Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(search): 补全 Cindy AI 网页搜索结果摘要并修正部分失败判定

Dash1f8fbdb8eb07
diff --git a/apps/desktop/src/main/mcp-integrations/__tests__/cindyProxySearch.test.ts b/apps/desktop/src/main/mcp-integrations/__tests__/cindyProxySearch.test.ts
index 06273b7c2f..98d2159538 100644
--- a/apps/desktop/src/main/mcp-integrations/__tests__/cindyProxySearch.test.ts
+++ b/apps/desktop/src/main/mcp-integrations/__tests__/cindyProxySearch.test.ts
@@ -110,6 +110,7 @@ describe('cindyProxySearch', () => {
expect(JSON.parse(String(init.body))).toEqual({
model: CINDY_SEARCH_MODEL_NAME,
max_tokens: 2048,
+ system: expect.stringContaining('output one line for each of the first 3 results'),
messages: [{ role: 'user', content: 'Cindy' }],
tools: [
{
@@ -143,6 +144,123 @@ describe('cindyProxySearch', () => {
});
});
+ it('结果缺 cited_text 时按模型输出的 `<url> | 摘要` 行回填,并清理链接与标题', async () => {
+ const service = createCindyProxySearchService({
+ getBaseUrl: () => 'https://gateway.example.test',
+ getApiKey: () => 'test-key',
+ fetchImpl: vi.fn(async () =>
+ response({
+ content: [
+ {
+ type: 'web_search_tool_result',
+ content: [
+ {
+ type: 'web_search_result',
+ title: 'Release',
+ url: 'https://example.test/release#1',
+ },
+ {
+ type: 'web_search_result',
+ title: '<p align="center">',
+ url: 'http://example.test/readme/',
+ },
+ { type: 'web_search_result', title: 'Channel', url: 'https://example.test/ch%0A' },
+ { type: 'web_search_result', title: 'Cited', url: 'https://example.test/cited' },
+ { type: 'web_search_result', title: 'Beyond limit', url: 'https://example.test/x' },
+ ],
+ },
+ {
+ type: 'text',
+ text: [
+ 'https://example.test/release#1 | Release notes for v1.',
+ '- https://example.test/readme | The README describes Cindy.',
+ '1. <https://example.test/ch> | Channel news | with a pipe.',
+ 'https://example.test/cited | Model note that must not replace cited_text.',
+ 'https://example.test/invented | A URL the search never returned.',
+ 'Unformatted closing remark.',
+ ].join('\n'),
+ citations: [
+ {
+ type: 'web_search_result_location',
+ url: 'https://example.test/cited',
+ cited_text: 'Quoted from the page',
+ },
+ ],
+ },
+ ],
+ }),
+ ) as unknown as typeof fetch,
+ });
+
+ await expect(service.search({ query: 'Cindy', limit: 4 })).resolves.toEqual({
+ ok: true,
+ results: [
+ {
+ title: 'Release',
+ url: 'https://example.test/release',
+ snippet: 'Release notes for v1.',
+ },
+ {
+ title: 'example.test',
+ url: 'http://example.test/readme/',
+ snippet: 'The README describes Cindy.',
+ },
+ {
+ title: 'Channel',
+ url: 'https://example.test/ch',
+ snippet: 'Channel news | with a pipe.',
+ },
+ { title: 'Cited', url: 'https://example.test/cited', snippet: 'Quoted from the page' },
+ ],
+ });
+ });
+
+ it('已有结果时,后续搜索的工具级错误不使整次搜索失败', async () => {
+ const result = { type: 'web_search_result', title: 'Kept', url: 'https://example.test/kept' };
+ for (const errorBlock of [
+ // Anthropic 标准形态:错误对象直接作为 content。
+ { type: 'web_search_tool_result_error', error_code: 'max_uses_exceeded' },
+ // 实测 Cindy 网关形态:错误项装在数组里。
+ [{ type: 'web_search_tool_result_error', error_code: 'max_uses_exceeded' }],
+ ]) {
+ const service = createCindyProxySearchService({
+ getBaseUrl: () => 'https://gateway.example.test',
+ getApiKey: () => 'test-key',
+ fetchImpl: vi.fn(async () =>
+ response({
+ content: [
+ { type: 'web_search_tool_result', content: [result] },
+ { type: 'web_search_tool_result', content: errorBlock },
+ ],
+ }),
+ ) as unknown as typeof fetch,
+ });
+ await expect(service.search({ query: 'Cindy', limit: 5 })).resolves.toMatchObject({
+ ok: true,
+ results: [{ title: 'Kept', url: 'https://example.test/kept', snippet: '' }],
+ });
+ }
+
+ const onlyError = createCindyProxySearchService({
+ getBaseUrl: () => 'https://gateway.example.test',
+ getApiKey: () => 'test-key',
+ fetchImpl: vi.fn(async () =>
+ response({
+ content: [
+ {
+ type: 'web_search_tool_result',
+ content: [{ type: 'web_search_tool_result_error', error_code: 'unavailable' }],
+ },
+ ],
+ }),
+ ) as unknown as typeof fetch,
+ });
+ await expect(onlyError.search({ query: 'Cindy', limit: 5 })).resolves.toMatchObject({
+ ok: false,
+ errorCode: 'UPSTREAM_UNAVAILABLE',
+ });
+ });
+
it('未配置 endpoint/key 时 fail closed,且不发请求', async () => {
const fetchImpl = vi.fn() as unknown as typeof fetch;
for (const [baseUrl, apiKey] of [
@@ -189,7 +307,12 @@ describe('cindyProxySearch', () => {
});
it('网关 401 时上报本次请求携带的 Key,其他状态不上报', async () => {
- for (const [status, expectedReports] of [[401, 1], [403, 0], [429, 0], [200, 0]] as const) {
+ for (const [status, expectedReports] of [
+ [401, 1],
+ [403, 0],
+ [429, 0],
+ [200, 0],
+ ] as const) {
const onKeyRejected = vi.fn();
const service = createCindyProxySearchService({
getBaseUrl: () => 'https://gateway.example.test',
diff --git a/apps/desktop/src/main/mcp-integrations/cindyProxySearch.ts b/apps/desktop/src/main/mcp-integrations/cindyProxySearch.ts
index 9dbb14ea32..500f0df776 100644
--- a/apps/desktop/src/main/mcp-integrations/cindyProxySearch.ts
+++ b/apps/desktop/src/main/mcp-integrations/cindyProxySearch.ts
@@ -13,6 +13,24 @@ export const CINDY_SEARCH_MODEL_NAME = 'cindy/web-search';
const CINDY_SEARCH_WEB_TOOL_TYPE = 'web_search_20250305';
const CINDY_SEARCH_MAX_TOKENS = 2_048;
const CINDY_SEARCH_TIMEOUT_MS = 30_000;
+const CINDY_SEARCH_MAX_SNIPPET_CHARS = 500;
+
+/**
+ * 搜索后端的结果只带标题、链接与加密正文,没有 cited_text / citations,摘要会全空;
+ * 后端模型也会无视 max_uses 再搜一次。让模型只搜一次,并为前 limit 条结果各写一行
+ * `<url> | 摘要`,由 parseSearchResponse 按链接回填 snippet。
+ */
+function buildSearchSystemPrompt(limit: number): string {
+ return [
+ 'You are a web search backend. Call web_search exactly once with the user query, then stop searching.',
+ `After the search, output one line for each of the first ${limit} results, in result order, formatted exactly as:`,
+ '<url> | <one or two sentences stating what that page says about the query>',
+ 'Copy each URL verbatim from the search result.',
+ // 只写「用查询的语言」时,后端模型会对英文查询输出中文摘要;写明映射后才稳定。
+ 'Write each summary in the same language as the user query (English query -> English, Chinese query -> Chinese).',
+ 'Output nothing else.',
+ ].join('\n');
+}
/**
* Join the Anthropic-compatible Messages endpoint for either a gateway root or
@@ -238,12 +256,25 @@ function normalizedHttpUrl(value: unknown): string | null {
try {
const url = new URL(value.trim());
if (url.protocol !== 'http:' && url.protocol !== 'https:') return null;
+ // 后端会给链接追加 `#1` 这类序号锚点、把换行编码进路径末尾,同一页面因此去重失败。
+ if (/^#\d+$/.test(url.hash)) url.hash = '';
+ url.pathname = url.pathname.replace(/(?:%0A|%0D|%09|%20)+$/i, '');
return url.toString();
} catch {
return null;
}
}
+/** 模型复述链接时可能改写协议或末尾斜杠,回填摘要按忽略这些差异的键匹配。 */
+function sourceMatchKey(url: string): string {
+ try {
+ const parsed = new URL(url);
+ return `${parsed.hostname.toLowerCase()}${parsed.pathname.replace(/\/+$/, '')}${parsed.search}`;
+ } catch {
+ return url;
+ }
+}
+
function fallbackTitle(url: string): string {
try {
return new URL(url).hostname;
@@ -252,6 +283,33 @@ function fallbackTitle(url: string): string {
}
}
+function cleanTitle(value: unknown, url: string): string {
+ // 抓取到的标题偶尔是页面里的 HTML 片段(如 `<p align="center">`)。
+ const title =
+ typeof value === 'string'
+ ? value
+ .replace(/<[^>]*>/g, ' ')
+ .replace(/\s+/g, ' ')
+ .trim()
+ : '';
+ return title || fallbackTitle(url);
+}
+
+const SNIPPET_LINE_PATTERN =
+ /^\s*(?:[-*>]|\d+[.)]|\[\d+\])?\s*<?(https?:\/\/[^\s|<>]+)>?\s*\|\s*(.+?)\s*$/;
+
+/** 解析 buildSearchSystemPrompt 要求的 `<url> | 摘要` 行;只回填已有结果,不据此新增来源。 */
+function collectSnippetLines(text: string, into: Map<string, string>): void {
+ for (const line of text.split('\n')) {
+ const match = SNIPPET_LINE_PATTERN.exec(line);
+ if (!match) continue;
+ const url = normalizedHttpUrl(match[1]);
+ if (!url) continue;
+ const key = sourceMatchKey(url);
+ if (!into.has(key)) into.set(key, match[2].slice(0, CINDY_SEARCH_MAX_SNIPPET_CHARS));
+ }
+}
+
interface SearchSource {
url: string;
title: string;
@@ -272,10 +330,7 @@ type ParsedSearchResponse =
function sourceFromRecord(value: Record<string, unknown>): SearchSource | null {
const url = normalizedHttpUrl(value.url);
if (!url) return null;
- const title =
- typeof value.title === 'string' && value.title.trim().length > 0
- ? value.title.trim()
- : fallbackTitle(url);
+ const title = cleanTitle(value.title, url);
const snippet =
typeof value.cited_text === 'string'
? value.cited_text.trim()
@@ -344,6 +399,11 @@ function parseSearchResponse(raw: unknown, limit: number): ParsedSearchResponse
if (!existing.snippet && source.snippet) existing.snippet = source.snippet;
};
+ // 后端在一次响应里可能先给出结果、再因后续搜索报错(如 max_uses_exceeded)。
+ // 工具级错误只在一条结果都没有时才代表本次搜索失败。
+ let toolError: { errorCode: unknown } | null = null;
+ const snippetLines = new Map<string, string>();
+
for (const block of raw.content) {
if (!isRecord(block)) continue;
@@ -351,32 +411,44 @@ function parseSearchResponse(raw: unknown, limit: number): ParsedSearchResponse
if (Array.isArray(block.content)) {
if (block.content.length > 0) sawCandidate = true;
for (const item of block.content) {
- if (!isRecord(item) || item.type !== 'web_search_result') continue;
+ if (!isRecord(item)) continue;
+ if (item.type === 'web_search_tool_result_error') {
+ toolError ??= { errorCode: item.error_code };
+ continue;
+ }
+ if (item.type !== 'web_search_result') continue;
const source = sourceFromRecord(item);
if (source) mergeSource(source);
}
} else if (isRecord(block.content) && block.content.type === 'web_search_tool_result_error') {
- return searchToolFailure(block.content.error_code);
+ toolError ??= { errorCode: block.content.error_code };
} else {
return invalidSearchResponse();
}
}
- if (block.type === 'text' && Array.isArray(block.citations)) {
- if (block.citations.length > 0) sawCandidate = true;
- for (const citation of block.citations) {
- if (!isRecord(citation)) continue;
- const source = sourceFromRecord(citation);
- if (source) mergeSource(source);
+ if (block.type === 'text') {
+ if (typeof block.text === 'string') collectSnippetLines(block.text, snippetLines);
+ if (Array.isArray(block.citations)) {
+ if (block.citations.length > 0) sawCandidate = true;
+ for (const citation of block.citations) {
+ if (!isRecord(citation)) continue;
+ const source = sourceFromRecord(citation);
+ if (source) mergeSource(source);
+ }
}
}
}
- if (sawCandidate && sources.size === 0) return invalidSearchResponse();
- return {
- ok: true,
- results: Array.from(sources.values()).slice(0, limit),
- };
+ if (sources.size === 0) {
+ if (toolError) return searchToolFailure(toolError.errorCode);
+ if (sawCandidate) return invalidSearchResponse();
+ }
+ const results = Array.from(sources.values()).slice(0, limit);
+ for (const result of results) {
+ if (!result.snippet) result.snippet = snippetLines.get(sourceMatchKey(result.url)) ?? '';
+ }
+ return { ok: true, results };
}
function webSearchRequestsOf(raw: unknown): number | undefined {
@@ -415,6 +487,7 @@ export function createCindyProxySearchService(deps: CindyProxySearchDeps): Cindy
body: JSON.stringify({
model: CINDY_SEARCH_MODEL_NAME,
max_tokens: CINDY_SEARCH_MAX_TOKENS,
+ system: buildSearchSystemPrompt(limit),
messages: [{ role: 'user', content: query }],
tools: [
{