANALYSIS_RETRY_POLICY),
outputLocale: options.outputLocale === "en" ? "en" : "zh-CN",
promptVersion: ANALYSIS_PROMPT_VERSION,
profileFingerprint: preview.profileFingerprint,
+ concurrency: normalizeConcurrency(options.concurrency, ANALYSIS_BATCH_CONCURRENCY),
catalogRevision: Number(options.catalogRevision) || 0,
resultCatalogRevision: null,
totalCharacters: preview.totalCharacters,
@@ -158,9 +165,13 @@ export function createVisionBatchJob(entries = [], options = {}) {
updatedAt: now,
outputLocale: options.outputLocale === "en" ? "en" : "zh-CN",
providerType: preview.providerType,
+ providerId: String(options.providerId ?? "").trim(),
model: preview.model,
+ outputProtocol: String(options.outputProtocol ?? "json_object").trim(),
+ retryPolicy: structuredClone(ANALYSIS_RETRY_POLICY),
includeAllImages: preview.includeAllImages,
reanalyze: preview.reanalyze,
+ concurrency: normalizeConcurrency(options.concurrency, VISION_BATCH_CONCURRENCY),
requestCount: preview.requestCount,
skippedAnalyzedCount: preview.skippedAnalyzedCount,
items: preview.items.map((item) => ({
@@ -179,8 +190,8 @@ export function createVisionBatchJob(entries = [], options = {}) {
export function normalizeAnalysisBatchJob(value) {
if (!value || ![1, ANALYSIS_BATCH_VERSION].includes(value.version) || !value.id || !Array.isArray(value.items)) return null;
- const statuses = new Set(["pending", "running", "succeeded", "failed"]);
- const jobStatuses = new Set(["running", "paused", "completed", "canceled"]);
+ const statuses = new Set(["pending", "running", "succeeded", "partial", "failed"]);
+ const jobStatuses = new Set(["running", "paused", "completed", "partial", "failed", "canceled"]);
return {
version: ANALYSIS_BATCH_VERSION,
kind: value.kind === "vision" ? "vision" : "text_tags",
@@ -190,11 +201,15 @@ export function normalizeAnalysisBatchJob(value) {
createdAt: String(value.createdAt ?? ""),
updatedAt: String(value.updatedAt ?? ""),
analysisModel: String(value.analysisModel ?? "") || DEFAULT_ANALYSIS_MODEL,
+ providerId: String(value.providerId ?? ""),
+ outputProtocol: String(value.outputProtocol ?? "json_object"),
+ retryPolicy: normalizeRetryPolicy(value.retryPolicy),
outputLocale: value.outputLocale === "en" ? "en" : "zh-CN",
providerType: value.providerType === "compatible" ? "compatible" : "openai",
model: String(value.model ?? ""),
includeAllImages: value.includeAllImages === true,
reanalyze: value.reanalyze === true,
+ concurrency: normalizeConcurrency(value.concurrency, value.kind === "vision" ? VISION_BATCH_CONCURRENCY : ANALYSIS_BATCH_CONCURRENCY),
requestCount: Math.max(0, Number(value.requestCount) || value.items.length),
skippedAnalyzedCount: Math.max(0, Number(value.skippedAnalyzedCount) || 0),
promptVersion: Number(value.promptVersion) || ANALYSIS_PROMPT_VERSION,
@@ -214,7 +229,10 @@ export function normalizeAnalysisBatchJob(value) {
attempts: Math.max(0, Number(item.attempts) || 0),
claimId: String(item.claimId ?? ""),
error: String(item.error ?? ""),
- statusCode: Math.max(0, Number(item.statusCode) || 0)
+ statusCode: Math.max(0, Number(item.statusCode) || 0),
+ serviceRequests: Math.max(0, Number(item.serviceRequests) || 0),
+ outputCorrectionRequests: Math.max(0, Number(item.outputCorrectionRequests) || 0),
+ cacheHit: item.cacheHit === true
}] : []),
usage: normalizeUsage(value.usage)
};
@@ -224,7 +242,7 @@ export function claimAnalysisItems(value, limit, idFactory = () => globalThis.cr
const job = requireJob(value);
if (job.status !== "running") return { job, claims: [] };
const maximum = limit === undefined
- ? job.kind === "vision" ? 1 : ANALYSIS_BATCH_CONCURRENCY
+ ? job.concurrency
: Math.max(1, Math.floor(limit));
const claims = [];
for (const item of job.items) {
@@ -247,13 +265,29 @@ export function claimAnalysisItems(value, limit, idFactory = () => globalThis.cr
return { job, claims };
}
-export function succeedAnalysisItem(value, entryId, claimId, usage, catalogRevision) {
+export function succeedAnalysisItem(value, entryId, claimId, usage, catalogRevision, metadata = {}) {
const job = requireJob(value);
const item = claimedItem(job, entryId, claimId);
item.status = "succeeded";
item.claimId = "";
item.error = "";
item.statusCode = 0;
+ recordItemExecution(item, metadata);
+ job.usage = addUsage(job.usage, usage);
+ job.resultCatalogRevision = Number.isInteger(catalogRevision) ? catalogRevision : job.resultCatalogRevision;
+ finishIfSettled(job);
+ touch(job);
+ return job;
+}
+
+export function partiallySucceedAnalysisItem(value, entryId, claimId, usage, catalogRevision, metadata = {}) {
+ const job = requireJob(value);
+ const item = claimedItem(job, entryId, claimId);
+ item.status = "partial";
+ item.claimId = "";
+ item.error = "分析结果待补全";
+ item.statusCode = 0;
+ recordItemExecution(item, metadata);
job.usage = addUsage(job.usage, usage);
job.resultCatalogRevision = Number.isInteger(catalogRevision) ? catalogRevision : job.resultCatalogRevision;
finishIfSettled(job);
@@ -268,9 +302,10 @@ export function failAnalysisItem(value, entryId, claimId, error = {}) {
item.claimId = "";
item.error = String(error.message ?? "分析失败");
item.statusCode = Math.max(0, Number(error.status) || 0);
+ recordItemExecution(item, error);
job.usage = addUsage(job.usage, error.usage);
finishIfSettled(job);
- if ([401, 402].includes(item.statusCode)) job.status = "paused";
+ if ([401, 402, 403].includes(item.statusCode)) job.status = "paused";
touch(job);
return job;
}
@@ -307,9 +342,14 @@ export function stageAnalysisRebuildResults(jobValue, stagingValue, state, resul
fingerprint: String(result.fingerprint ?? ""),
textRevision: Math.max(1, Number(result.textRevision) || 1),
model: String(result.model ?? ""),
+ normalizationDiagnostics: structuredClone(Array.isArray(result.normalizationDiagnostics) ? result.normalizationDiagnostics : []),
+ attempts: {
+ serviceRequests: Math.max(0, Number(result.attempts?.serviceRequests) || 0),
+ outputCorrectionRequests: Math.max(0, Number(result.attempts?.outputCorrectionRequests) || 0)
+ },
usage: normalizeUsage(result.usage)
};
- job = succeedAnalysisItem(job, result.entryId, result.claimId, result.usage, state.facetCatalog.revision);
+ job = succeedAnalysisItem(job, result.entryId, result.claimId, result.usage, state.facetCatalog.revision, result);
}
return { job, staging };
}
@@ -376,10 +416,13 @@ export function retryFailedAnalysisItems(value) {
const job = requireJob(value);
let count = 0;
for (const item of job.items) {
- if (item.status !== "failed") continue;
+ if (!["failed", "partial"].includes(item.status)) continue;
item.status = "pending";
item.error = "";
item.statusCode = 0;
+ item.serviceRequests = 0;
+ item.outputCorrectionRequests = 0;
+ item.cacheHit = false;
count += 1;
}
if (!count) throw new Error("没有可重试的失败案例");
@@ -414,10 +457,11 @@ export function reconcileVisionBatchResults(value, entries = []) {
if (!["pending", "running"].includes(item.status)) continue;
const analysis = visuals.get(`${item.entryId}:${item.visualId}`)?.visionAnalysis;
if (!analysis?.description || analysis.batchJobId !== job.id) continue;
- item.status = "succeeded";
+ item.status = analysis.quality === "partial" ? "partial" : "succeeded";
item.claimId = "";
item.error = "";
item.statusCode = 0;
+ recordItemExecution(item, analysis);
job.usage = addUsage(job.usage, {
promptTokens: analysis.usage?.inputTokens,
completionTokens: analysis.usage?.outputTokens,
@@ -435,9 +479,22 @@ export function reconcileVisionBatchResults(value, entries = []) {
export function analysisBatchSummary(value) {
const job = normalizeAnalysisBatchJob(value);
if (!job) return null;
- const counts = { pending: 0, running: 0, succeeded: 0, failed: 0 };
+ const counts = { pending: 0, running: 0, succeeded: 0, partial: 0, failed: 0 };
for (const item of job.items) counts[item.status] += 1;
- return { ...job, counts, total: job.items.length };
+ const failureCategories = {};
+ for (const item of job.items.filter((candidate) => candidate.status === "failed")) {
+ const category = failureCategory(item.statusCode);
+ failureCategories[category] = (failureCategories[category] ?? 0) + 1;
+ }
+ return {
+ ...job,
+ counts,
+ total: job.items.length,
+ requestAttempts: job.items.reduce((sum, item) => sum + item.serviceRequests, 0),
+ outputCorrectionRequests: job.items.reduce((sum, item) => sum + item.outputCorrectionRequests, 0),
+ cacheHitCount: job.items.filter((item) => item.cacheHit).length,
+ failureCategories
+ };
}
export function analysisRebuildRecovery(value, stagingValue) {
@@ -454,7 +511,7 @@ export function analysisRebuildRecovery(value, stagingValue) {
return {
stagedResultCount: stagedIds.length,
stagingValid,
- recoverable: job.status === "completed" && counts.failed > 0 && !job.partialApplied && stagingValid
+ recoverable: ["completed", "partial"].includes(job.status) && counts.failed > 0 && !job.partialApplied && stagingValid
};
}
@@ -510,7 +567,39 @@ function claimedItem(job, entryId, claimId) {
function finishIfSettled(job) {
if (job.items.some((item) => item.status === "pending" || item.status === "running")) return;
- job.status = "completed";
+ const succeeded = job.items.filter((item) => item.status === "succeeded").length;
+ const partial = job.items.filter((item) => item.status === "partial").length;
+ const failed = job.items.filter((item) => item.status === "failed").length;
+ job.status = failed === job.items.length ? "failed" : failed || partial ? "partial" : succeeded ? "completed" : "failed";
+}
+
+function normalizeConcurrency(value, fallback) {
+ const number = Number(value);
+ return Number.isInteger(number) && number >= 2 ? number : fallback;
+}
+
+function normalizeRetryPolicy(value) {
+ return {
+ serviceRetries: Number(value?.serviceRetries) === 2 ? 2 : ANALYSIS_RETRY_POLICY.serviceRetries,
+ outputCorrectionRequests: Number(value?.outputCorrectionRequests) === 1 ? 1 : ANALYSIS_RETRY_POLICY.outputCorrectionRequests,
+ backoffMs: [...ANALYSIS_RETRY_POLICY.backoffMs],
+ obeyRetryAfter: true
+ };
+}
+
+function recordItemExecution(item, metadata = {}) {
+ item.serviceRequests = Math.max(0, Number(metadata.attempts?.serviceRequests ?? metadata.serviceRequests) || 0);
+ item.outputCorrectionRequests = Math.max(0, Number(metadata.attempts?.outputCorrectionRequests ?? metadata.outputCorrectionRequests) || 0);
+ item.cacheHit = metadata.cacheHit === true;
+}
+
+function failureCategory(status) {
+ if ([401, 402, 403].includes(status)) return "authorization";
+ if (status === 429) return "rate_limit";
+ if (status === 408) return "timeout";
+ if (status >= 500) return "service";
+ if (status === 0) return "network";
+ return "output";
}
function requireJob(value) {
@@ -524,7 +613,7 @@ function touch(job) {
}
function emptyUsage() {
- return { promptTokens: 0, completionTokens: 0, totalTokens: 0, cacheHitTokens: 0, cacheMissTokens: 0 };
+ return { promptTokens: 0, completionTokens: 0, totalTokens: 0, cacheHitTokens: 0, cacheMissTokens: 0, cacheHits: 0 };
}
function pickEntryAnalysisState(entry) {
@@ -551,6 +640,11 @@ function rebuildAnalysisMeta(staged, job, entry, analyzedAt) {
model: staged.model,
analyzedAt,
profileFingerprint: job.profileFingerprint,
+ normalizationDiagnostics: structuredClone(Array.isArray(staged.normalizationDiagnostics) ? staged.normalizationDiagnostics : []),
+ attempts: {
+ serviceRequests: Math.max(0, Number(staged.attempts?.serviceRequests) || 0),
+ outputCorrectionRequests: Math.max(0, Number(staged.attempts?.outputCorrectionRequests) || 0)
+ },
usage: normalizeUsage(staged.usage)
};
}
diff --git a/analysis-candidates.js b/analysis-candidates.js
index fb2db8a..425ce9b 100644
--- a/analysis-candidates.js
+++ b/analysis-candidates.js
@@ -102,6 +102,16 @@ export function applyVisionAnalysis(state = {}, entryId, result = {}, metadata =
entry.visionAnalysis = {
version: Math.max(1, Number(metadata.version) || 1),
description,
+ quality: result.quality === "partial" ? "partial" : "complete",
+ missingFields: Array.isArray(result.missingFields) ? [...new Set(result.missingFields.map(String).filter(Boolean))] : [],
+ normalizationDiagnostics: Array.isArray(result.normalizationDiagnostics)
+ ? result.normalizationDiagnostics.map((item) => ({
+ field: String(item?.field ?? ""),
+ ...(String(item?.message ?? "").trim() ? { message: String(item.message).trim() } : {}),
+ ...(String(item?.code ?? "").trim() ? { code: String(item.code).trim() } : {}),
+ ...(Number(item?.count) > 0 ? { count: Number(item.count) } : {})
+ })).filter((item) => item.field)
+ : [],
...(result.canvas ? { canvas: structuredClone(result.canvas) } : {}),
...(Array.isArray(result.elements) ? { elements: structuredClone(result.elements) } : {}),
...(Array.isArray(result.dimensions) ? { dimensions: structuredClone(result.dimensions) } : {}),
@@ -113,10 +123,16 @@ export function applyVisionAnalysis(state = {}, entryId, result = {}, metadata =
locale: metadata.locale === "en" ? "en" : "zh-CN",
imageFingerprint: String(metadata.imageFingerprint ?? "").trim(),
profileFingerprint: String(metadata.profileFingerprint ?? "").trim(),
+ catalogRevision: Math.max(0, Number(metadata.catalogRevision) || 0),
analyzedAt: String(metadata.analyzedAt ?? "").trim() || new Date().toISOString(),
providerType: metadata.providerType === "compatible" ? "compatible" : "openai",
model: String(metadata.model ?? "").trim(),
...(metadata.usage && typeof metadata.usage === "object" ? { usage: structuredClone(metadata.usage) } : {}),
+ cacheHit: metadata.cacheHit === true,
+ attempts: {
+ serviceRequests: Math.max(0, Number(metadata.attempts?.serviceRequests) || 0),
+ outputCorrectionRequests: Math.max(0, Number(metadata.attempts?.outputCorrectionRequests) || 0)
+ },
...(String(metadata.batchJobId ?? "").trim() ? { batchJobId: String(metadata.batchJobId).trim() } : {}),
userEdited: false
};
diff --git a/analysis-retry-policy.js b/analysis-retry-policy.js
new file mode 100644
index 0000000..64d6496
--- /dev/null
+++ b/analysis-retry-policy.js
@@ -0,0 +1,6 @@
+export const ANALYSIS_RETRY_POLICY = Object.freeze({
+ serviceRetries: 2,
+ outputCorrectionRequests: 1,
+ backoffMs: Object.freeze([1000, 3000]),
+ obeyRetryAfter: true
+});
diff --git a/analysis-scheduler.js b/analysis-scheduler.js
new file mode 100644
index 0000000..dc2e87a
--- /dev/null
+++ b/analysis-scheduler.js
@@ -0,0 +1,88 @@
+import { ANALYSIS_RETRY_POLICY } from "./analysis-retry-policy.js";
+
+const schedulers = new Map();
+const inFlight = new Map();
+const RETRYABLE_STATUSES = new Set([0, 408, 429, 500, 502, 503, 504]);
+
+export function scheduleAnalysis(keyValue, concurrencyValue, task) {
+ const key = String(keyValue ?? "").trim();
+ if (!key || typeof task !== "function") throw new Error("分析调度参数无效");
+ const concurrency = normalizedConcurrency(concurrencyValue);
+ const state = schedulers.get(key) ?? { active: 0, limit: concurrency, effectiveLimit: concurrency, queue: [], throttledUntil: 0 };
+ state.limit = concurrency;
+ state.effectiveLimit = state.throttledUntil
+ ? Math.min(state.effectiveLimit, concurrency)
+ : concurrency;
+ if (!schedulers.has(key)) schedulers.set(key, state);
+ return new Promise((resolve, reject) => {
+ state.queue.push({ task, resolve, reject });
+ drain(state);
+ });
+}
+
+export function coalesceAnalysisRequest(keyValue, task, options = {}) {
+ const key = String(keyValue ?? "").trim();
+ if (!key) return task();
+ if (inFlight.has(key)) {
+ options.onCoalesced?.();
+ return inFlight.get(key);
+ }
+ const promise = Promise.resolve().then(task).finally(() => inFlight.delete(key));
+ inFlight.set(key, promise);
+ return promise;
+}
+
+export async function runScheduledAnalysisWithRetries(options = {}) {
+ const wait = options.wait ?? defaultWait;
+ const jitter = options.jitter ?? (() => Math.floor(Math.random() * 501));
+ let retries = 0;
+ for (;;) {
+ try {
+ return await scheduleAnalysis(options.key, options.concurrency, options.task);
+ } catch (error) {
+ if (!isRetryable(error) || retries >= ANALYSIS_RETRY_POLICY.serviceRetries) throw error;
+ const retryAfter = Number(error?.retryAfterMs) || 0;
+ const delay = retryAfter || ANALYSIS_RETRY_POLICY.backoffMs[retries] + Math.max(0, Number(jitter()) || 0);
+ retries += 1;
+ await wait(delay);
+ }
+ }
+}
+
+function drain(state) {
+ while (state.active < state.effectiveLimit && state.queue.length) {
+ const item = state.queue.shift();
+ state.active += 1;
+ Promise.resolve().then(item.task).then((value) => {
+ if (state.throttledUntil && Date.now() >= state.throttledUntil) {
+ state.effectiveLimit = Math.min(state.limit, state.effectiveLimit + 1);
+ if (state.effectiveLimit === state.limit) state.throttledUntil = 0;
+ }
+ item.resolve(value);
+ }, (error) => {
+ if ([429, 503].includes(Number(error?.status))) {
+ state.effectiveLimit = Math.max(1, Math.floor(state.effectiveLimit / 2));
+ state.throttledUntil = Date.now() + 30_000;
+ }
+ item.reject(error);
+ }).finally(() => {
+ state.active -= 1;
+ drain(state);
+ });
+ }
+}
+
+function isRetryable(error) {
+ if (error?.status === undefined || error?.status === null) return false;
+ const status = Number(error.status) || 0;
+ return error?.name !== "AbortError" && RETRYABLE_STATUSES.has(status);
+}
+
+function normalizedConcurrency(value) {
+ const number = Number(value);
+ return Number.isInteger(number) && number >= 2 ? number : 2;
+}
+
+function defaultWait(milliseconds) {
+ return new Promise((resolve) => setTimeout(resolve, milliseconds));
+}
diff --git a/automatic-vision.js b/automatic-vision.js
index 6be0b0e..7c7f869 100644
--- a/automatic-vision.js
+++ b/automatic-vision.js
@@ -10,7 +10,10 @@ export function buildAutomaticVisionJob(entries, entryIds, options = {}, current
includeAllImages: true,
reanalyze: false,
providerType: options.providerType,
- model: options.model
+ providerId: options.providerId,
+ model: options.model,
+ outputProtocol: options.outputProtocol,
+ concurrency: options.concurrency
});
if (!preview.requestCount) return null;
@@ -34,7 +37,10 @@ export function buildAutomaticVisionJob(entries, entryIds, options = {}, current
includeAllImages: true,
reanalyze: false,
providerType: options.providerType,
+ providerId: options.providerId,
model: options.model,
+ outputProtocol: options.outputProtocol,
+ concurrency: options.concurrency,
outputLocale: options.outputLocale,
now: options.now,
id: options.id
diff --git a/background.js b/background.js
index ce9073d..d6d863a 100644
--- a/background.js
+++ b/background.js
@@ -68,6 +68,7 @@ import {
finalizePartialAnalysisRebuild,
normalizeAnalysisBatchJob,
pauseAnalysisBatch,
+ partiallySucceedAnalysisItem,
previewAnalysisBatch,
previewVisionBatch,
reconcileVisionBatchResults,
@@ -79,6 +80,8 @@ import {
textFingerprint
} from "./analysis-batch.js";
import { buildAutomaticVisionJob } from "./automatic-vision.js";
+import { coalesceAnalysisRequest, runScheduledAnalysisWithRetries } from "./analysis-scheduler.js";
+import { findPersistedVisionAnalysis } from "./vision-analysis-cache.js";
import {
analysisRevisionMeta,
entryTextRevision,
@@ -304,7 +307,8 @@ import {
analyzeImageWithVision,
blobToDataUrl,
imageFingerprint,
- publicVisionSettings
+ publicVisionSettings,
+ visionAnalysisProfileFingerprint
} from "./vision.js";
import {
activateCreativeResultContext,
@@ -800,7 +804,24 @@ async function handleMessage(message, interaction = {}) {
case "ANALYZE_CREATIVE_OUTPUT":
return analyzeCreativeOutput(message.runId, message.visualId);
case "ANALYZE_ENTRY_IMAGE":
- return analyzeEntryImage(message.entryId, message.visualId, message.outputLocale, message.batchJobId);
+ try {
+ return await analyzeEntryImage(
+ message.entryId,
+ message.visualId,
+ message.outputLocale,
+ message.batchJobId,
+ message.bypassCache === true,
+ message.assignment
+ );
+ } catch (error) {
+ return {
+ ok: false,
+ message: userMessage(error),
+ status: Number(error?.status) || 0,
+ usage: error?.usage,
+ attempts: error?.attempts
+ };
+ }
case "ANALYZE_ENTRY_VISUAL_SET":
return analyzeEntryVisualSet(message);
case "ANALYZE_ENTRY_VIDEO":
@@ -3637,8 +3658,6 @@ async function saveTempReferenceAsCaseAction(message) {
}
async function analyzeTempReferencesAction(message) {
- if (visionAnalysisInFlight) return { ok: false, message: "已有图片正在分析,请等待完成" };
- visionAnalysisInFlight = true;
try {
const sessionId = String(message.sessionId ?? "").trim();
const requestedIds = [...new Set((Array.isArray(message.tempReferenceIds) ? message.tempReferenceIds : [])
@@ -3646,7 +3665,8 @@ async function analyzeTempReferencesAction(message) {
if (!sessionId || !requestedIds.length) return { ok: false, message: "没有可分析的临时图片" };
const stored = await chrome.storage.local.get([
STORAGE_KEYS.composerSessions,
- STORAGE_KEYS.facetCatalog
+ STORAGE_KEYS.facetCatalog,
+ STORAGE_KEYS.entries
]);
const sessions = normalizeComposerSessions(stored[STORAGE_KEYS.composerSessions]);
const session = sessions.find((item) => item.id === sessionId);
@@ -3659,37 +3679,45 @@ async function analyzeTempReferencesAction(message) {
if (references.length !== requested.size) {
return { ok: false, message: "参考图片已经变化,请确认后重试" };
}
- const settings = resolveVisionTaskSettings("imageAnalysis", await loadAiConfiguration());
+ const configuration = await loadAiConfiguration();
const locale = message.outputLocale === "en" ? "en" : "zh-CN";
- const analyzed = [];
- for (const reference of references) {
+ const settledReferences = await Promise.allSettled(references.map(async (reference) => {
const imageAssets = unreadReferenceImageAssets(reference);
- const descriptions = [];
- const fingerprints = [];
- const analyses = [];
- for (const asset of imageAssets) {
+ const assetResults = await Promise.all(imageAssets.map(async (asset) => {
const blob = await getTempReferenceVisionBlob(asset);
if (!blob) throw new Error(`参考图片已经失效:${asset.name || "未命名图片"}`);
- const fingerprint = await imageFingerprint(blob);
- const result = await analyzeImageWithVision({
- imageDataUrl: await blobToDataUrl(blob),
+ const { result, fingerprint } = await analyzeVisionBlobWithScheduler({
+ blob,
catalog: stored[STORAGE_KEYS.facetCatalog],
locale,
- settings
- });
- descriptions.push(result.description);
- fingerprints.push({ assetId: asset.assetId, fingerprint });
- analyses.push({
- assetId: asset.assetId,
- imageFingerprint: fingerprint,
- analysisImageFingerprint: fingerprint,
- analysisVersion: VISION_ANALYSIS_VERSION,
- analysisFingerprint: result.profileFingerprint,
- reconstructionPrompt: result.reconstructionPrompt
+ configuration,
+ entries: stored[STORAGE_KEYS.entries]
});
- }
- analyzed.push({ referenceId: reference.entryId, descriptions, fingerprints, analyses });
- }
+ return {
+ description: result.description,
+ fingerprint: { assetId: asset.assetId, fingerprint },
+ analysis: {
+ assetId: asset.assetId,
+ imageFingerprint: fingerprint,
+ analysisImageFingerprint: fingerprint,
+ analysisVersion: VISION_ANALYSIS_VERSION,
+ analysisFingerprint: result.profileFingerprint,
+ reconstructionPrompt: result.reconstructionPrompt
+ }
+ };
+ }));
+ return {
+ referenceId: reference.entryId,
+ descriptions: assetResults.map((item) => item.description),
+ fingerprints: assetResults.map((item) => item.fingerprint),
+ analyses: assetResults.map((item) => item.analysis)
+ };
+ }));
+ const analyzed = settledReferences.flatMap((settled) => settled.status === "fulfilled" ? [settled.value] : []);
+ const failures = settledReferences.flatMap((settled, index) => settled.status === "rejected"
+ ? [{ referenceId: references[index].entryId, message: userMessage(settled.reason) }]
+ : []);
+ if (!analyzed.length) throw new Error(failures[0]?.message || "临时图片分析失败");
return await enqueue(async () => {
const latestStored = await chrome.storage.local.get(STORAGE_KEYS.composerSessions);
@@ -3731,15 +3759,18 @@ async function analyzeTempReferencesAction(message) {
await commitLocalChanges({ [STORAGE_KEYS.composerSessions]: next });
return {
ok: true,
- message: `已完成 ${analyzed.length} 项临时图片分析`,
+ quality: failures.length ? "partial" : "complete",
+ failedCount: failures.length,
+ failures,
+ message: failures.length
+ ? `已保存 ${analyzed.length} 项分析,另有 ${failures.length} 项失败,可直接重试`
+ : `已完成 ${analyzed.length} 项临时图片分析`,
session: updatedSession,
summaries: next.map(sessionSummary)
};
});
} catch (error) {
return { ok: false, message: userMessage(error) };
- } finally {
- visionAnalysisInFlight = false;
}
}
@@ -4626,10 +4657,7 @@ function canonicalAiTaskId(value) {
return map[id] || id;
}
-async function analyzeEntryImage(entryId, visualIdValue, outputLocale, batchJobIdValue = "") {
- if (visionAnalysisInFlight) return { ok: false, message: "已有一张图片正在分析,请等待完成" };
- visionAnalysisInFlight = true;
- try {
+async function analyzeEntryImage(entryId, visualIdValue, outputLocale, batchJobIdValue = "", bypassCache = false, assignmentOverride = null) {
const state = await readState();
const entry = findEntry(state, entryId);
const visualId = String(visualIdValue ?? "").trim() || primaryVisual(entry)?.id;
@@ -4637,14 +4665,18 @@ async function analyzeEntryImage(entryId, visualIdValue, outputLocale, batchJobI
if (!visual) return { ok: false, message: "这条案例没有可分析的截图" };
const blob = await getVisionImageBlob(visual.id);
if (!blob) return { ok: false, message: "没有读取到这条案例的截图" };
- const fingerprint = await imageFingerprint(blob);
- const settings = resolveVisionTaskSettings("imageAnalysis", await loadAiConfiguration());
- const result = await analyzeImageWithVision({
- imageDataUrl: await blobToDataUrl(blob),
+ const loadedConfiguration = await loadAiConfiguration();
+ const configuration = configurationForAssignment(loadedConfiguration, "imageAnalysis", assignmentOverride);
+ const locale = outputLocale === "en" ? "en" : "zh-CN";
+ const { result, fingerprint } = await analyzeVisionBlobWithScheduler({
+ blob,
catalog: state.facetCatalog,
- locale: outputLocale === "en" ? "en" : "zh-CN",
+ locale,
measuredCanvas: { width: visual.width, height: visual.height },
- settings
+ configuration,
+ previousAnalysis: visual.visionAnalysis,
+ entries: state.entries,
+ bypassCache
});
return await enqueue(async () => {
@@ -4664,10 +4696,13 @@ async function analyzeEntryImage(entryId, visualIdValue, outputLocale, batchJobI
visualId: visual.id,
imageFingerprint: fingerprint,
profileFingerprint: result.profileFingerprint,
+ catalogRevision: state.facetCatalog.revision,
locale: outputLocale,
providerType: result.providerType,
model: result.model,
usage: result.usage,
+ cacheHit: result.cacheHit,
+ attempts: result.attempts,
batchJobId: String(batchJobIdValue ?? "").trim()
});
const analyzed = applied.state.entries.find((item) => item.id === current.id);
@@ -4689,18 +4724,99 @@ async function analyzeEntryImage(entryId, visualIdValue, outputLocale, batchJobI
});
return {
ok: true,
- message: result.tagDiagnostics?.rejectedCount
+ message: result.quality === "partial"
+ ? `画面分析已保存可用部分,仍有 ${result.missingFields.length} 项待补全`
+ : result.tagDiagnostics?.rejectedCount
? `画面分析已保存,并写入 ${applied.appliedCount} 个检索标签;另有 ${result.tagDiagnostics.rejectedCount} 个不符合当前分类体系的可选标签未采用`
: `画面分析已保存,并写入 ${applied.appliedCount} 个检索标签`,
entry: applied.state.entries.find((item) => item.id === current.id),
usage: result.usage,
+ cacheHit: result.cacheHit,
+ quality: result.quality,
+ missingFields: result.missingFields,
...publicDomainState(applied.state),
visionUndoEntryIds: Object.keys(undoStore)
};
});
- } finally {
- visionAnalysisInFlight = false;
+}
+
+function configurationForAssignment(configuration, taskId, assignmentOverride) {
+ const providerId = String(assignmentOverride?.providerId ?? "").trim();
+ const model = String(assignmentOverride?.model ?? "").trim();
+ if (!providerId && !model) return configuration;
+ return {
+ ...configuration,
+ assignments: normalizeAiTaskAssignments({
+ ...configuration.assignments,
+ [taskId]: {
+ ...configuration.assignments[taskId],
+ providerId: providerId || configuration.assignments[taskId]?.providerId,
+ model: model || configuration.assignments[taskId]?.model,
+ concurrency: assignmentOverride?.concurrency ?? configuration.assignments[taskId]?.concurrency
+ }
+ }, configuration.registry)
+ };
+}
+
+async function analyzeVisionBlobWithScheduler({ blob, catalog, locale, measuredCanvas, configuration, previousAnalysis, entries = [], bypassCache = false }) {
+ const fingerprint = await imageFingerprint(blob);
+ const assignment = configuration.assignments.imageAnalysis;
+ const settings = resolveVisionTaskSettings("imageAnalysis", configuration);
+ const profileFingerprint = await visionAnalysisProfileFingerprint(settings, locale);
+ const cached = findPersistedVisionAnalysis(entries, {
+ fingerprint,
+ profileFingerprint,
+ locale,
+ catalogRevision: catalog?.revision
+ });
+ if (!bypassCache && cached?.quality === "complete") return {
+ fingerprint,
+ result: {
+ ...cached,
+ cacheHit: true,
+ attempts: { serviceRequests: 0, outputCorrectionRequests: 0 },
+ usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0 }
+ }
+ };
+ const schedulerKey = `${assignment.providerId}:${assignment.model}:imageAnalysis`;
+ const requestKey = `${schedulerKey}:${fingerprint}:${locale}:${catalog?.revision ?? 0}`;
+ let serviceRequests = 0;
+ let coalesced = false;
+ let result;
+ try {
+ result = await coalesceAnalysisRequest(requestKey, () => runScheduledAnalysisWithRetries({
+ key: schedulerKey,
+ concurrency: assignment.concurrency,
+ task: async () => {
+ serviceRequests += 1;
+ return analyzeImageWithVision({
+ imageDataUrl: await blobToDataUrl(blob),
+ catalog,
+ locale,
+ measuredCanvas,
+ previousAnalysis: !bypassCache && previousAnalysis?.quality === "partial" ? previousAnalysis : !bypassCache ? cached : null,
+ settings
+ });
+ }
+ }), { onCoalesced: () => { coalesced = true; } });
+ } catch (error) {
+ error.attempts = {
+ serviceRequests,
+ outputCorrectionRequests: Math.max(0, Number(error?.attempts?.outputCorrectionRequests) || 0)
+ };
+ throw error;
}
+ return {
+ result: {
+ ...result,
+ cacheHit: result.cacheHit === true || coalesced,
+ attempts: {
+ serviceRequests,
+ outputCorrectionRequests: Number(result.attempts?.outputCorrectionRequests) || 0
+ }
+ },
+ fingerprint
+ };
}
async function analyzeEntryVisualSet(message) {
@@ -5095,7 +5211,8 @@ async function createDeepSeekBatch(outputLocale, mode = "incremental") {
STORAGE_KEYS.batchJob,
STORAGE_KEYS.analysisRebuildStaging
]);
- const settings = resolveTextTaskSettings("textTags", await loadAiConfiguration());
+ const configuration = await loadAiConfiguration();
+ const settings = resolveTextTaskSettings("textTags", configuration);
if (!publicAiSettings(settings).configured) return { ok: false, message: "请先完成所选 AI 服务配置" };
if (!settings.consent) return { ok: false, message: "请先确认:主动分析时会发送案例文字到所选 AI 服务" };
const previous = normalizeAnalysisBatchJob(stored[STORAGE_KEYS.batchJob]);
@@ -5121,7 +5238,10 @@ async function createDeepSeekBatch(outputLocale, mode = "incremental") {
profileFingerprint,
mode,
catalogRevision: state.facetCatalog.revision,
- fixedTaxonomyCharacters: analysisTaxonomyPrompt(state.facetCatalog, locale).length
+ fixedTaxonomyCharacters: analysisTaxonomyPrompt(state.facetCatalog, locale).length,
+ concurrency: configuration.assignments.textTags.concurrency,
+ providerId: configuration.assignments.textTags.providerId,
+ outputProtocol: "json_object"
});
const changes = { [STORAGE_KEYS.batchJob]: job };
if (job.mode === "rebuild") {
@@ -5200,7 +5320,8 @@ async function commitDeepSeekBatchItem(message) {
message.entryId,
message.claimId,
message.usage,
- applied.state.facetCatalog.revision
+ applied.state.facetCatalog.revision,
+ message
);
await commitLocalChanges({
...storagePayload(applied.state),
@@ -5255,7 +5376,7 @@ async function commitDeepSeekBatchItems(message) {
...result,
profileFingerprint: job.profileFingerprint
}, result.fingerprint, updated.analyzedAt, updated);
- job = succeedAnalysisItem(job, result.entryId, result.claimId, result.usage, working.facetCatalog.revision);
+ job = succeedAnalysisItem(job, result.entryId, result.claimId, result.usage, working.facetCatalog.revision, result);
}
await commitLocalChanges({
...storagePayload(working),
@@ -5429,7 +5550,8 @@ async function undoDeepSeekBatch(jobId) {
async function previewVisionBatchTask(message) {
const state = await readState();
- const settings = resolveVisionTaskSettings("imageAnalysis", await loadAiConfiguration(), { requireConfigured: false });
+ const configuration = await loadAiConfiguration();
+ const settings = resolveVisionTaskSettings("imageAnalysis", configuration, { requireConfigured: false });
const provider = settings[settings.activeProvider];
const preview = previewVisionBatch(state.entries, {
entryIds: message.entryIds,
@@ -5444,7 +5566,8 @@ async function previewVisionBatchTask(message) {
async function createVisionBatchTask(message) {
const state = await readState();
const stored = await chrome.storage.local.get(STORAGE_KEYS.batchJob);
- const settings = resolveVisionTaskSettings("imageAnalysis", await loadAiConfiguration(), { requireConfigured: false });
+ const configuration = await loadAiConfiguration();
+ const settings = resolveVisionTaskSettings("imageAnalysis", configuration, { requireConfigured: false });
const provider = settings[settings.activeProvider];
if (!provider.apiKey) return { ok: false, message: "请先保存当前图片服务的 API Key" };
if (!settings.consent) return { ok: false, message: "请先确认主动分析时会发送当前截图" };
@@ -5458,8 +5581,11 @@ async function createVisionBatchTask(message) {
includeAllImages: message.includeAllImages,
reanalyze: message.reanalyze,
providerType: settings.activeProvider,
+ providerId: configuration.assignments.imageAnalysis.providerId,
model: provider.model,
- outputLocale: message.outputLocale
+ outputProtocol: settings.activeProvider === "compatible" ? settings.compatible.structuredOutput : "json_schema",
+ outputLocale: message.outputLocale,
+ concurrency: configuration.assignments.imageAnalysis.concurrency
});
await commitLocalChanges({ [STORAGE_KEYS.batchJob]: job });
await chrome.storage.local.remove(STORAGE_KEYS.analysisBatchUndo);
@@ -5481,19 +5607,22 @@ async function claimVisionBatchItem(jobId) {
if (reconciled.recoveredCount) {
await commitLocalChanges({ [STORAGE_KEYS.batchJob]: job });
}
- const settings = resolveVisionTaskSettings("imageAnalysis", await loadAiConfiguration(), { requireConfigured: false });
- const provider = settings[settings.activeProvider];
- if (settings.activeProvider !== job.providerType || provider.model !== job.model) {
+ const loadedConfiguration = await loadAiConfiguration();
+ const configuration = configurationForAssignment(loadedConfiguration, "imageAnalysis", job);
+ const settings = resolveVisionTaskSettings("imageAnalysis", configuration, { requireConfigured: false });
+ const snapshotReady = publicVisionSettings(settings)[settings.activeProvider]?.configured === true;
+ if (!snapshotReady || !settings.consent) {
const paused = pauseAnalysisBatch(job);
await commitLocalChanges({ [STORAGE_KEYS.batchJob]: paused });
- return { ok: false, message: "图片服务或模型已变化,任务已暂停,请重新创建预览" };
+ return { ok: false, message: "任务快照所用服务当前不可用,任务已暂停" };
}
- const claimed = claimAnalysisItems(job, 1);
+ const claimed = claimAnalysisItems(job);
if (claimed.claims.length) {
await commitLocalChanges({ [STORAGE_KEYS.batchJob]: claimed.job });
}
return {
ok: true,
+ claims: claimed.claims,
claim: claimed.claims[0] ?? null,
visionBatchJob: analysisBatchSummary(claimed.job)
};
@@ -5506,9 +5635,12 @@ async function completeVisionBatchItem(message) {
const usage = {
promptTokens: Math.max(0, Number(message.usage?.inputTokens) || 0),
completionTokens: Math.max(0, Number(message.usage?.outputTokens) || 0),
- totalTokens: Math.max(0, Number(message.usage?.totalTokens) || 0)
+ totalTokens: Math.max(0, Number(message.usage?.totalTokens) || 0),
+ cacheHits: message.cacheHit === true ? 1 : 0
};
- const next = succeedAnalysisItem(job, message.entryId, message.claimId, usage);
+ const next = message.quality === "partial"
+ ? partiallySucceedAnalysisItem(job, message.entryId, message.claimId, usage, undefined, message)
+ : succeedAnalysisItem(job, message.entryId, message.claimId, usage, undefined, message);
await commitLocalChanges({ [STORAGE_KEYS.batchJob]: next });
return { ok: true, visionBatchJob: analysisBatchSummary(next) };
}
@@ -5547,14 +5679,18 @@ async function queueAutomaticVisionAnalysis(entryIdsValue, options = {}) {
readState(),
chrome.storage.local.get(STORAGE_KEYS.automaticVisionBatchJob)
]);
- const settings = resolveVisionTaskSettings("imageAnalysis", await loadAiConfiguration(), { requireConfigured: false });
+ const configuration = await loadAiConfiguration();
+ const settings = resolveVisionTaskSettings("imageAnalysis", configuration, { requireConfigured: false });
const publicSettings = publicVisionSettings(settings);
if ((options.requireAutoImportSetting !== false && !settings.autoAnalyzeImports) ||
!settings.consent || !publicSettings[settings.activeProvider].configured) return false;
const provider = settings[settings.activeProvider];
const job = buildAutomaticVisionJob(state.entries, entryIds, {
providerType: settings.activeProvider,
+ providerId: configuration.assignments.imageAnalysis.providerId,
model: provider.model,
+ outputProtocol: settings.activeProvider === "compatible" ? settings.compatible.structuredOutput : "json_schema",
+ concurrency: configuration.assignments.imageAnalysis.concurrency,
outputLocale: resolveLocale(state.uiPreferences, chrome.i18n.getUILanguage()) === "en" ? "en" : "zh-CN"
}, stored[STORAGE_KEYS.automaticVisionBatchJob]);
if (!job) return false;
@@ -5581,44 +5717,58 @@ async function runAutomaticVisionItem() {
STORAGE_KEYS.automaticVisionBatchJob
]);
let job = normalizeAnalysisBatchJob(stored[STORAGE_KEYS.automaticVisionBatchJob]);
- const settings = resolveVisionTaskSettings("imageAnalysis", await loadAiConfiguration(), { requireConfigured: false });
- const configured = publicVisionSettings(settings)[settings.activeProvider].configured;
- if (!job || job.kind !== "vision" || job.status !== "running" ||
- !settings.autoAnalyzeImports || !settings.consent || !configured) {
+ if (!job || job.kind !== "vision" || job.status !== "running") {
await ensureAutomaticVisionAlarm(false);
return;
}
- if (visionAnalysisInFlight) {
- continueRunning = true;
+ const loadedConfiguration = await loadAiConfiguration();
+ const configuration = configurationForAssignment(loadedConfiguration, "imageAnalysis", job);
+ const settings = resolveVisionTaskSettings("imageAnalysis", configuration, { requireConfigured: false });
+ const configured = publicVisionSettings(settings)[settings.activeProvider].configured;
+ if (!settings.autoAnalyzeImports || !settings.consent || !configured) {
+ job = pauseAnalysisBatch(job);
+ await commitLocalChanges({ [STORAGE_KEYS.automaticVisionBatchJob]: job });
+ await ensureAutomaticVisionAlarm(false);
return;
}
const state = await readState();
job = resumeAnalysisBatch(job);
job = reconcileVisionBatchResults(job, state.entries).job;
- const claimed = claimAnalysisItems(job, 1);
+ const claimed = claimAnalysisItems(job);
job = claimed.job;
await commitLocalChanges({ [STORAGE_KEYS.automaticVisionBatchJob]: job });
- const claim = claimed.claims[0];
- if (!claim) {
+ if (!claimed.claims.length) {
await ensureAutomaticVisionAlarm(false);
return;
}
-
- const result = await analyzeEntryImage(claim.entryId, claim.visualId, job.outputLocale, job.id);
- await enqueue(async () => {
- const latestStored = await chrome.storage.local.get(STORAGE_KEYS.automaticVisionBatchJob);
- const latest = requireAnalysisBatch(latestStored[STORAGE_KEYS.automaticVisionBatchJob], job.id, "vision");
- requireClaim(latest, claim.entryId, claim.claimId, claim.visualId);
- const next = result.ok
- ? succeedAnalysisItem(latest, claim.entryId, claim.claimId, {
- promptTokens: result.usage?.inputTokens,
- completionTokens: result.usage?.outputTokens,
- totalTokens: result.usage?.totalTokens
- })
- : { ...failAnalysisItem(latest, claim.entryId, claim.claimId, { message: result.message }), status: "paused" };
- await commitLocalChanges({ [STORAGE_KEYS.automaticVisionBatchJob]: next });
- job = next;
- });
+ await Promise.all(claimed.claims.map(async (claim) => {
+ let result;
+ try {
+ result = await analyzeEntryImage(claim.entryId, claim.visualId, job.outputLocale, job.id, false, job);
+ } catch (error) {
+ result = { ok: false, message: userMessage(error), status: Number(error?.status) || 0, usage: error?.usage };
+ }
+ await enqueue(async () => {
+ const latestStored = await chrome.storage.local.get(STORAGE_KEYS.automaticVisionBatchJob);
+ const latest = requireAnalysisBatch(latestStored[STORAGE_KEYS.automaticVisionBatchJob], job.id, "vision");
+ requireClaim(latest, claim.entryId, claim.claimId, claim.visualId);
+ const next = result.ok
+ ? (result.quality === "partial" ? partiallySucceedAnalysisItem : succeedAnalysisItem)(latest, claim.entryId, claim.claimId, {
+ promptTokens: result.usage?.inputTokens,
+ completionTokens: result.usage?.outputTokens,
+ totalTokens: result.usage?.totalTokens,
+ cacheHits: result.cacheHit === true ? 1 : 0
+ }, undefined, result)
+ : failAnalysisItem(latest, claim.entryId, claim.claimId, {
+ message: result.message,
+ status: result.status,
+ usage: result.usage,
+ attempts: result.attempts
+ });
+ await commitLocalChanges({ [STORAGE_KEYS.automaticVisionBatchJob]: next });
+ job = next;
+ });
+ }));
continueRunning = job.status === "running";
await ensureAutomaticVisionAlarm(continueRunning);
} catch (error) {
@@ -6048,6 +6198,15 @@ function analysisMeta(message, fingerprint, analyzedAt, entry = {}) {
model: String(message.model ?? ""),
analyzedAt,
profileFingerprint: String(message.profileFingerprint ?? "").trim(),
+ normalizationDiagnostics: (Array.isArray(message.normalizationDiagnostics) ? message.normalizationDiagnostics : []).map((item) => ({
+ field: String(item?.field ?? ""),
+ code: String(item?.code ?? ""),
+ count: Math.max(0, Number(item?.count) || 0)
+ })).filter((item) => item.field && item.code),
+ attempts: {
+ serviceRequests: Math.max(0, Number(message.attempts?.serviceRequests) || 0),
+ outputCorrectionRequests: Math.max(0, Number(message.attempts?.outputCorrectionRequests) || 0)
+ },
usage: {
promptTokens: Math.max(0, Number(message.usage?.promptTokens) || 0),
completionTokens: Math.max(0, Number(message.usage?.completionTokens) || 0),
diff --git a/deepseek.js b/deepseek.js
index 9b0ccbb..1666fb4 100644
--- a/deepseek.js
+++ b/deepseek.js
@@ -3,6 +3,8 @@ import {
validateAnalysisTagResponse,
validateDetailOrganizationResponse
} from "./tag-taxonomy.js";
+import { parseStructuredObject } from "./structured-output.js";
+import { ANALYSIS_RETRY_POLICY } from "./analysis-retry-policy.js";
import {
normalizeComposerAiProfile,
normalizePlannerResult,
@@ -20,7 +22,7 @@ export const DEEPSEEK_ENDPOINT = "https://api.deepseek.com/chat/completions";
export const DEFAULT_ANALYSIS_MODEL = "deepseek-v4-flash";
export const DEFAULT_COMPOSER_STREAM_TIMEOUT_MS = 120_000;
export const DEFAULT_COMPOSER_REQUEST_TIMEOUT_MS = 120_000;
-export const ANALYSIS_SERVICE_RETRY_LIMIT = 3;
+export const ANALYSIS_SERVICE_RETRY_LIMIT = ANALYSIS_RETRY_POLICY.serviceRetries;
export const ANALYSIS_CLAIM_TIMEOUT_MS = DEFAULT_COMPOSER_REQUEST_TIMEOUT_MS * (ANALYSIS_SERVICE_RETRY_LIMIT + 1) + 30_000;
const ANALYSIS_OUTPUT_CORRECTION_TIMEOUT_MS = DEFAULT_COMPOSER_REQUEST_TIMEOUT_MS / 2;
export const ANALYSIS_PROMPT_VERSION = 9;
@@ -180,10 +182,13 @@ export async function analyzeTextDetailedWithDeepSeek(entry, catalogValue, setti
elapsedMs: Date.now() - requestStartedAt,
tagCount
});
- const tags = validateAnalysisTagResponse(parsed, catalogValue);
+ const normalizationDiagnostics = [];
+ const tags = validateAnalysisTagResponse(parsed, catalogValue, { diagnostics: normalizationDiagnostics });
emitAnalysisDiagnostic(requestOptions, "validation_succeeded", { attempt, tagCount: tags.length });
return {
tags,
+ normalizationDiagnostics,
+ attempts: { outputCorrectionRequests: outputAttempt },
usage,
model: result.model || settings.analysisModel,
finishReason: result.finishReason
@@ -554,9 +559,8 @@ async function requestDeepSeek(body, settings, options = {}) {
options.signal?.removeEventListener("abort", onExternalAbort);
}
if (!response.ok) {
- const retryAfter = Number(response.headers?.get?.("retry-after"));
throw new DeepSeekApiError(apiError(payload, response.status), response.status, {
- retryAfterMs: Number.isFinite(retryAfter) ? retryAfter * 1000 : 0
+ retryAfterMs: retryAfterMilliseconds(response.headers?.get?.("retry-after"))
});
}
const choice = payload?.choices?.[0];
@@ -689,11 +693,8 @@ function deepSeekErrorKind(status) {
}
function parseJsonObject(content, message) {
- const cleaned = String(content).trim().replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/, "");
try {
- const parsed = JSON.parse(cleaned);
- if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) throw new Error();
- return parsed;
+ return parseStructuredObject(content, message);
} catch {
throw new Error(message);
}
@@ -727,6 +728,13 @@ function finite(value) {
return Number.isFinite(number) && number > 0 ? Math.round(number) : 0;
}
+function retryAfterMilliseconds(value) {
+ const seconds = Number(value);
+ if (Number.isFinite(seconds) && seconds >= 0) return seconds * 1000;
+ const date = Date.parse(String(value ?? ""));
+ return Number.isFinite(date) ? Math.max(0, date - Date.now()) : 0;
+}
+
function apiError(payload, status) {
const detail = String(payload?.error?.message ?? "").trim();
if (status === 401) return "DeepSeek API Key 无效或已失效";
diff --git a/i18n.js b/i18n.js
index 2c4d6ac..2c8fd6a 100644
--- a/i18n.js
+++ b/i18n.js
@@ -154,6 +154,7 @@ const EN = Object.freeze({
"提示词": "Prompt",
"分析主图": "Analyze primary image",
"重新分析主图": "Reanalyze primary image",
+ "补全主图分析": "Complete primary image analysis",
"编辑当前图片": "Edit current image",
"编辑共享提示词": "Edit shared prompt",
"添加视频链接": "Add video link",
@@ -368,8 +369,17 @@ const EN = Object.freeze({
"不生成整组总结": "No group summary",
"正在分析…": "Analyzing…",
"本次已完成": "Completed this run",
+ "待补全 · 已保存有效结果": "Needs completion · Valid results saved",
"失败 · {message}": "Failed · {message}",
"{failed} 张失败,已成功 {completed} 张且不会重复请求。可直接重试失败项。": "{failed} images failed. {completed} succeeded and will not be requested again. You can retry only the failed items.",
+ "{failed} 张失败,{partial} 张待补全,已成功 {completed} 张且不会重复请求。可直接重试失败或待补全项。": "{failed} images failed, {partial} need completion, and {completed} succeeded without repeat requests. Retry only failed or incomplete items.",
+ "批量分析并发数": "Batch analysis concurrency",
+ "文字任务默认 20,图片和视频任务默认 10;调整只对新任务生效。": "Text tasks default to 20; image and video tasks default to 10. Changes apply only to new jobs.",
+ "我确认高于 20 的并发可能明显增加瞬时费用、内存占用和错误数量": "I understand concurrency above 20 may sharply increase burst cost, memory use, and errors",
+ "官方并发上限:{count}": "Official concurrency limit: {count}",
+ "服务商未提供固定并发上限": "The provider has not published a fixed concurrency limit",
+ "批量分析并发数必须是大于或等于 2 的整数": "Batch analysis concurrency must be an integer of at least 2",
+ "并发高于 20,请先确认费用和稳定性风险": "Confirm the cost and stability risks before using concurrency above 20",
"待确认 {count}": "{count} to review",
"归档这些标签?": "Archive these tags?",
"确认归档": "Archive",
diff --git a/image-prompt.js b/image-prompt.js
index 8111553..8f5f95b 100644
--- a/image-prompt.js
+++ b/image-prompt.js
@@ -2,7 +2,7 @@ import { entryMediaAssets } from "./media.js";
export function validReconstructionPrompt(assetValue = {}) {
const analysis = assetValue?.visionAnalysis;
- if (!analysis || analysis.invalidated || Number(analysis.version) !== 2) return "";
+ if (!analysis || analysis.invalidated || analysis.quality === "partial" || Number(analysis.version) !== 2) return "";
const prompt = String(analysis.reconstructionPrompt ?? "").trim();
if (!prompt) return "";
const currentFingerprint = String(assetValue.contentHash ?? "").trim() || String(analysis.imageFingerprint ?? "").trim();
diff --git a/library.js b/library.js
index 5a41b82..5b626cd 100644
--- a/library.js
+++ b/library.js
@@ -446,7 +446,7 @@ function renderAnalysisBatch() {
const job = analysisBatchJob;
const preview = analysisBatchPreview;
const active = Boolean(job && ["running", "paused"].includes(job.status));
- const unfinishedRebuild = Boolean(job?.mode === "rebuild" && job?.status === "completed" && job?.counts.failed && !job.partialApplied);
+ const unfinishedRebuild = Boolean(job?.mode === "rebuild" && ["completed", "partial"].includes(job?.status) && job?.counts.failed && !job.partialApplied);
const partialRebuildApplied = Boolean(job?.mode === "rebuild" && job?.partialApplied);
if (preview && !active && !unfinishedRebuild) {
elements.batchStatusBadge.textContent = t("本次预览");
@@ -463,11 +463,11 @@ function renderAnalysisBatch() {
elements.batchStatusBadge.textContent = t("尚未开始");
elements.analysisBatchSummary.textContent = t("先生成预览,不会直接产生 API 请求。");
} else {
- const statusNames = { running: "分析中", paused: "已暂停", completed: unfinishedRebuild ? "重建待完成" : partialRebuildApplied ? "成功结果已应用" : "上次任务", canceled: "上次已取消" };
+ const statusNames = { running: "分析中", paused: "已暂停", completed: unfinishedRebuild ? "重建待完成" : partialRebuildApplied ? "成功结果已应用" : "已完成", partial: "部分完成", failed: "分析失败", canceled: "上次已取消" };
elements.batchStatusBadge.textContent = t(statusNames[job.status] || job.status);
const progress = currentLocale() === "en"
- ? [`${job.counts.succeeded}/${job.total} completed`, `${job.counts.running} active`, `${job.counts.pending} queued`, `${job.counts.failed} failed`, `input ${job.usage.promptTokens} / output ${job.usage.completionTokens} tokens`, `cache hits ${job.usage.cacheHitTokens} tokens`].join(" · ")
- : [`${job.counts.succeeded}/${job.total} 已完成`, `${job.counts.running} 处理中`, `${job.counts.pending} 等待中`, `${job.counts.failed} 失败`, `输入 ${job.usage.promptTokens} / 输出 ${job.usage.completionTokens} tokens`, `缓存命中 ${job.usage.cacheHitTokens} tokens`].join(" · ");
+ ? [`${job.counts.succeeded}/${job.total} completed`, `${job.counts.partial || 0} partial`, `${job.counts.running} active`, `${job.counts.pending} queued`, `${job.counts.failed} failed`, `${job.requestAttempts || 0} actual requests`, `${job.outputCorrectionRequests || 0} output corrections`, `input ${job.usage.promptTokens} / output ${job.usage.completionTokens} tokens`, `${job.cacheHitCount || 0} result cache hits`, `prompt cache ${job.usage.cacheHitTokens} tokens`, analysisFailureCategorySummary(job, "en")].filter(Boolean).join(" · ")
+ : [`${job.counts.succeeded}/${job.total} 已完成`, `${job.counts.partial || 0} 待补全`, `${job.counts.running} 处理中`, `${job.counts.pending} 等待中`, `${job.counts.failed} 失败`, `真实请求 ${job.requestAttempts || 0} 次`, `结构补全 ${job.outputCorrectionRequests || 0} 次`, `输入 ${job.usage.promptTokens} / 输出 ${job.usage.completionTokens} tokens`, `结果缓存命中 ${job.cacheHitCount || 0} 次`, `提示缓存命中 ${job.usage.cacheHitTokens} tokens`, analysisFailureCategorySummary(job, "zh-CN")].filter(Boolean).join(" · ");
const recovery = unfinishedRebuild
? (currentLocale() === "en"
? `${job.stagedResultCount ?? job.counts.succeeded} successful results are safely staged. ${analysisFailureSummary(job, "en")}`
@@ -491,8 +491,10 @@ function renderAnalysisBatch() {
elements.cancelAnalysisBatch.hidden = !active;
elements.retryAnalysisFailures.textContent = unfinishedRebuild
? (currentLocale() === "en" ? `Finish rebuild (${job.counts.failed})` : `继续完成重建(${job.counts.failed} 条)`)
- : t("重试失败项");
- elements.retryAnalysisFailures.hidden = !job?.counts.failed || running || partialRebuildApplied || (unfinishedRebuild && job.stagingValid !== true);
+ : currentLocale() === "en"
+ ? `Retry failed/incomplete (${(job?.counts.failed || 0) + (job?.counts.partial || 0)} expected items)`
+ : `重试失败/待补全(预计新增 ${(job?.counts.failed || 0) + (job?.counts.partial || 0)} 项)`;
+ elements.retryAnalysisFailures.hidden = !(job?.counts.failed || job?.counts.partial) || running || partialRebuildApplied || (unfinishedRebuild && job.stagingValid !== true);
elements.applyStagedAnalysisRebuild.textContent = unfinishedRebuild
? (currentLocale() === "en"
? `Apply ${job.stagedResultCount ?? job.counts.succeeded} successful results (${job.counts.failed} pending)`
@@ -502,7 +504,7 @@ function renderAnalysisBatch() {
elements.undoAnalysisBatch.hidden = !job || !canUndoAnalysisBatch || running || paused;
elements.analysisProgress.hidden = !active;
elements.analysisProgressBar.max = Math.max(1, job?.total || 1);
- elements.analysisProgressBar.value = Math.min(job?.counts?.succeeded + job?.counts?.failed || 0, job?.total || 0);
+ elements.analysisProgressBar.value = Math.min((job?.counts?.succeeded || 0) + (job?.counts?.partial || 0) + (job?.counts?.failed || 0), job?.total || 0);
}
function analysisFailureSummary(job, locale) {
@@ -517,6 +519,16 @@ function analysisFailureSummary(job, locale) {
return locale === "en" ? `Failures: ${summaries.join("; ")}` : `失败原因:${summaries.join(";")}`;
}
+function analysisFailureCategorySummary(job, locale) {
+ const labels = locale === "en"
+ ? { authorization: "authorization", rate_limit: "rate limit", timeout: "timeout", service: "service", network: "network", output: "output" }
+ : { authorization: "授权", rate_limit: "限流", timeout: "超时", service: "服务", network: "网络", output: "输出格式" };
+ const values = Object.entries(job?.failureCategories ?? {}).filter(([, count]) => count > 0);
+ if (!values.length) return "";
+ const summary = values.map(([category, count]) => `${labels[category] || category} ${count}`).join(locale === "en" ? ", " : "、");
+ return locale === "en" ? `Failure types: ${summary}` : `失败类型:${summary}`;
+}
+
function bindEvents() {
elements.searchInput.addEventListener("compositionstart", () => { searchComposing = true; });
elements.searchInput.addEventListener("compositionend", () => { searchComposing = false; scheduleSearchRender(); });
@@ -2426,8 +2438,8 @@ function renderVisionBatchDialog(preview = null) {
const caseCount = preview?.caseCount ?? new Set(job?.items?.map((item) => item.entryId) ?? []).size;
elements.visionBatchSummary.textContent = job
? currentLocale() === "en"
- ? `${counts.succeeded ?? 0}/${requestCount} completed · ${counts.failed ?? 0} failed`
- : `${counts.succeeded ?? 0}/${requestCount} 已完成 · ${counts.failed ?? 0} 失败`
+ ? `${counts.succeeded ?? 0}/${requestCount} completed · ${counts.partial ?? 0} partial · ${counts.failed ?? 0} failed · ${job.requestAttempts || 0} actual requests · ${job.outputCorrectionRequests || 0} output corrections · ${job.cacheHitCount || 0} cache hits${analysisFailureCategorySummary(job, "en") ? ` · ${analysisFailureCategorySummary(job, "en")}` : ""}`
+ : `${counts.succeeded ?? 0}/${requestCount} 已完成 · ${counts.partial ?? 0} 待补全 · ${counts.failed ?? 0} 失败 · 真实请求 ${job.requestAttempts || 0} 次 · 结构补全 ${job.outputCorrectionRequests || 0} 次 · 缓存命中 ${job.cacheHitCount || 0} 次${analysisFailureCategorySummary(job, "zh-CN") ? ` · ${analysisFailureCategorySummary(job, "zh-CN")}` : ""}`
: currentLocale() === "en"
? `${caseCount} cases · ${requestCount} image requests`
: `${caseCount} 个案例 · ${requestCount} 次图片请求`;
@@ -2438,13 +2450,17 @@ function renderVisionBatchDialog(preview = null) {
elements.visionBatchStart.disabled = !requestCount;
elements.visionBatchPause.hidden = !job || job.status !== "running" || !visionBatchRunnerActive;
elements.visionBatchResume.hidden = !job || !active || visionBatchRunnerActive;
- elements.visionBatchRetry.hidden = !job || !counts.failed || visionBatchRunnerActive;
+ elements.visionBatchRetry.hidden = !job || !(counts.failed || counts.partial) || visionBatchRunnerActive;
+ elements.visionBatchRetry.textContent = currentLocale() === "en"
+ ? `Retry failed/incomplete (${(counts.failed || 0) + (counts.partial || 0)} expected items)`
+ : `重试失败/待补全(预计新增 ${(counts.failed || 0) + (counts.partial || 0)} 项)`;
elements.visionBatchCancel.hidden = !active;
if (job && !active) {
const total = job.usage?.totalTokens ?? 0;
- showVisionBatchFeedback(currentLocale() === "en"
- ? `Task finished · ${total} total tokens reported by the service`
- : `任务已结束 · 服务实际返回 ${total} tokens`);
+ const label = job.status === "completed" ? (currentLocale() === "en" ? "Completed" : "全部完成")
+ : job.status === "failed" ? (currentLocale() === "en" ? "Failed" : "全部失败")
+ : (currentLocale() === "en" ? "Partially completed" : "部分完成,可重试失败或待补全项");
+ showVisionBatchFeedback(`${label} · ${total} tokens` , job.status === "failed");
}
}
@@ -2482,37 +2498,50 @@ async function runVisionBatch() {
});
if (!claimed?.ok) throw new Error(claimed?.message || "无法领取下一张图片");
visionBatchJob = claimed.visionBatchJob;
- const claim = claimed.claim;
- if (!claim) break;
+ const claims = claimed.claims ?? (claimed.claim ? [claimed.claim] : []);
+ if (!claims.length) break;
showVisionBatchFeedback(currentLocale() === "en"
- ? `Analysing image ${visionBatchJob.counts.succeeded + visionBatchJob.counts.failed + 1} of ${visionBatchJob.requestCount}…`
- : `正在分析第 ${visionBatchJob.counts.succeeded + visionBatchJob.counts.failed + 1}/${visionBatchJob.requestCount} 张图片…`);
- const result = await chrome.runtime.sendMessage({
- type: "ANALYZE_ENTRY_IMAGE",
- entryId: claim.entryId,
- visualId: claim.visualId,
- outputLocale: visionBatchJob.outputLocale,
- batchJobId: visionBatchJob.id
- });
- const update = result?.ok
- ? await chrome.runtime.sendMessage({
- type: "COMPLETE_VISION_BATCH_ITEM",
- jobId: visionBatchJob.id,
- entryId: claim.entryId,
- visualId: claim.visualId,
- claimId: claim.claimId,
- usage: result.usage
- })
- : await chrome.runtime.sendMessage({
- type: "FAIL_VISION_BATCH_ITEM",
- jobId: visionBatchJob.id,
+ ? `Analysing ${claims.length} images concurrently…`
+ : `正在并发分析 ${claims.length} 张图片…`);
+ const settled = await Promise.allSettled(claims.map(async (claim) => {
+ const result = await chrome.runtime.sendMessage({
+ type: "ANALYZE_ENTRY_IMAGE",
entryId: claim.entryId,
visualId: claim.visualId,
- claimId: claim.claimId,
- error: { message: result?.message || "图片分析失败", status: 0 }
+ outputLocale: visionBatchJob.outputLocale,
+ batchJobId: visionBatchJob.id,
+ bypassCache: visionBatchJob.reanalyze === true,
+ assignment: {
+ providerId: visionBatchJob.providerId,
+ model: visionBatchJob.model,
+ concurrency: visionBatchJob.concurrency
+ }
});
- if (!update?.ok) throw new Error(update?.message || "无法保存批量任务进度");
- visionBatchJob = update.visionBatchJob;
+ const update = result?.ok
+ ? await chrome.runtime.sendMessage({
+ type: "COMPLETE_VISION_BATCH_ITEM",
+ jobId: visionBatchJob.id,
+ entryId: claim.entryId,
+ visualId: claim.visualId,
+ claimId: claim.claimId,
+ usage: result.usage,
+ cacheHit: result.cacheHit,
+ attempts: result.attempts,
+ quality: result.quality
+ })
+ : await chrome.runtime.sendMessage({
+ type: "FAIL_VISION_BATCH_ITEM",
+ jobId: visionBatchJob.id,
+ entryId: claim.entryId,
+ visualId: claim.visualId,
+ claimId: claim.claimId,
+ error: { message: result?.message || "图片分析失败", status: Number(result?.status) || 0 }
+ });
+ if (!update?.ok) throw new Error(update?.message || "无法保存批量任务进度");
+ visionBatchJob = update.visionBatchJob;
+ }));
+ const rejected = settled.find((item) => item.status === "rejected");
+ if (rejected) throw rejected.reason;
renderVisionBatchDialog();
if (cancelVisionBatchAfterCurrent) {
await updateVisionBatchAction("CANCEL_VISION_BATCH");
@@ -5188,6 +5217,7 @@ function createVisionDescription(entry) {
const section = el("section", "detail-section vision-description");
const heading = el("div", "vision-description-heading");
heading.append(textEl("h3", "", "画面描述"));
+ if (vision.quality === "partial") heading.append(textEl("span", "batch-status-badge", `部分完成 · ${vision.missingFields?.length || 0} 项待补全`));
const actions = el("div", "vision-description-actions");
const copy = textEl("button", "button-secondary", "复制");
copy.addEventListener("click", () => copyTextWithFeedback(
@@ -5228,7 +5258,13 @@ async function analyzeEntryVision(entry, button) {
button.textContent = t("正在分析画面…");
setVisionAnalysisStatus(entry.id, "loading", "正在发送当前截图,请保持页面打开。");
try {
- const response = await chrome.runtime.sendMessage({ type: "ANALYZE_ENTRY_IMAGE", entryId: entry.id, visualId: visual.id, outputLocale: currentLocale() });
+ const response = await chrome.runtime.sendMessage({
+ type: "ANALYZE_ENTRY_IMAGE",
+ entryId: entry.id,
+ visualId: visual.id,
+ outputLocale: currentLocale(),
+ bypassCache: Boolean(vision && vision.quality !== "partial")
+ });
if (!response?.ok) throw new Error(response?.message || "图片分析失败");
if (response.canUndoVisionAnalysis) visionUndoEntryIds.add(entry.id);
const replacement = visualAnalysisPromptReplacement(response.entry, visual.id);
@@ -6543,7 +6579,7 @@ function createPromptSection(entry, options = {}) {
const vision = primaryVisionAnalysis(entry);
const primary = primaryVisual(entry);
const analyzeVisual = primary?.kind === "image" && !entry.compoundCase
- ? textEl("button", "button-secondary", vision ? "重新分析主图" : "分析主图")
+ ? textEl("button", "button-secondary", vision?.quality === "partial" ? "补全主图分析" : vision ? "重新分析主图" : "分析主图")
: null;
const analyze = textEl("button", "button-secondary", "分析检索标签");
const edit = textEl("button", "button-secondary", separatesCurrentAndShared ? "编辑当前图片" : "编辑");
@@ -6625,7 +6661,8 @@ async function analyzeEntryVisualSet(entry, button) {
const fingerprintMatches = Boolean(currentFingerprint && analysis?.imageFingerprint
&& currentFingerprint === analysis.imageFingerprint);
if (analysis?.invalidated || !fingerprintMatches) return { valid: false, label: "分析已过期" };
- if (Number(analysis?.version) === 2 && analysis?.reconstructionPrompt) return { valid: true, label: "有效 V2,可直接复用" };
+ if (Number(analysis?.version) === 2 && analysis?.quality !== "partial" && analysis?.reconstructionPrompt) return { valid: true, label: "有效 V2,可直接复用" };
+ if (analysis?.quality === "partial") return { valid: false, label: "部分完成,待补全" };
return { valid: false, label: "未分析" };
};
const missingAssets = assets.filter((asset) => !stateFor(asset).valid);
@@ -6721,23 +6758,34 @@ async function analyzeEntryVisualSet(entry, button) {
}
const promptSuggestions = [];
const failures = [];
- for (const asset of selectedAssets) {
+ const partials = [];
+ await Promise.allSettled(selectedAssets.map(async (asset) => {
const imageIndex = assets.findIndex((item) => item.id === asset.id);
const card = cardByAssetId.get(asset.id);
const statusLabel = card?.querySelector(".visual-analysis-card-status");
if (card) card.dataset.state = "processing";
if (statusLabel) statusLabel.textContent = t("正在分析…");
try {
- const response = await chrome.runtime.sendMessage({ type: "ANALYZE_ENTRY_IMAGE", entryId: entry.id, visualId: asset.id, outputLocale: currentLocale() });
+ const response = await chrome.runtime.sendMessage({
+ type: "ANALYZE_ENTRY_IMAGE",
+ entryId: entry.id,
+ visualId: asset.id,
+ outputLocale: currentLocale(),
+ bypassCache: stateFor(asset).valid
+ });
if (!response?.ok) throw new Error(response?.message || "逐图分析失败");
- completedAssetIds.add(asset.id);
+ const partial = response.quality === "partial";
+ if (partial) partials.push({ index: imageIndex + 1, missingFields: response.missingFields ?? [] });
+ else completedAssetIds.add(asset.id);
const input = dialogControls?.get(`asset_${imageIndex}`);
- if (input) {
+ if (input && !partial) {
input.checked = false;
input.disabled = true;
}
- if (card) card.dataset.state = "completed";
- if (statusLabel) statusLabel.textContent = t("本次已完成");
+ if (card) card.dataset.state = partial ? "partial" : "completed";
+ if (statusLabel) statusLabel.textContent = partial
+ ? t("待补全 · 已保存有效结果")
+ : t("本次已完成");
updateSelectionSummary();
const replacement = visualAnalysisPromptReplacement(response.entry, asset.id);
if (replacement) savedPromptSuggestions.set(asset.id, { ...replacement, index: imageIndex + 1 });
@@ -6747,11 +6795,15 @@ async function analyzeEntryVisualSet(entry, button) {
if (card) card.dataset.state = "failed";
if (statusLabel) statusLabel.textContent = t("失败 · {message}", { message });
}
- }
- if (failures.length) {
+ }));
+ if (failures.length || partials.length) {
const statusLine = status();
statusLine.classList.add("error");
- statusLine.textContent = t("{failed} 张失败,已成功 {completed} 张且不会重复请求。可直接重试失败项。", { failed: failures.length, completed: completedAssetIds.size });
+ statusLine.textContent = t("{failed} 张失败,{partial} 张待补全,已成功 {completed} 张且不会重复请求。可直接重试失败或待补全项。", {
+ failed: failures.length,
+ partial: partials.length,
+ completed: completedAssetIds.size
+ });
await refreshLibrary();
return false;
}
@@ -7326,12 +7378,19 @@ async function openAiTaskAssignmentDialog(taskId) {
? current.providerId : capable[0].id;
const modelOptions = (providerId) => taskModelOptions(aiProviderRegistry.providers?.[providerId], taskId);
const initialModels = modelOptions(selectedProviderId);
+ const supportsBatchConcurrency = ["textTags", "imageAnalysis"].includes(taskId);
+ const defaultConcurrency = ["textTags", "skillExtraction", "creativePlanning"].includes(taskId) ? 20 : 10;
+ const initialConcurrency = Number.isInteger(Number(current.concurrency)) ? Number(current.concurrency) : defaultConcurrency;
const result = await showAppDialog({
title: t(task.label),
description: t("这里设置全局默认;创作台中的本轮切换不会改动此处。"),
fields: [
{ id: "providerId", label: t("AI 服务"), type: "select", value: selectedProviderId, options: capable.map((profile) => ({ value: profile.id, label: providerDisplayLabel(profile) })) },
- { id: "model", label: t("模型"), type: "select", value: initialModels.some((item) => item.value === current.model) ? current.model : initialModels[0]?.value || "", options: initialModels }
+ { id: "model", label: t("模型"), type: "select", value: initialModels.some((item) => item.value === current.model) ? current.model : initialModels[0]?.value || "", options: initialModels },
+ ...(supportsBatchConcurrency ? [
+ { id: "concurrency", label: t("批量分析并发数"), type: "number", min: 2, step: 1, value: initialConcurrency, help: t("文字任务默认 20,图片和视频任务默认 10;调整只对新任务生效。") },
+ { id: "highConcurrencyConfirmed", label: t("我确认高于 20 的并发可能明显增加瞬时费用、内存占用和错误数量"), type: "checkbox", value: false }
+ ] : [])
],
confirmLabel: t("保存任务默认"),
cancelLabel: t("取消"),
@@ -7343,13 +7402,26 @@ async function openAiTaskAssignmentDialog(taskId) {
localizeAiDialogClose(dialog);
const provider = controls.get("providerId");
const model = controls.get("model");
+ const concurrency = controls.get("concurrency");
+ const highConcurrencyConfirmed = controls.get("highConcurrencyConfirmed");
const modelField = model?.closest(".app-dialog-field");
const capabilityHelp = document.createElement("small");
capabilityHelp.className = "model-capability-help";
modelField?.append(capabilityHelp);
const syncCapabilityHelp = () => {
- capabilityHelp.textContent = modelAssignmentHelp(aiProviderRegistry.providers?.[provider.value], taskId, model.value);
+ const profile = aiProviderRegistry.providers?.[provider.value];
+ const descriptor = profile?.discoveredModels?.find((item) => item.id === model.value);
+ const officialLimit = Number(descriptor?.concurrencyLimit?.value);
+ const limitText = Number.isInteger(officialLimit)
+ ? t("官方并发上限:{count}", { count: officialLimit })
+ : t("服务商未提供固定并发上限");
+ capabilityHelp.textContent = [modelAssignmentHelp(profile, taskId, model.value), limitText].filter(Boolean).join(" · ");
capabilityHelp.hidden = !capabilityHelp.textContent;
+ if (concurrency) concurrency.max = Number.isInteger(officialLimit) ? String(officialLimit) : "";
+ if (highConcurrencyConfirmed) {
+ const wrapper = highConcurrencyConfirmed.closest(".app-dialog-field");
+ if (wrapper) wrapper.hidden = Number(concurrency?.value) <= 20;
+ }
};
provider?.addEventListener("change", () => {
const options = modelOptions(provider.value);
@@ -7365,13 +7437,17 @@ async function openAiTaskAssignmentDialog(taskId) {
syncCapabilityHelp();
});
model?.addEventListener("change", syncCapabilityHelp);
+ concurrency?.addEventListener("input", syncCapabilityHelp);
syncCapabilityHelp();
},
onSubmit: async (values) => {
if (!String(values.model ?? "").trim()) throw new Error(t("请选择或填写模型"));
+ const concurrency = supportsBatchConcurrency ? Number(values.concurrency) : initialConcurrency;
+ if (supportsBatchConcurrency && (!Number.isInteger(concurrency) || concurrency < 2)) throw new Error(t("批量分析并发数必须是大于或等于 2 的整数"));
+ if (supportsBatchConcurrency && concurrency > 20 && values.highConcurrencyConfirmed !== true) throw new Error(t("并发高于 20,请先确认费用和稳定性风险"));
const response = await chrome.runtime.sendMessage({
type: "UPDATE_AI_PROVIDER_CONFIGURATION",
- assignments: { ...aiTaskAssignments, [taskId]: { providerId: values.providerId, model: values.model } }
+ assignments: { ...aiTaskAssignments, [taskId]: { providerId: values.providerId, model: values.model, concurrency } }
});
if (!response?.ok) throw new Error(translateUiMessage(response?.message) || t("任务默认保存失败"));
applyAiConfigurationResponse(response);
@@ -8021,7 +8097,11 @@ async function runAnalysisBatch() {
const recovered = await chrome.runtime.sendMessage({ type: "RECOVER_ANALYSIS_BATCH" });
if (recovered?.analysisBatchJob) analysisBatchJob = recovered.analysisBatchJob;
recordAnalysisDiagnostic("job_recovered");
- const settingsValue = await privateAiSettings();
+ const settingsValue = await privateAiSettings({
+ providerId: analysisBatchJob.providerId,
+ model: analysisBatchJob.model || analysisBatchJob.analysisModel,
+ concurrency: analysisBatchJob.concurrency
+ });
const currentProfile = await analysisProfileFingerprint(settingsValue, analysisBatchJob.outputLocale);
if (analysisBatchJob.profileFingerprint && currentProfile !== analysisBatchJob.profileFingerprint) {
const paused = await chrome.runtime.sendMessage({ type: "PAUSE_ANALYSIS_BATCH", jobId: analysisBatchJob.id });
@@ -8114,21 +8194,23 @@ async function runAnalysisClaim(claim, catalog, settingsValue, signal) {
fingerprint: claim.fingerprint,
textRevision: claim.textRevision,
tags: result.tags,
+ normalizationDiagnostics: result.normalizationDiagnostics,
+ attempts: result.attempts,
usage: result.usage,
model: result.model
};
} catch (error) {
if (error?.name === "AbortError") throw error;
- return analysisClaimFailure(claim, error.message || "分析失败", error.status || 0, error.usage);
+ return analysisClaimFailure(claim, error.message || "分析失败", error.status || 0, error.usage, error.attempts);
}
}
-function analysisClaimFailure(claim, message, status, usage) {
+function analysisClaimFailure(claim, message, status, usage, attempts) {
return {
entryId: claim.entryId,
claimId: claim.claimId,
textRevision: claim.textRevision,
- error: { message, status, usage }
+ error: { message, status, usage, attempts }
};
}
@@ -8139,17 +8221,28 @@ async function analyzeEntryWithRetry(entry, catalog, settingsValue, outputLocale
: undefined;
for (;;) {
try {
- return await analyzeTextDetailedWithDeepSeek(entry, catalog, { ...settingsValue, outputLocale }, fetch, {
+ const result = await analyzeTextDetailedWithDeepSeek(entry, catalog, { ...settingsValue, outputLocale }, fetch, {
signal,
onDiagnostic: reportDiagnostic
});
+ return {
+ ...result,
+ attempts: {
+ serviceRequests: serviceRetries + 1,
+ outputCorrectionRequests: Number(result.attempts?.outputCorrectionRequests) || 0
+ }
+ };
} catch (error) {
if (isRetryableDeepSeekError(error) && serviceRetries < ANALYSIS_SERVICE_RETRY_LIMIT) {
- const delay = error.retryAfterMs || 1000 * (2 ** serviceRetries);
+ const delay = error.retryAfterMs || [1000, 3000][serviceRetries] + Math.floor(Math.random() * 501);
serviceRetries += 1;
await wait(Math.min(delay, 30_000), signal);
continue;
}
+ error.attempts = {
+ serviceRequests: serviceRetries + 1,
+ outputCorrectionRequests: Math.max(0, Number(error?.attempts?.outputCorrectionRequests) || 0)
+ };
throw error;
}
}
@@ -8168,6 +8261,8 @@ async function analyzeSingleEntry(entry, button) {
fingerprint,
textRevision: entryTextRevision(entry),
tags: result.tags,
+ normalizationDiagnostics: result.normalizationDiagnostics,
+ attempts: result.attempts,
usage: result.usage,
model: result.model,
profileFingerprint
@@ -8182,8 +8277,8 @@ async function analyzeSingleEntry(entry, button) {
}
}
-async function privateAiSettings() {
- const response = await chrome.runtime.sendMessage({ type: "GET_AI_TASK_RUNTIME", taskId: "textTags" });
+async function privateAiSettings(assignment = null) {
+ const response = await chrome.runtime.sendMessage({ type: "GET_AI_TASK_RUNTIME", taskId: "textTags", assignment });
if (!response?.ok) throw new Error(response?.message || "无法读取文字标签服务");
return normalizeAiSettings(response.aiSettings);
}
diff --git a/structured-output.js b/structured-output.js
new file mode 100644
index 0000000..82c3be8
--- /dev/null
+++ b/structured-output.js
@@ -0,0 +1,114 @@
+export function parseStructuredObject(value, message = "模型返回的结构化 JSON 无效") {
+ const parsed = parseStructuredValue(value, 0);
+ if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) throw new Error(message);
+ return parsed;
+}
+
+function parseStructuredValue(value, depth) {
+ if (depth > 2) throw new Error("模型返回的结构化 JSON 重复编码过多");
+ if (value && typeof value === "object") return value;
+ const source = String(value ?? "").replace(/^\uFEFF/, "").trim();
+ if (!source) throw new Error("模型没有返回结构化 JSON");
+ for (const candidate of jsonCandidates(source)) {
+ try {
+ const parsed = JSON.parse(removeTrailingCommas(candidate));
+ return typeof parsed === "string" ? parseStructuredValue(parsed, depth + 1) : parsed;
+ } catch {}
+ }
+ throw new Error("模型返回的结构化 JSON 无效");
+}
+
+function jsonCandidates(source) {
+ const candidates = [];
+ const unfenced = source.replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/i, "").trim();
+ candidates.push(unfenced);
+ const object = firstBalancedObject(unfenced);
+ if (object && object !== unfenced) candidates.push(object);
+ if (!object) candidates.push(...recoverableTruncatedObjectPrefixes(unfenced));
+ return [...new Set(candidates)];
+}
+
+function recoverableTruncatedObjectPrefixes(source) {
+ const start = source.indexOf("{");
+ if (start < 0) return [];
+ const object = source.slice(start);
+ const safeEnds = [];
+ const stack = [];
+ let quote = false;
+ let escaped = false;
+ for (let index = 0; index < object.length; index += 1) {
+ const character = object[index];
+ if (quote) {
+ if (escaped) escaped = false;
+ else if (character === "\\") escaped = true;
+ else if (character === '"') quote = false;
+ continue;
+ }
+ if (character === '"') {
+ quote = true;
+ continue;
+ }
+ if (character === "{" || character === "[") stack.push(character);
+ else if (character === "}" || character === "]") stack.pop();
+ else if (character === "," && stack.length === 1 && stack[0] === "{") safeEnds.push(index);
+ }
+ const candidates = [`${object}}`];
+ for (const end of safeEnds.reverse()) candidates.push(`${object.slice(0, end)}}`);
+ return candidates;
+}
+
+function firstBalancedObject(source) {
+ let start = -1;
+ let depth = 0;
+ let quote = false;
+ let escaped = false;
+ for (let index = 0; index < source.length; index += 1) {
+ const character = source[index];
+ if (quote) {
+ if (escaped) escaped = false;
+ else if (character === "\\") escaped = true;
+ else if (character === '"') quote = false;
+ continue;
+ }
+ if (character === '"') {
+ quote = true;
+ continue;
+ }
+ if (character === "{") {
+ if (depth === 0) start = index;
+ depth += 1;
+ } else if (character === "}" && depth) {
+ depth -= 1;
+ if (depth === 0) return source.slice(start, index + 1);
+ }
+ }
+ return "";
+}
+
+function removeTrailingCommas(source) {
+ let result = "";
+ let quote = false;
+ let escaped = false;
+ for (let index = 0; index < source.length; index += 1) {
+ const character = source[index];
+ if (quote) {
+ result += character;
+ if (escaped) escaped = false;
+ else if (character === "\\") escaped = true;
+ else if (character === '"') quote = false;
+ continue;
+ }
+ if (character === '"') {
+ quote = true;
+ result += character;
+ continue;
+ }
+ if (character === ",") {
+ let next = index + 1;
+ while (/\s/.test(source[next] ?? "")) next += 1;
+ if (["}", "]"].includes(source[next])) continue;
+ }
+ result += character;
+ }
+ return result;
+}
diff --git a/tag-taxonomy.js b/tag-taxonomy.js
index 4bddf02..2e309f5 100644
--- a/tag-taxonomy.js
+++ b/tag-taxonomy.js
@@ -162,12 +162,12 @@ export function validateAnalysisTagResponse(value, catalogValue, options = {}) {
if (!value || typeof value !== "object" || Array.isArray(value) || !Array.isArray(value.tags)) {
throw new Error("AI 返回格式无效,本次没有写入任何标签");
}
- if (Object.keys(value).some((key) => key !== "tags")) {
- throw new Error("AI 返回了未允许字段,本次没有写入");
- }
const minimum = options.allowEmpty === true ? 0 : ANALYSIS_TAG_MIN;
const maximum = Math.min(ANALYSIS_TAG_MAX, Math.max(minimum, Number(options.maxTags) || ANALYSIS_TAG_MAX));
- if (value.tags.length < minimum || value.tags.length > maximum) {
+ const diagnostics = Array.isArray(options.diagnostics) ? options.diagnostics : null;
+ const extraRootFields = Object.keys(value).filter((key) => key !== "tags").length;
+ if (extraRootFields) diagnostics?.push({ field: "root", code: "extra_fields_ignored", count: extraRootFields });
+ if (value.tags.length < minimum) {
throw new Error(`AI 必须返回 ${minimum}–${maximum} 个标签,本次没有写入`);
}
const catalog = catalogValue && Array.isArray(catalogValue.nodes) ? catalogValue : createFixedFacetCatalog();
@@ -175,20 +175,43 @@ export function validateAnalysisTagResponse(value, catalogValue, options = {}) {
.filter((item) => item.status !== "archived" && !item.parentId && item.kind !== "detail")
.map((item) => [item.id, item]));
const seen = new Set();
- const tags = value.tags.map((item) => {
- if (!item || typeof item !== "object" || Array.isArray(item)) throw new Error("AI 标签格式无效,本次没有写入");
- const keys = Object.keys(item);
- if (keys.some((key) => !["g", "t"].includes(key))) throw new Error("AI 返回了未允许字段,本次没有写入");
- const groupId = String(item.g ?? "").trim();
- if (!groups.has(groupId)) throw new Error("AI 返回了未知分类路径,本次没有写入");
- const hasDetail = Object.hasOwn(item, "t") && String(item.t ?? "").trim();
- const detail = hasDetail ? normalizeDetailLabel(item.t) : "";
- if (detail.length > ANALYSIS_DETAIL_MAX_LENGTH) throw new Error("AI 标签过长,本次没有写入");
+ const tags = [];
+ let unknownCount = 0;
+ for (let index = 0; index < value.tags.length; index += 1) {
+ const item = value.tags[index];
+ if (!item || typeof item !== "object" || Array.isArray(item)) {
+ diagnostics?.push({ field: "tags", code: "invalid_tag_dropped", count: 1 });
+ continue;
+ }
+ const extraFields = Object.keys(item).filter((key) => !["g", "groupId", "group_id", "group", "t", "detail", "tag", "label"].includes(key)).length;
+ if (extraFields) diagnostics?.push({ field: "tags", code: "extra_fields_ignored", count: extraFields });
+ const groupId = String(item.g ?? item.groupId ?? item.group_id ?? item.group ?? "").trim();
+ if (!groups.has(groupId)) {
+ unknownCount += 1;
+ diagnostics?.push({ field: "tags", code: "unknown_path_dropped", count: 1 });
+ continue;
+ }
+ const detailValue = item.t ?? item.detail ?? item.tag ?? item.label;
+ const detail = String(detailValue ?? "").trim()
+ ? [...normalizeDetailLabel(detailValue)].slice(0, ANALYSIS_DETAIL_MAX_LENGTH).join("")
+ : "";
const key = `${groupId}:${detailNormalizationKey(detail)}`;
- if (seen.has(key)) throw new Error("AI 返回了重复标签,本次没有写入");
+ if (seen.has(key)) {
+ diagnostics?.push({ field: "tags", code: "duplicate_dropped", count: 1 });
+ continue;
+ }
seen.add(key);
- return detail ? { g: groupId, t: detail } : { g: groupId };
- });
+ tags.push(detail ? { g: groupId, t: detail } : { g: groupId });
+ if (tags.length >= maximum) {
+ const capped = value.tags.length - index - 1;
+ if (capped) diagnostics?.push({ field: "tags", code: "over_limit_dropped", count: capped });
+ break;
+ }
+ }
+ if (tags.length < minimum || (!tags.length && value.tags.length)) {
+ if (unknownCount) throw new Error("AI 返回了未知分类路径,本次没有写入");
+ throw new Error(`AI 必须返回 ${minimum}–${maximum} 个标签,本次没有写入`);
+ }
return tags;
}
diff --git a/test/ai-provider-module.test.js b/test/ai-provider-module.test.js
index fd40916..dc936aa 100644
--- a/test/ai-provider-module.test.js
+++ b/test/ai-provider-module.test.js
@@ -139,7 +139,7 @@ test("Gemini discovery grants Nano Banana image generation only to exact officia
});
test("Nano Banana capability metadata exposes official Interactions parameters and per-model reference limits", () => {
- assert.deepEqual(AI_MODEL_CAPABILITIES.map((item) => item.id), [
+ assert.deepEqual(AI_MODEL_CAPABILITIES.filter((item) => item.providerId === "gemini").map((item) => item.id), [
"gemini-3.1-flash-lite-image",
"gemini-3.1-flash-image",
"gemini-3-pro-image",
diff --git a/test/ai-provider-registry.test.js b/test/ai-provider-registry.test.js
index 31408a1..1680560 100644
--- a/test/ai-provider-registry.test.js
+++ b/test/ai-provider-registry.test.js
@@ -13,6 +13,7 @@ import {
resolveAiProviderAssignment
} from "../ai-provider-registry.js";
import { AI_PROVIDER_PRESETS, getAiProviderPreset } from "../ai-provider-presets.js";
+import { getAiModelCapability } from "../ai-model-capabilities.js";
function configuredRegistry() {
return normalizeAiProviderRegistry({ providers: {
@@ -115,7 +116,8 @@ test("task model candidates separate declared catalog capability from explicit m
}, registry).imageAnalysis, {
providerId: "deepseek",
model: "opaque-account-model",
- evidence: "manual_unverified"
+ evidence: "manual_unverified",
+ concurrency: 10
});
assert.throws(() => createAiTaskAssignment(
"imageAnalysis", "deepseek", "declared-text", registry
@@ -161,12 +163,12 @@ test("a manually assigned catalog-visible DeepSeek model resolves for image anal
), /当前模型目录中没有/);
});
-test("registry v4 exposes provider categories and leaves all new-install tasks unassigned", () => {
+test("registry v5 exposes task concurrency defaults and official model limits", () => {
const registry = normalizeAiProviderRegistry({ providers: {
kimi: { discovery: { etag: "catalog-version" } }
} });
const assignments = normalizeAiTaskAssignments({}, registry);
- assert.equal(registry.version, 4);
+ assert.equal(registry.version, 5);
assert.equal(registry.providers.kimi.endpoint, "https://api.moonshot.cn/v1/chat/completions");
assert.deepEqual(registry.providers.kimi.discovery, {
adapter: "kimi",
@@ -177,9 +179,13 @@ test("registry v4 exposes provider categories and leaves all new-install tasks u
error: ""
});
assert.equal(publicAiProviderRegistry(registry).providers.kimi.category, "official");
- assert.deepEqual(assignments, Object.fromEntries(AI_ASSIGNMENT_TASKS.map(({ id }) => [id, {
- providerId: "", model: ""
- }])));
+ assert.equal(assignments.textTags.concurrency, 20);
+ assert.equal(assignments.imageAnalysis.concurrency, 10);
+ assert.equal(assignments.videoAnalysis.concurrency, 10);
+ const deepSeekVision = getAiModelCapability("deepseek", "deepseek-v4-flash-vision-exp");
+ assert.equal(deepSeekVision.inputModalities.includes("image"), true);
+ assert.equal(deepSeekVision.concurrencyLimit.value, 2500);
+ assert.match(deepSeekVision.concurrencyLimit.source.url, /deepseek\.com\/quick_start\/pricing/);
assert.throws(() => resolveAiProviderAssignment("textTags", registry, assignments), /未分配/);
const bounded = normalizeAiProviderRegistry({ providers: {
kimi: { discoveredModels: [{ id: "declared-model", tasks: ["videoAnalysis", "imageGeneration", "videoGeneration"] }] }
@@ -203,32 +209,32 @@ test("unknown saved provider ids do not become hidden v4 connections", () => {
assert.equal(Object.hasOwn(providers, "retired-provider"), false);
});
-test("empty v4 storage normalizes without inventing task assignments", () => {
+test("empty v5 storage normalizes without inventing provider or model assignments", () => {
const registry = normalizeAiProviderRegistry({});
const assignments = normalizeAiTaskAssignments({});
- assert.equal(registry.version, 4);
- assert.deepEqual(assignments, Object.fromEntries(AI_ASSIGNMENT_TASKS.map(({ id }) => [id, {
- providerId: "",
- model: ""
- }])));
+ assert.equal(registry.version, 5);
+ assert.equal(Object.values(assignments).every((assignment) => !assignment.providerId && !assignment.model), true);
+ assert.equal(assignments.textTags.concurrency, 20);
+ assert.equal(assignments.imageAnalysis.concurrency, 10);
});
-test("obsolete shared text routes are ignored instead of entering v4 assignments", () => {
+test("obsolete shared text routes are ignored instead of entering v5 assignments", () => {
const assignments = normalizeAiTaskAssignments({
text: { providerId: "deepseek", model: "obsolete-shared-model" }
});
- assert.deepEqual(assignments, Object.fromEntries(AI_ASSIGNMENT_TASKS.map(({ id }) => [id, {
- providerId: "",
- model: ""
- }])));
+ assert.equal(Object.values(assignments).every((assignment) => !assignment.providerId && !assignment.model), true);
});
-test("explicit task assignments survive v4 normalization unchanged", () => {
+test("explicit task assignments survive v5 normalization with task defaults", () => {
const explicit = Object.fromEntries(AI_ASSIGNMENT_TASKS.map(({ id }, index) => [id, {
providerId: index % 2 ? "openai" : "deepseek",
model: `account-model-${index}`
}]));
- assert.deepEqual(normalizeAiTaskAssignments(explicit), explicit);
+ const normalized = normalizeAiTaskAssignments(explicit);
+ assert.equal(AI_ASSIGNMENT_TASKS.every(({ id }) => normalized[id].providerId === explicit[id].providerId
+ && normalized[id].model === explicit[id].model), true);
+ assert.equal(normalized.textTags.concurrency, 20);
+ assert.equal(normalized.imageAnalysis.concurrency, 10);
assert.equal(configuredRegistry().providers.deepseek.apiKey, "deepseek-secret");
});
diff --git a/test/ai-runtime.test.js b/test/ai-runtime.test.js
index 763eba6..839bc3c 100644
--- a/test/ai-runtime.test.js
+++ b/test/ai-runtime.test.js
@@ -66,7 +66,7 @@ test("obsolete AI storage keys are ignored instead of becoming a second configur
});
assert.equal(configuration.registry.providers.deepseek.apiKey, "");
- assert.deepEqual(configuration.assignments.textTags, { providerId: "", model: "" });
+ assert.deepEqual(configuration.assignments.textTags, { providerId: "", model: "", concurrency: 20 });
assert.equal(JSON.stringify(configuration).includes("obsolete-key"), false);
});
diff --git a/test/analysis-batch.test.js b/test/analysis-batch.test.js
index 160780b..dd66be6 100644
--- a/test/analysis-batch.test.js
+++ b/test/analysis-batch.test.js
@@ -11,6 +11,7 @@ import {
createAnalysisBatchJob,
finalizeAnalysisRebuild,
failAnalysisItem,
+ partiallySucceedAnalysisItem,
previewAnalysisBatch,
finalizePartialAnalysisRebuild,
resumeAnalysisBatch,
@@ -196,7 +197,7 @@ test("recovering an old stalled rebuild preserves completed work and queues only
const recovered = resumeAnalysisBatch(job);
const summary = analysisBatchSummary(recovered);
- assert.deepEqual(summary.counts, { pending: 5, running: 0, succeeded: 586, failed: 0 });
+ assert.deepEqual(summary.counts, { pending: 5, running: 0, succeeded: 586, partial: 0, failed: 0 });
assert.equal(summary.status, "running");
});
@@ -291,7 +292,7 @@ test("batch success totals usage and failures can be retried", async () => {
job = succeedAnalysisItem(claimed.job, "a", "claim-1", { totalTokens: 12, cacheHitTokens: 7 }, 4);
job = failAnalysisItem(job, "b", "claim-2", { message: "busy", status: 503 });
let summary = analysisBatchSummary(job);
- assert.equal(summary.status, "completed");
+ assert.equal(summary.status, "partial");
assert.equal(summary.counts.succeeded, 1);
assert.equal(summary.counts.failed, 1);
assert.equal(summary.usage.totalTokens, 12);
@@ -326,6 +327,76 @@ test("authentication failures pause the whole batch", async () => {
assert.equal(job.items[0].status, "failed");
});
+test("settled batch status distinguishes complete, partial, and failed outcomes", async () => {
+ let mixed = await createAnalysisBatchJob([{ id: "a", text: "one" }, { id: "b", text: "two" }], { id: "mixed" });
+ let claims = claimAnalysisItems(mixed, 2, (() => { let index = 0; return () => `mixed-${++index}`; })());
+ mixed = succeedAnalysisItem(claims.job, "a", "mixed-1", { totalTokens: 1 });
+ mixed = failAnalysisItem(mixed, "b", "mixed-2", { message: "bad json", status: 422 });
+ assert.equal(mixed.status, "partial");
+
+ let failed = await createAnalysisBatchJob([{ id: "a", text: "one" }], { id: "failed" });
+ claims = claimAnalysisItems(failed, 1, () => "failed-1");
+ failed = failAnalysisItem(claims.job, "a", "failed-1", { message: "bad json", status: 422 });
+ assert.equal(failed.status, "failed");
+
+ let complete = await createAnalysisBatchJob([{ id: "a", text: "one" }], { id: "complete" });
+ claims = claimAnalysisItems(complete, 1, () => "complete-1");
+ complete = succeedAnalysisItem(claims.job, "a", "complete-1", { totalTokens: 1 });
+ assert.equal(complete.status, "completed");
+});
+
+test("a saved partial item remains retryable and only the incomplete item is requeued", async () => {
+ let job = await createAnalysisBatchJob([{ id: "a", text: "one" }, { id: "b", text: "two" }], { id: "saved-partial" });
+ const claims = claimAnalysisItems(job, 2, (() => { let index = 0; return () => `partial-${++index}`; })());
+ job = succeedAnalysisItem(claims.job, "a", "partial-1", { totalTokens: 5 });
+ job = partiallySucceedAnalysisItem(job, "b", "partial-2", { totalTokens: 7 });
+ assert.equal(job.status, "partial");
+ assert.deepEqual(analysisBatchSummary(job).counts, { pending: 0, running: 0, succeeded: 1, partial: 1, failed: 0 });
+
+ job = retryFailedAnalysisItems(job);
+ assert.equal(job.items[0].status, "succeeded");
+ assert.equal(job.items[1].status, "pending");
+ assert.equal(job.status, "running");
+});
+
+test("batch summary exposes actual requests, output corrections, cache hits, and failure categories", async () => {
+ let job = await createAnalysisBatchJob([{ id: "a", text: "one" }, { id: "b", text: "two" }], { id: "execution-diagnostics" });
+ const claims = claimAnalysisItems(job, 2, (() => { let index = 0; return () => `diag-${++index}`; })());
+ job = succeedAnalysisItem(claims.job, "a", "diag-1", {}, undefined, {
+ attempts: { serviceRequests: 0, outputCorrectionRequests: 0 }, cacheHit: true
+ });
+ job = failAnalysisItem(job, "b", "diag-2", {
+ message: "rate limited", status: 429, attempts: { serviceRequests: 3, outputCorrectionRequests: 1 }
+ });
+ const summary = analysisBatchSummary(job);
+ assert.equal(summary.requestAttempts, 3);
+ assert.equal(summary.outputCorrectionRequests, 1);
+ assert.equal(summary.cacheHitCount, 1);
+ assert.deepEqual(summary.failureCategories, { rate_limit: 1 });
+});
+
+test("batch jobs snapshot configured concurrency and vision claims use it", () => {
+ const job = createVisionBatchJob([{
+ id: "one",
+ primaryMediaId: "image-1",
+ mediaAssets: Array.from({ length: 12 }, (_, index) => ({ id: `image-${index + 1}`, kind: "image", usage: "content" }))
+ }], {
+ id: "vision-concurrency",
+ entryIds: ["one"],
+ includeAllImages: true,
+ concurrency: 10
+ });
+ assert.equal(job.concurrency, 10);
+ assert.deepEqual(job.retryPolicy, {
+ serviceRetries: 2,
+ outputCorrectionRequests: 1,
+ backoffMs: [1000, 3000],
+ obeyRetryAfter: true
+ });
+ assert.equal(job.outputProtocol, "json_object");
+ assert.equal(claimAnalysisItems(job).claims.length, 10);
+});
+
test("vision batch defaults to one primary image per selected case and can include every unanalysed image", async () => {
const entries = [{
id: "one",
diff --git a/test/analysis-candidates.test.js b/test/analysis-candidates.test.js
index 454664c..46e125a 100644
--- a/test/analysis-candidates.test.js
+++ b/test/analysis-candidates.test.js
@@ -506,7 +506,7 @@ test("screenshot replacement invalidates vision references and undo restores the
assert.deepEqual(restored.facetAssignments.map((item) => item.source).sort(), ["manual", "vision_model"]);
});
-test("vision analysis stores zero to six tags and rejects an over-limit result as a whole", () => {
+test("vision analysis stores zero to six tags and safely caps an over-limit result", () => {
for (const returned of [0, 3, 6]) {
const applied = applyVisionAnalysis({
facetCatalog: createDefaultFacetCatalog(),
@@ -518,11 +518,13 @@ test("vision analysis stores zero to six tags and rejects an over-limit result a
assert.equal(applied.appliedCount, returned);
assert.equal(applied.state.entries[0].facetAssignments.length, returned);
}
- assert.throws(() => applyVisionAnalysis({
+ const capped = applyVisionAnalysis({
facetCatalog: createDefaultFacetCatalog(), entries: [{ id: "too-many", facetAssignments: [] }]
}, "too-many", {
description: "description", tags: Array.from({ length: 7 }, (_, index) => ({ g: "style.render", t: `标签${index}` }))
- }), /0–6/);
+ });
+ assert.equal(capped.appliedCount, 6);
+ assert.equal(capped.state.entries[0].facetAssignments.length, 6);
});
test("image analysis can retain a reusable breakdown while promoting only six results", () => {
diff --git a/test/analysis-scheduler.test.js b/test/analysis-scheduler.test.js
new file mode 100644
index 0000000..1710642
--- /dev/null
+++ b/test/analysis-scheduler.test.js
@@ -0,0 +1,52 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { coalesceAnalysisRequest, runScheduledAnalysisWithRetries, scheduleAnalysis } from "../analysis-scheduler.js";
+
+const delay = (milliseconds) => new Promise((resolve) => setTimeout(resolve, milliseconds));
+
+test("analysis scheduler enforces the shared provider-model-task limit", async () => {
+ let active = 0;
+ let maximum = 0;
+ await Promise.all(Array.from({ length: 8 }, (_, index) => scheduleAnalysis("deepseek:model:imageAnalysis", 3, async () => {
+ active += 1;
+ maximum = Math.max(maximum, active);
+ await delay(4);
+ active -= 1;
+ return index;
+ })));
+ assert.equal(maximum, 3);
+});
+
+test("analysis scheduler coalesces identical in-flight work and retries transient failures twice", async () => {
+ let calls = 0;
+ let coalesced = 0;
+ const work = () => coalesceAnalysisRequest("same-fingerprint", () => runScheduledAnalysisWithRetries({
+ key: "deepseek:model:imageAnalysis",
+ concurrency: 10,
+ wait: async () => undefined,
+ jitter: () => 0,
+ task: async () => {
+ calls += 1;
+ if (calls < 3) throw Object.assign(new Error("busy"), { status: 503 });
+ return "ok";
+ }
+ }), { onCoalesced: () => { coalesced += 1; } });
+ assert.deepEqual(await Promise.all([work(), work()]), ["ok", "ok"]);
+ assert.equal(calls, 3);
+ assert.equal(coalesced, 1);
+});
+
+test("analysis scheduler does not mistake local JSON validation errors for network failures", async () => {
+ let calls = 0;
+ await assert.rejects(() => runScheduledAnalysisWithRetries({
+ key: "deepseek:model:imageAnalysis:local-output-error",
+ concurrency: 10,
+ wait: async () => undefined,
+ task: async () => {
+ calls += 1;
+ throw new Error("模型 JSON 无效");
+ }
+ }), /JSON 无效/);
+ assert.equal(calls, 1);
+});
diff --git a/test/structured-output.test.js b/test/structured-output.test.js
new file mode 100644
index 0000000..e87defd
--- /dev/null
+++ b/test/structured-output.test.js
@@ -0,0 +1,33 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { parseStructuredObject } from "../structured-output.js";
+
+test("structured output accepts fenced JSON, surrounding prose, a trailing comma, and double encoding", () => {
+ assert.deepEqual(parseStructuredObject("\uFEFF```json\n{\"tags\":[{\"g\":\"style.render\",}],}\n```"), {
+ tags: [{ g: "style.render" }]
+ });
+ assert.deepEqual(parseStructuredObject("模型结果如下: {\"tags\":[{\"g\":\"scene.place\"}]} 谢谢"), {
+ tags: [{ g: "scene.place" }]
+ });
+ assert.deepEqual(parseStructuredObject(JSON.stringify('{"tags":[{"g":"light.direction"}]}')), {
+ tags: [{ g: "light.direction" }]
+ });
+});
+
+test("structured output never evaluates executable text or invents a missing object", () => {
+ globalThis.__structuredOutputExecuted = false;
+ assert.throws(() => parseStructuredObject("globalThis.__structuredOutputExecuted = true"), /JSON|结构化/);
+ assert.equal(globalThis.__structuredOutputExecuted, false);
+ delete globalThis.__structuredOutputExecuted;
+});
+
+test("structured output recovers only complete top-level fields from a truncated object", () => {
+ const recovered = parseStructuredObject('{"description":"visible","tags":[{"g":"style.render","t":"CG"}],"canvas":{"width":');
+ assert.deepEqual(recovered, {
+ description: "visible",
+ tags: [{ g: "style.render", t: "CG" }]
+ });
+ assert.deepEqual(parseStructuredObject('{"description":"visible"'), { description: "visible" });
+ assert.throws(() => parseStructuredObject('{"description":"vis'), /JSON/);
+});
diff --git a/test/tag-taxonomy.test.js b/test/tag-taxonomy.test.js
index 10c55a6..113abbf 100644
--- a/test/tag-taxonomy.test.js
+++ b/test/tag-taxonomy.test.js
@@ -39,17 +39,23 @@ test("each protected other group cannot be renamed, archived, or merged away", (
}), /必须保留/);
});
-test("analysis response rejects empty, unknown, duplicate, and over-limit results as a whole", () => {
+test("analysis response keeps valid tags while dropping unknown, duplicate, extra, and over-limit output", () => {
const catalog = createFixedFacetCatalog();
+ const diagnostics = [];
assert.throws(() => validateAnalysisTagResponse({ tags: [] }, catalog), /1–10/);
assert.throws(() => validateAnalysisTagResponse({ tags: [{ g: "unknown", t: "值" }] }, catalog), /未知分类路径/);
- assert.throws(() => validateAnalysisTagResponse({ tags: [
+ assert.deepEqual(validateAnalysisTagResponse({ extra: "ignored", tags: [
{ g: "style.render", t: "Cel-Shading" },
- { g: "style.render", t: "cel shading" }
- ] }, catalog), /重复标签/);
- assert.throws(() => validateAnalysisTagResponse({
- tags: Array.from({ length: ANALYSIS_TAG_MAX + 1 }, (_, index) => ({ g: "style.render", t: `标签${index}` }))
- }, catalog), /1–10/);
+ { g: "style.render", t: "cel shading" },
+ { g: "unknown", t: "discarded" }
+ ] }, catalog, { diagnostics }), [{ g: "style.render", t: "Cel-Shading" }]);
+ assert.ok(diagnostics.some((item) => item.code === "extra_fields_ignored"));
+ assert.ok(diagnostics.some((item) => item.code === "duplicate_dropped"));
+ assert.ok(diagnostics.some((item) => item.code === "unknown_path_dropped"));
+ const bounded = validateAnalysisTagResponse({
+ tags: Array.from({ length: ANALYSIS_TAG_MAX + 3 }, (_, index) => ({ g: "style.render", t: `标签${index}` }))
+ }, catalog);
+ assert.equal(bounded.length, ANALYSIS_TAG_MAX);
});
test("analysis can assign a group directly and reuses normalized detail labels", () => {
diff --git a/test/vision-analysis-cache.test.js b/test/vision-analysis-cache.test.js
new file mode 100644
index 0000000..f44976a
--- /dev/null
+++ b/test/vision-analysis-cache.test.js
@@ -0,0 +1,53 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { findPersistedVisionAnalysis } from "../vision-analysis-cache.js";
+
+function entryWithAnalysis(id, quality, overrides = {}) {
+ return {
+ id,
+ primaryMediaId: `${id}-image`,
+ mediaAssets: [{
+ id: `${id}-image`,
+ kind: "image",
+ storageMode: "managed",
+ mimeType: "image/png",
+ byteSize: 10,
+ visionAnalysis: {
+ version: 2,
+ quality,
+ description: `${quality} result`,
+ imageFingerprint: "same-image",
+ profileFingerprint: "same-profile",
+ locale: "zh-CN",
+ catalogRevision: 7,
+ ...overrides
+ }
+ }]
+ };
+}
+
+test("persisted vision cache prefers a complete matching result and can use a partial result as completion input", () => {
+ const options = { fingerprint: "same-image", profileFingerprint: "same-profile", locale: "zh-CN", catalogRevision: 7 };
+ const partial = findPersistedVisionAnalysis([entryWithAnalysis("partial", "partial")], options);
+ assert.equal(partial.quality, "partial");
+
+ const complete = findPersistedVisionAnalysis([
+ entryWithAnalysis("partial", "partial"),
+ entryWithAnalysis("complete", "complete")
+ ], options);
+ assert.equal(complete.quality, "complete");
+});
+
+test("persisted vision cache rejects stale model, locale, taxonomy, and protocol matches", () => {
+ const entry = entryWithAnalysis("stale", "complete");
+ assert.equal(findPersistedVisionAnalysis([entry], {
+ fingerprint: "same-image", profileFingerprint: "different-profile", locale: "zh-CN", catalogRevision: 7
+ }), null);
+ assert.equal(findPersistedVisionAnalysis([entry], {
+ fingerprint: "same-image", profileFingerprint: "same-profile", locale: "en", catalogRevision: 7
+ }), null);
+ assert.equal(findPersistedVisionAnalysis([entry], {
+ fingerprint: "same-image", profileFingerprint: "same-profile", locale: "zh-CN", catalogRevision: 8
+ }), null);
+});
diff --git a/test/vision.test.js b/test/vision.test.js
index 2c61bbb..ee5b1a3 100644
--- a/test/vision.test.js
+++ b/test/vision.test.js
@@ -210,7 +210,7 @@ test("localhost compatible services allow HTTP without an API key and list model
assert.deepEqual(probe.models, ["local-vision", "text-only"]);
});
-test("vision result requires a complete reconstruction record and keeps at most six optional compact tags", () => {
+test("vision result keeps partial paid output visible and keeps at most six optional compact tags", () => {
const values = Array.from({ length: 6 }, (_, index) => ({ g: "light.direction", t: `标签${index + 1}` }));
const result = normalizeVisionResult(completeVisionResult(" 主体处于逆光中。 ", values));
assert.equal(result.description, "主体处于逆光中。");
@@ -219,8 +219,27 @@ test("vision result requires a complete reconstruction record and keeps at most
const overLimit = normalizeVisionResult(completeVisionResult("有效描述", [...values, values[0]]));
assert.equal(overLimit.tags.length, 6);
assert.equal(overLimit.tagDiagnostics.rejectedCount, 1);
- assert.throws(() => normalizeVisionResult(completeVisionResult("")), /画面描述/);
- assert.throws(() => normalizeVisionResult({ description: "旧版简述", tags: [] }), /画布信息/);
+ const missingDescription = normalizeVisionResult(completeVisionResult(""));
+ assert.equal(missingDescription.quality, "partial");
+ assert.ok(missingDescription.missingFields.includes("description"));
+ const partial = normalizeVisionResult({ description: "旧版简述", tags: [] });
+ assert.equal(partial.quality, "partial");
+ assert.equal(partial.description, "旧版简述");
+ assert.ok(partial.missingFields.includes("canvas"));
+ assert.ok(partial.missingFields.includes("reconstructionPrompt"));
+});
+
+test("partial vision normalization accepts root casing and snake-case aliases without inventing fields", () => {
+ const partial = normalizeVisionResult({
+ Description: "可见人物站在窗边。",
+ reconstruction_prompt: "人物站在窗边,侧逆光。",
+ Tags: [{ group_id: "light.direction", detail: "侧逆光" }],
+ extra_model_note: "ignored"
+ });
+ assert.equal(partial.quality, "partial");
+ assert.equal(partial.reconstructionPrompt, "人物站在窗边,侧逆光。");
+ assert.deepEqual(partial.tags, [{ g: "light.direction", t: "侧逆光" }]);
+ assert.ok(partial.missingFields.includes("elements"));
});
test("an invalid optional search tag cannot discard an otherwise complete paid visual analysis", () => {
@@ -369,13 +388,15 @@ test("compatible json_object vision requests keep image input and local full-sch
assert.match(instruction, /"ocr":\{"type":"array"/);
assert.equal(success.description, "一张完整分析。");
- await assert.rejects(() => analyzeImageWithVision(input, async () => ({
+ const partial = await analyzeImageWithVision(input, async () => ({
ok: true,
json: async () => ({ choices: [{ message: { content: JSON.stringify({ description: "不完整" }) } }] })
- })), /画布信息|缺少|必须/);
+ }));
+ assert.equal(partial.quality, "partial");
+ assert.ok(partial.missingFields.includes("canvas"));
});
-test("compatible empty vision results expose finish reasons without retrying or falling back", async () => {
+test("compatible empty vision results get exactly one structured-output correction request", async () => {
const input = {
imageDataUrl: "data:image/png;base64,AAAA",
catalog: createDefaultFacetCatalog(),
@@ -408,7 +429,7 @@ test("compatible empty vision results expose finish reasons without retrying or
};
}), expected);
}
- assert.equal(calls, cases.length);
+ assert.equal(calls, cases.length * 2);
});
test("compatible Responses request sends the OpenAI image shape and parses structured output", async () => {
@@ -542,14 +563,21 @@ test("vision service refusal, invalid JSON, missing description and HTTP errors
ok: true,
json: async () => ({ output: [{ content: [{ type: "refusal", refusal: "not allowed" }] }] })
})), /拒绝分析/);
- await assert.rejects(() => analyzeImageWithVision(input, async () => ({
- ok: true,
- json: async () => ({ output_text: "not-json" })
- })), /JSON 无效/);
- await assert.rejects(() => analyzeImageWithVision(input, async () => ({
+ let invalidJsonCalls = 0;
+ await assert.rejects(() => analyzeImageWithVision(input, async () => {
+ invalidJsonCalls += 1;
+ return {
+ ok: true,
+ json: async () => ({ output_text: "not-json" })
+ };
+ }), /JSON 无效/);
+ assert.equal(invalidJsonCalls, 2);
+ const partial = await analyzeImageWithVision(input, async () => ({
ok: true,
json: async () => ({ output_text: JSON.stringify(completeVisionResult("")) })
- })), /没有返回画面描述/);
+ }));
+ assert.equal(partial.quality, "partial");
+ assert.ok(partial.missingFields.includes("description"));
await assert.rejects(() => analyzeImageWithVision(input, async () => ({
ok: false,
status: 429,
diff --git a/test/visual-analysis.test.js b/test/visual-analysis.test.js
index 246b6af..ea8bf14 100644
--- a/test/visual-analysis.test.js
+++ b/test/visual-analysis.test.js
@@ -5,6 +5,7 @@ import {
VISUAL_ANALYSIS_DIMENSIONS,
VISUAL_MODEL_RESPONSE_SCHEMA,
compileVisualAnalysisInstruction,
+ mergePartialVisualAnalysis,
normalizeVisualAnalysisV2,
normalizeVisualModelResponse,
prepareVisualSetSummary
@@ -111,22 +112,42 @@ test("known root field casing is normalized before strict visual validation", ()
assert.equal(Object.hasOwn(result, "OCR"), false);
});
-test("root field casing conflicts and unknown fields remain rejected", () => {
+test("root field casing conflicts remain rejected while unknown fields are ignored", () => {
const conflict = completeModelResponse();
conflict.OCR = conflict.ocr;
assert.throws(() => normalizeVisualModelResponse(conflict), /字段大小写冲突:ocr/);
const unknown = completeModelResponse({ audit: [] });
- assert.throws(() => normalizeVisualModelResponse(unknown), /未允许字段:audit/);
+ assert.equal(normalizeVisualModelResponse(unknown).quality, "complete");
});
-test("invalid model boxes fail without clipping or rewriting the saved analysis", () => {
+test("invalid model boxes produce a visible partial result without clipping or rewriting input", () => {
const invalid = completeModelResponse();
invalid.elements[0].box_2d = [150, 250, 140, 900];
- assert.throws(() => normalizeVisualModelResponse(invalid), /画面位置关系不完整/);
+ const result = normalizeVisualModelResponse(invalid);
+ assert.equal(result.quality, "partial");
+ assert.ok(result.missingFields.includes("elements"));
assert.deepEqual(invalid.elements[0].box_2d, [150, 250, 140, 900]);
});
+test("completion merges only missing fields and preserves fields already validated from a paid partial result", () => {
+ const previous = normalizeVisualModelResponse({
+ description: "已经保存的可靠描述",
+ reconstructionPrompt: "已经保存的可靠重建提示词",
+ tags: []
+ });
+ const incoming = normalizeVisualModelResponse(completeModelResponse({
+ description: "补全请求中的重复描述",
+ reconstructionPrompt: "补全请求中的重复提示词"
+ }));
+ const merged = mergePartialVisualAnalysis(previous, incoming);
+ assert.equal(merged.quality, "complete");
+ assert.deepEqual(merged.missingFields, []);
+ assert.equal(merged.description, "已经保存的可靠描述");
+ assert.equal(merged.reconstructionPrompt, "已经保存的可靠重建提示词");
+ assert.deepEqual(merged.canvas, incoming.canvas);
+});
+
test("visual instruction requests quantitative reconstruction in one call and exposes only fixed taxonomy paths", () => {
const catalog = createDefaultFacetCatalog();
catalog.nodes.push({
diff --git a/ui-dialogs.js b/ui-dialogs.js
index 6a539fd..74b617c 100644
--- a/ui-dialogs.js
+++ b/ui-dialogs.js
@@ -208,6 +208,9 @@ function createField(field = {}) {
if (field.required === true) input.required = true;
if (clean(field.pattern)) input.pattern = clean(field.pattern);
if (Number.isFinite(Number(field.minLength))) input.minLength = Number(field.minLength);
+ if (Number.isFinite(Number(field.min))) input.min = String(Number(field.min));
+ if (Number.isFinite(Number(field.max))) input.max = String(Number(field.max));
+ if (Number.isFinite(Number(field.step))) input.step = String(Number(field.step));
if (field.disabled === true) input.disabled = true;
if (field.type === "checkbox") wrapper.append(input, label);
else wrapper.append(label, input);
diff --git a/vision-analysis-cache.js b/vision-analysis-cache.js
new file mode 100644
index 0000000..2ab81c1
--- /dev/null
+++ b/vision-analysis-cache.js
@@ -0,0 +1,22 @@
+import { VISION_ANALYSIS_VERSION } from "./vision.js";
+import { normalizeEntryVisuals } from "./visuals.js";
+
+export function findPersistedVisionAnalysis(entries, options = {}) {
+ const fingerprint = String(options.fingerprint ?? "").trim();
+ const profileFingerprint = String(options.profileFingerprint ?? "").trim();
+ const locale = options.locale === "en" ? "en" : "zh-CN";
+ const catalogRevision = Number(options.catalogRevision);
+ if (!fingerprint || !profileFingerprint || !Number.isFinite(catalogRevision)) return null;
+ let partial = null;
+ for (const entry of Array.isArray(entries) ? entries : []) {
+ for (const visual of normalizeEntryVisuals(entry).visuals) {
+ const analysis = visual.visionAnalysis;
+ if (!analysis || Number(analysis.version) !== VISION_ANALYSIS_VERSION) continue;
+ if (analysis.imageFingerprint !== fingerprint || analysis.profileFingerprint !== profileFingerprint) continue;
+ if (analysis.locale !== locale || Number(analysis.catalogRevision) !== catalogRevision) continue;
+ if (analysis.quality === "complete") return structuredClone(analysis);
+ if (analysis.quality === "partial") partial = structuredClone(analysis);
+ }
+ }
+ return partial;
+}
diff --git a/vision.js b/vision.js
index 6250945..4f4d6b5 100644
--- a/vision.js
+++ b/vision.js
@@ -3,9 +3,11 @@ import {
VISUAL_MODEL_PROTOCOL_VERSION,
VISUAL_MODEL_RESPONSE_SCHEMA,
compileVisualAnalysisInstruction,
+ mergePartialVisualAnalysis,
normalizeVisualModelResponse,
visualModelResponseSchema
} from "./visual-analysis.js";
+import { parseStructuredObject } from "./structured-output.js";
import { MICU_COMPATIBLE_PROVIDER_PRESET } from "./compatible-provider-presets.js";
export const VISION_ANALYSIS_VERSION = VISUAL_ANALYSIS_VERSION;
@@ -21,6 +23,21 @@ export const MICU_DEFAULT_IMAGE_SIZE = MICU_COMPATIBLE_PROVIDER_PRESET.imageGene
export const MICU_IMAGE_RESULT_PERMISSION = "https://oss.filenest.top/*";
export const MAX_VISION_TAGS = 6;
export const VISION_MAX_OUTPUT_TOKENS = 12000;
+export class VisionApiError extends Error {
+ constructor(message, status = 0, options = {}) {
+ super(message, options);
+ this.name = "VisionApiError";
+ this.status = Number(status) || 0;
+ this.retryAfterMs = Number(options.retryAfterMs) || 0;
+ }
+}
+class VisionOutputError extends Error {
+ constructor(message, options = {}) {
+ super(message, options);
+ this.name = "VisionOutputError";
+ this.usage = options.usage;
+ }
+}
export const VISION_REQUEST_TIMEOUT_MS = 120_000;
export const DEFAULT_VISION_INSTRUCTIONS_BY_LOCALE = Object.freeze({
@@ -302,28 +319,58 @@ export async function analyzeImageWithVision(input = {}, fetchImpl = fetch) {
const settings = requireVisionSettings(input.settings);
const imageDataUrl = requireImageDataUrl(input.imageDataUrl);
const locale = input.locale === "en" ? "en" : "zh-CN";
- const instruction = compileVisualAnalysisInstruction({
+ const previousAnalysis = input.previousAnalysis?.quality === "partial" ? input.previousAnalysis : null;
+ const baseInstruction = compileVisualAnalysisInstruction({
catalog: input.catalog,
customInstruction: settings.instructionsByLocale[locale],
locale,
measuredCanvas: input.measuredCanvas
});
+ const instruction = previousAnalysis
+ ? `${baseInstruction}\nThis is a completion request. Focus on repairing these missing or invalid root fields: ${JSON.stringify(previousAnalysis.missingFields ?? [])}. Previously validated fields are preserved locally and must not be weakened.`
+ : baseInstruction;
const requestOptions = {
maxOutputTokens: settings.maxOutputTokens,
schema: visualModelResponseSchema(input.catalog)
};
- const response = settings.nativeProvider?.id === "gemini"
- ? await requestGeminiVision(settings.nativeProvider, imageDataUrl, instruction, fetchImpl, requestOptions)
+ const request = async (requestInstruction) => settings.nativeProvider?.id === "gemini"
+ ? requestGeminiVision(settings.nativeProvider, imageDataUrl, requestInstruction, fetchImpl, requestOptions)
: settings.activeProvider === "compatible"
- ? await requestCompatible(settings, imageDataUrl, instruction, fetchImpl, requestOptions)
- : await requestOpenAI(settings, imageDataUrl, instruction, fetchImpl, requestOptions);
- const parsed = parseJson(response.content);
+ ? requestCompatible(settings, imageDataUrl, requestInstruction, fetchImpl, requestOptions)
+ : requestOpenAI(settings, imageDataUrl, requestInstruction, fetchImpl, requestOptions);
+ const perform = async (requestInstruction) => {
+ const response = await request(requestInstruction);
+ try {
+ return { response, normalized: normalizeVisionResult(parseJson(response.content), input.catalog) };
+ } catch (error) {
+ throw new VisionOutputError(error.message, { cause: error, usage: response.usage });
+ }
+ };
+ let firstUsage;
+ let outputCorrectionRequests = 0;
+ let completed;
+ try {
+ completed = await perform(instruction);
+ } catch (error) {
+ if (!(error instanceof VisionOutputError)) throw error;
+ firstUsage = error.usage;
+ outputCorrectionRequests = 1;
+ try {
+ completed = await perform(`${instruction}\nThe previous response was empty or structurally unusable. Return one corrected JSON object now; do not add commentary or markdown.`);
+ } catch (correctionError) {
+ correctionError.attempts = { outputCorrectionRequests };
+ correctionError.usage = addVisionUsage(firstUsage, correctionError.usage);
+ throw correctionError;
+ }
+ }
+ const { response, normalized } = completed;
return {
- ...normalizeVisionResult(parsed, input.catalog),
+ ...mergePartialVisualAnalysis(previousAnalysis, normalized),
profileFingerprint: await visionAnalysisProfileFingerprint(settings, locale),
providerType: settings.nativeProvider?.id || settings.activeProvider,
model: response.model,
- usage: response.usage
+ attempts: { outputCorrectionRequests },
+ usage: addVisionUsage(firstUsage, response.usage)
};
}
@@ -423,7 +470,7 @@ async function requestResponses(provider, imageDataUrl, instruction, fetchImpl,
const refusal = extractOpenAIRefusal(payload);
if (refusal) throw new Error(`${provider.serviceName} 拒绝分析这张图片:${refusal}`);
const content = String(payload.output_text ?? extractOpenAIText(payload) ?? "").trim();
- if (!content) throw new Error(options.emptyResultMessage || `${provider.serviceName} 没有返回可用的画面描述`);
+ if (!content) throw new VisionOutputError(options.emptyResultMessage || `${provider.serviceName} 没有返回可用的画面描述`);
return { content, model: String(payload.model ?? provider.model), usage: normalizeOpenAIUsage(payload.usage) };
}
@@ -460,7 +507,7 @@ async function requestCompatible(settings, imageDataUrl, instruction, fetchImpl,
const choice = payload?.choices?.[0];
const content = choice?.message?.content;
if (typeof content !== "string" || !content.trim()) {
- throw new Error(compatibleEmptyResultMessage(choice?.finish_reason, options.emptyResultMessage));
+ throw new VisionOutputError(compatibleEmptyResultMessage(choice?.finish_reason, options.emptyResultMessage));
}
return { content, model: String(payload.model ?? settings.compatible.model), usage: normalizeCompatibleUsage(payload.usage) };
}
@@ -502,15 +549,19 @@ async function requestGeminiVision(provider, imageDataUrl, instruction, fetchImp
signal: controller.signal
});
} catch (error) {
- if (controller.signal.aborted) throw new Error("Gemini 图片分析超过 2 分钟未完成", { cause: error });
- throw new Error("无法连接 Gemini 图片分析服务", { cause: error });
+ if (controller.signal.aborted) throw new VisionApiError("Gemini 图片分析超过 2 分钟未完成", 408, { cause: error });
+ throw new VisionApiError("无法连接 Gemini 图片分析服务", 0, { cause: error });
} finally {
clearTimeout(timeoutId);
}
const payload = await response.json().catch(() => ({}));
- if (!response.ok) throw new Error(`Gemini 图片分析失败:${String(payload?.error?.message ?? `HTTP ${response.status}`)}`);
+ if (!response.ok) throw new VisionApiError(
+ `Gemini 图片分析失败:${String(payload?.error?.message ?? `HTTP ${response.status}`)}`,
+ response.status,
+ { retryAfterMs: retryAfterMilliseconds(response.headers?.get?.("retry-after")) }
+ );
const content = (payload?.candidates?.[0]?.content?.parts ?? []).map((part) => part?.text || "").join("").trim();
- if (!content) throw new Error(options.emptyResultMessage || "Gemini 没有返回可用的画面描述");
+ if (!content) throw new VisionOutputError(options.emptyResultMessage || "Gemini 没有返回可用的画面描述");
return {
content,
model: provider.model,
@@ -535,15 +586,17 @@ async function requestJson(url, apiKey, body, fetchImpl) {
signal: controller.signal
});
} catch (error) {
- if (controller.signal.aborted) throw new Error("图片分析超过 2 分钟未完成,本次已停止,请手动重试", { cause: error });
- throw new Error("无法连接图片分析服务,请检查网络、地址和权限", { cause: error });
+ if (controller.signal.aborted) throw new VisionApiError("图片分析超过 2 分钟未完成,本次已停止,请手动重试", 408, { cause: error });
+ throw new VisionApiError("无法连接图片分析服务,请检查网络、地址和权限", 0, { cause: error });
} finally {
clearTimeout(timeoutId);
}
const payload = await response.json().catch(() => ({}));
if (!response.ok) {
const detail = String(payload?.error?.message ?? payload?.message ?? "").trim();
- throw new Error(`图片分析失败${detail ? `:${detail}` : `(HTTP ${response.status})`}`);
+ throw new VisionApiError(`图片分析失败${detail ? `:${detail}` : `(HTTP ${response.status})`}`, response.status, {
+ retryAfterMs: retryAfterMilliseconds(response.headers?.get?.("retry-after"))
+ });
}
return payload;
}
@@ -658,9 +711,8 @@ function normalizeCreativeEvaluationResult(value) {
}
function parseJson(content) {
- const cleaned = String(content ?? "").trim().replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/, "");
try {
- return JSON.parse(cleaned);
+ return parseStructuredObject(content, "视觉模型返回的 JSON 无效,本次没有写入");
} catch {
throw new Error("视觉模型返回的 JSON 无效,本次没有写入");
}
@@ -700,6 +752,20 @@ function normalizeCompatibleUsage(value = {}) {
};
}
+function addVisionUsage(left = {}, right = {}) {
+ return Object.fromEntries(["inputTokens", "outputTokens", "totalTokens"].map((key) => [
+ key,
+ finite(left?.[key]) + finite(right?.[key])
+ ]));
+}
+
+function retryAfterMilliseconds(value) {
+ const seconds = Number(value);
+ if (Number.isFinite(seconds) && seconds >= 0) return seconds * 1000;
+ const date = Date.parse(String(value ?? ""));
+ return Number.isFinite(date) ? Math.max(0, date - Date.now()) : 0;
+}
+
function finite(value) {
const number = Number(value);
return Number.isFinite(number) && number > 0 ? Math.round(number) : 0;
diff --git a/visual-analysis.js b/visual-analysis.js
index c19ad36..b5822f9 100644
--- a/visual-analysis.js
+++ b/visual-analysis.js
@@ -25,6 +25,10 @@ const ROOT_FIELDS = new Set([
"limitations", "completeness", "tags"
]);
const ROOT_FIELDS_BY_CASE = new Map([...ROOT_FIELDS].map((field) => [field.toLocaleLowerCase("en-US"), field]));
+const ROOT_FIELD_ALIASES = new Map([...ROOT_FIELDS].flatMap((field) => [
+ [field.toLocaleLowerCase("en-US"), field],
+ [field.replace(/[A-Z]/g, (character) => `_${character.toLocaleLowerCase("en-US")}`), field]
+]));
const BBOX_SCHEMA = Object.freeze({
type: "object",
@@ -198,20 +202,79 @@ export function normalizeVisualModelResponse(value, catalogValue) {
const tagDiagnostics = partitionModelTags(source.tags, catalogValue);
const converted = {
...source,
- elements: (Array.isArray(source.elements) ? source.elements : []).map((item) => convertModelBox(item)),
- ocr: (Array.isArray(source.ocr) ? source.ocr : []).map((item) => convertModelBox(item)),
+ elements: (Array.isArray(source.elements) ? source.elements : []).map(convertModelBoxIfValid),
+ ocr: (Array.isArray(source.ocr) ? source.ocr : []).map(convertModelBoxIfValid),
tags: tagDiagnostics.accepted
};
- return {
- ...normalizeVisualAnalysisV2(converted, catalogValue),
- tagDiagnostics: { rejectedCount: tagDiagnostics.rejectedCount }
+ try {
+ return {
+ ...normalizeVisualAnalysisV2(converted, catalogValue),
+ quality: "complete",
+ missingFields: [],
+ normalizationDiagnostics: tagDiagnostics.rejections,
+ tagDiagnostics: { rejectedCount: tagDiagnostics.rejectedCount, rejections: tagDiagnostics.rejections }
+ };
+ } catch (error) {
+ return normalizePartialVisualAnalysis(converted, catalogValue, tagDiagnostics, error);
+ }
+}
+
+export function mergePartialVisualAnalysis(previousValue, incomingValue) {
+ const previous = previousValue && typeof previousValue === "object" ? previousValue : {};
+ const incoming = incomingValue && typeof incomingValue === "object" ? incomingValue : {};
+ if (previous.quality !== "partial") return incoming;
+ const previousMissing = new Set(Array.isArray(previous.missingFields) ? previous.missingFields : []);
+ const incomingMissing = new Set(Array.isArray(incoming.missingFields) ? incoming.missingFields : []);
+ const fields = ["description", "canvas", "elements", "dimensions", "ocr", "reconstructionPrompt", "completeness"];
+ const merged = {};
+ const missingFields = [];
+ for (const field of fields) {
+ const previousValid = !previousMissing.has(field) && hasVisualField(previous, field);
+ const incomingValid = !incomingMissing.has(field) && hasVisualField(incoming, field);
+ if (previousValid) merged[field] = structuredClone(previous[field]);
+ else if (incomingValid) merged[field] = structuredClone(incoming[field]);
+ else {
+ if (hasVisualField(incoming, field)) merged[field] = structuredClone(incoming[field]);
+ else if (hasVisualField(previous, field)) merged[field] = structuredClone(previous[field]);
+ missingFields.push(field);
+ }
+ }
+ merged.limitations = [...new Set([...(previous.limitations ?? []), ...(incoming.limitations ?? [])].map(String).filter(Boolean))];
+ const tagKeys = new Set();
+ merged.tags = [...(incoming.tags ?? []), ...(previous.tags ?? [])].filter((tag) => {
+ const key = `${tag?.g ?? ""}\n${tag?.t ?? ""}`;
+ if (!tag?.g || !tag?.t || tagKeys.has(key) || tagKeys.size >= 6) return false;
+ tagKeys.add(key);
+ return true;
+ }).map((tag) => ({ g: tag.g, t: tag.t }));
+ const unresolved = new Set(missingFields);
+ merged.normalizationDiagnostics = [
+ ...(incoming.normalizationDiagnostics ?? []),
+ ...(previous.normalizationDiagnostics ?? [])
+ ].filter((item, index, all) => (item?.field === "tags" || unresolved.has(item?.field))
+ && all.findIndex((other) => other?.field === item?.field && other?.code === item?.code && other?.message === item?.message) === index);
+ merged.quality = missingFields.length ? "partial" : "complete";
+ merged.missingFields = missingFields;
+ merged.tagDiagnostics = {
+ rejectedCount: Number(previous.tagDiagnostics?.rejectedCount || 0) + Number(incoming.tagDiagnostics?.rejectedCount || 0),
+ rejections: [...(previous.tagDiagnostics?.rejections ?? []), ...(incoming.tagDiagnostics?.rejections ?? [])]
};
+ return merged;
+}
+
+function hasVisualField(value, field) {
+ const fieldValue = value?.[field];
+ if (Array.isArray(fieldValue)) return true;
+ if (fieldValue && typeof fieldValue === "object") return true;
+ return typeof fieldValue === "string" && Boolean(fieldValue.trim());
}
function normalizeKnownRootFieldCasing(value) {
const normalized = {};
for (const [field, fieldValue] of Object.entries(value)) {
- const canonical = ROOT_FIELDS_BY_CASE.get(field.toLocaleLowerCase("en-US")) || field;
+ const normalizedKey = field.toLocaleLowerCase("en-US").replace(/[\s-]+/g, "_");
+ const canonical = ROOT_FIELD_ALIASES.get(normalizedKey) || ROOT_FIELDS_BY_CASE.get(normalizedKey);
+ if (!canonical) continue;
if (Object.hasOwn(normalized, canonical)) {
throw new Error(`视觉分析字段大小写冲突:${canonical}`);
}
@@ -220,6 +283,71 @@ function normalizeKnownRootFieldCasing(value) {
return normalized;
}
+function normalizePartialVisualAnalysis(value, catalogValue, tagDiagnostics, fullError) {
+ const missingFields = [];
+ const normalizationDiagnostics = [];
+ const result = {};
+ const record = (field, normalize) => {
+ try {
+ const normalized = normalize();
+ if (normalized === undefined || normalized === null || normalized === "") throw new Error(`${field} 为空`);
+ result[field] = normalized;
+ } catch (error) {
+ missingFields.push(field);
+ normalizationDiagnostics.push({ field, message: String(error?.message || `${field} 无效`) });
+ }
+ };
+ record("description", () => requiredText(value.description, "视觉模型没有返回画面描述"));
+ record("canvas", () => normalizeCanvas(value.canvas));
+ record("elements", () => normalizePartialElements(value.elements));
+ record("dimensions", () => normalizePartialDimensions(value.dimensions));
+ record("ocr", () => normalizePartialOcr(value.ocr));
+ record("reconstructionPrompt", () => requiredText(value.reconstructionPrompt, "视觉模型没有返回可独立使用的重建提示词"));
+ record("completeness", () => normalizeCompleteness(value.completeness));
+ result.limitations = cleanStrings(value.limitations);
+ result.tags = validateAnalysisTagResponse({ tags: tagDiagnostics.accepted }, catalogValue, { allowEmpty: true, maxTags: 6 });
+ normalizationDiagnostics.push(...tagDiagnostics.rejections);
+ if (!result.description && !result.reconstructionPrompt) throw fullError;
+ if (!result.description) result.description = result.reconstructionPrompt;
+ return {
+ ...result,
+ quality: "partial",
+ missingFields,
+ normalizationDiagnostics,
+ tagDiagnostics: { rejectedCount: tagDiagnostics.rejectedCount, rejections: tagDiagnostics.rejections }
+ };
+}
+
+function normalizePartialElements(values) {
+ const normalized = (Array.isArray(values) ? values : []).flatMap((item) => {
+ try { return [normalizeElements([convertModelBoxIfValid(item)])[0]]; }
+ catch { return []; }
+ });
+ if (!normalized.length) throw new Error("视觉分析没有可用的画面元素");
+ return normalized;
+}
+
+function normalizePartialDimensions(values) {
+ const byId = new Map((Array.isArray(values) ? values : []).map((item) => [String(item?.id ?? "").trim(), item]));
+ const normalized = VISUAL_ANALYSIS_DIMENSIONS.flatMap((id) => {
+ const item = byId.get(id);
+ if (!item || typeof item.applicable !== "boolean") return [];
+ const facts = cleanStrings(item.facts);
+ if (!item.applicable && facts.length) return [];
+ return [{ id, applicable: item.applicable, facts, measurements: cleanStrings(item.measurements) }];
+ });
+ if (!normalized.length) throw new Error("视觉分析没有可用的视觉维度");
+ return normalized;
+}
+
+function normalizePartialOcr(values) {
+ if (!Array.isArray(values)) throw new Error("视觉分析缺少 OCR 清单");
+ return values.flatMap((item) => {
+ try { return [normalizeOcr([convertModelBoxIfValid(item)])[0]]; }
+ catch { return []; }
+ });
+}
+
export function normalizeVisualAnalysisV2(value, catalogValue) {
if (!value || typeof value !== "object" || Array.isArray(value)) {
throw new Error("视觉模型返回格式不正确,本次没有写入");
@@ -251,7 +379,7 @@ export function prepareVisualSetSummary(values, locale = "zh-CN") {
for (const item of Array.isArray(values) ? values : []) {
const assetId = String(item?.assetId ?? "").trim();
if (!assetId) continue;
- if (!item?.analysis || Number(item.analysis.version ?? VISUAL_ANALYSIS_VERSION) !== VISUAL_ANALYSIS_VERSION) {
+ if (!item?.analysis || item.analysis.quality === "partial" || Number(item.analysis.version ?? VISUAL_ANALYSIS_VERSION) !== VISUAL_ANALYSIS_VERSION) {
missingAssetIds.push(assetId);
continue;
}
@@ -423,23 +551,28 @@ function partitionModelTags(values, catalog) {
const accepted = [];
const seen = new Set();
let rejectedCount = Array.isArray(values) ? 0 : 1;
+ const rejections = Array.isArray(values) ? [] : [{ field: "tags", code: "invalid_collection", count: 1 }];
for (const value of source) {
let tag;
try {
- [tag] = validateAnalysisTagResponse({ tags: [value] }, catalog, { allowEmpty: true, maxTags: 1 });
+ const diagnostics = [];
+ [tag] = validateAnalysisTagResponse({ tags: [value] }, catalog, { allowEmpty: true, maxTags: 1, diagnostics });
+ rejections.push(...diagnostics);
} catch {
rejectedCount += 1;
+ rejections.push({ field: "tags", code: "invalid_or_unknown_tag_dropped", count: 1 });
continue;
}
const key = JSON.stringify(tag);
if (seen.has(key) || accepted.length >= 6) {
rejectedCount += 1;
+ rejections.push({ field: "tags", code: seen.has(key) ? "duplicate_dropped" : "over_limit_dropped", count: 1 });
continue;
}
seen.add(key);
accepted.push(tag);
}
- return { accepted, rejectedCount };
+ return { accepted, rejectedCount, rejections };
}
function convertModelBox(item) {
@@ -466,6 +599,11 @@ function convertModelBox(item) {
};
}
+function convertModelBoxIfValid(item) {
+ try { return convertModelBox(item); }
+ catch { return item; }
+}
+
function assertOnlyFields(value, allowed, label) {
const unexpected = Object.keys(value).filter((key) => !allowed.has(key));
if (unexpected.length) throw new Error(`${label}包含未允许字段:${unexpected.join("、")}`);