Skip to content

LLM Wiki is a cross-platform desktop application that turns your documents into an organized, interlinked knowledge base — automatically. Instead of traditional RAG (retrieve-and-answer from scratch every time), the LLM incrementally builds and maintains a persistent wiki from your sources。

TypeScript

fix: caption images in the configured output language

b9bt5dp9hg-ship-itfe70eb975997
diff --git a/src/lib/image-caption-pipeline.test.ts b/src/lib/image-caption-pipeline.test.ts
index fccb4543..96ae462c 100644
--- a/src/lib/image-caption-pipeline.test.ts
+++ b/src/lib/image-caption-pipeline.test.ts
@@ -116,6 +116,23 @@ describe("captionMarkdownImages", () => {
})
})
+ it("forwards outputLanguage to captionImage", async () => {
+ mockReadBase64.mockResolvedValue({ base64: "AAAA", mimeType: "image/png" })
+ mockCaption.mockResolvedValue("ein rotes Quadrat")
+
+ await captionMarkdownImages("/proj", "![](/abs/img-1.png)", cfg, {
+ outputLanguage: "German",
+ })
+
+ expect(mockCaption).toHaveBeenCalledWith(
+ "AAAA",
+ "image/png",
+ cfg,
+ undefined,
+ expect.objectContaining({ outputLanguage: "German" }),
+ )
+ })
+
it("dedupes by SHA-256: two refs to the same bytes → one LLM call, both rewritten", async () => {
// Both URLs return the same base64 bytes — same hash → single
// caption call, both alt-texts populated.
diff --git a/src/lib/image-caption-pipeline.ts b/src/lib/image-caption-pipeline.ts
index 56b75f3b..4998c90a 100644
--- a/src/lib/image-caption-pipeline.ts
+++ b/src/lib/image-caption-pipeline.ts
@@ -260,6 +260,8 @@ export interface CaptionPipelineOptions {
* passes the user's chosen value.
*/
concurrency?: number
+ /** Language requested for factual captions (for example, "German"). */
+ outputLanguage?: string
/**
* Progress hook fired after each image finishes (ok or failed),
* with running counts. Used by ingest to update the activity
@@ -401,7 +403,11 @@ export async function captionMarkdownImages(
bytes.mimeType,
llmConfig,
options?.signal,
- { contextBefore: before, contextAfter: after },
+ {
+ contextBefore: before,
+ contextAfter: after,
+ outputLanguage: options?.outputLanguage,
+ },
)
cache[hash] = {
caption,
diff --git a/src/lib/ingest.ts b/src/lib/ingest.ts
index c428252d..4b8abfe0 100644
--- a/src/lib/ingest.ts
+++ b/src/lib/ingest.ts
@@ -324,9 +324,9 @@ function resolveCaptionConfig(
maxContextSize: mainLlm.maxContextSize,
}
}
-import { buildLanguageDirective } from "@/lib/output-language"
+import { buildLanguageDirective, getOutputLanguage } from "@/lib/output-language"
import { detectLanguage } from "@/lib/detect-language"
-import { sameScriptFamily } from "@/lib/language-metadata"
+import { getLanguagePromptName, sameScriptFamily } from "@/lib/language-metadata"
import {
loadProjectWikiSchemaRouting,
validateWikiPageRouting,
@@ -812,6 +812,7 @@ async function autoIngestImpl(
isSavedImagePromptUrl(pp, sourceSummarySlug, url),
urlToAbsPath: (url) => promptImageUrlToAbs(pp, url),
concurrency: mmCfg.concurrency,
+ outputLanguage: getLanguagePromptName(getOutputLanguage(sourceContent)),
onProgress: (done, total) =>
activity.updateItem(activityId, {
detail: `Captioning images... ${done}/${total}`,
@@ -961,6 +962,7 @@ async function autoIngestImpl(
shouldCaption: (url) => url.startsWith(ourMediaPrefix) || isSavedImagePromptUrl(pp, sourceSummarySlug, url),
urlToAbsPath: (url) => promptImageUrlToAbs(pp, url),
concurrency: mmCfg.concurrency,
+ outputLanguage: getLanguagePromptName(getOutputLanguage(enrichedSourceContent)),
onProgress: (done, total) =>
activity.updateItem(activityId, {
detail: `Captioning images... ${done}/${total}`,
diff --git a/src/lib/vision-caption.test.ts b/src/lib/vision-caption.test.ts
index 46ea9234..6dc74c88 100644
--- a/src/lib/vision-caption.test.ts
+++ b/src/lib/vision-caption.test.ts
@@ -173,6 +173,46 @@ describe("captionImage", () => {
expect(CAPTION_PROMPT).toMatch(/no markdown/)
})
+ it("adds the configured caption language while preserving visible source text", async () => {
+ await captionImage(TINY_B64, "image/png", cfg, undefined, {
+ outputLanguage: "German",
+ })
+ const messages = mockStreamChat.mock.calls[0]?.[1] as ChatMessage[]
+ const textBlock = Array.isArray(messages?.[0]?.content)
+ ? messages[0].content[0]
+ : null
+ expect(textBlock).toMatchObject({ type: "text" })
+ if (!textBlock || textBlock.type !== "text") {
+ throw new Error("expected text prompt block")
+ }
+ expect(textBlock.text).toContain("Write the description in German")
+ expect(textBlock.text).toContain("Preserve visible text verbatim")
+ })
+
+ it("appends the language instruction to the context-aware prompt too", async () => {
+ await captionImage(TINY_B64, "image/png", cfg, undefined, {
+ contextBefore: "Figure 3: Q2 revenue chart",
+ outputLanguage: "German",
+ })
+ const messages = mockStreamChat.mock.calls[0]?.[1] as ChatMessage[]
+ const blocks = messages[0].content as Array<{ type: string; text?: string }>
+ const promptText = blocks[0].text ?? ""
+ expect(promptText).toContain("Figure 3: Q2 revenue chart")
+ expect(promptText).toContain("Write the description in German")
+ })
+
+ it('treats "auto" (and absent) outputLanguage as no language directive', async () => {
+ await captionImage(TINY_B64, "image/png", cfg, undefined, {
+ outputLanguage: "auto",
+ })
+ const messages = mockStreamChat.mock.calls[0]?.[1] as ChatMessage[]
+ const blocks = messages[0].content as Array<{ type: string; text?: string }>
+ // "auto" means "follow the source" — resolving it is the caller's
+ // job (ingest resolves via detectLanguage), so here it must fall
+ // back to the unmodified prompt rather than "Write in auto".
+ expect(blocks[0].text).toBe(CAPTION_PROMPT)
+ })
+
it("uses the no-context prompt when context is empty / whitespace-only", async () => {
mockStreamChat.mockImplementation(async (_c, _m, cb) => {
cb.onDone()
diff --git a/src/lib/vision-caption.ts b/src/lib/vision-caption.ts
index 0b5b3763..140f2cb3 100644
--- a/src/lib/vision-caption.ts
+++ b/src/lib/vision-caption.ts
@@ -69,6 +69,12 @@ import { streamChat, type ChatMessage } from "./llm-client"
export const CAPTION_PROMPT =
"Describe this image factually for a knowledge-base index. Include: any visible text verbatim, chart axes and values, diagram structure (boxes/arrows/labels), key visual elements. Do NOT speculate or editorialize. 2 to 4 sentences. Output plain text only — no markdown, no preamble."
+function captionLanguageInstruction(outputLanguage?: string): string {
+ const language = outputLanguage?.trim()
+ if (!language || language.toLowerCase() === "auto") return ""
+ return `Write the description in ${language}. Preserve visible text verbatim in its original language.`
+}
+
/**
* Build the prompt that gets used WHEN the caller supplies
* surrounding text. Wraps the no-context prompt with an explicit
@@ -83,6 +89,7 @@ export const CAPTION_PROMPT =
export function buildCaptionPromptWithContext(
before: string,
after: string,
+ outputLanguage?: string,
): string {
const fmt = (s: string) => {
const trimmed = s.trim()
@@ -100,7 +107,8 @@ export function buildCaptionPromptWithContext(
"This surrounding text MAY help describe the image — for example, a sentence like \"Figure 3: Q2 revenue chart\" tells you what the chart actually plots. It MAY ALSO be unrelated body text that just happens to flank the image. Use your judgment: if a passage clearly identifies, references, or labels the image, anchor your caption to it; if not, ignore the surrounding text and describe what you see.",
"",
"Now describe the image factually for a knowledge-base index. Include: any visible text verbatim, chart axes and values, diagram structure (boxes/arrows/labels), key visual elements. If the surrounding text contains a relevant figure number / caption / referent, incorporate that specifically. Do NOT invent details that aren't visible in the image or directly stated in the surrounding text. 2 to 4 sentences. Output plain text only — no markdown, no preamble.",
- ].join("\n")
+ captionLanguageInstruction(outputLanguage),
+ ].filter(Boolean).join("\n")
}
export interface CaptionOptions {
@@ -134,6 +142,8 @@ export interface CaptionOptions {
*/
contextBefore?: string
contextAfter?: string
+ /** Output language for the description. Visible source text stays verbatim. */
+ outputLanguage?: string
}
/**
@@ -172,8 +182,10 @@ export async function captionImage(
const after = options?.contextAfter?.trim() ?? ""
const promptText =
before.length > 0 || after.length > 0
- ? buildCaptionPromptWithContext(before, after)
- : CAPTION_PROMPT
+ ? buildCaptionPromptWithContext(before, after, options?.outputLanguage)
+ : [CAPTION_PROMPT, captionLanguageInstruction(options?.outputLanguage)]
+ .filter(Boolean)
+ .join("\n")
const messages: ChatMessage[] = [
{