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: support spaced image caption paths
nash_sudbe720ec46db
diff --git a/src/lib/image-caption-pipeline.test.ts b/src/lib/image-caption-pipeline.test.tsindex 3e297d6a..e0320c77 100644--- a/src/lib/image-caption-pipeline.test.ts+++ b/src/lib/image-caption-pipeline.test.ts@@ -27,7 +27,12 @@ vi.mock("@/commands/fs", () => ({readFileAsBase64: (p: string) => mockReadBase64(p),}))-import { captionMarkdownImages, loadCaptionCache, __test } from "./image-caption-pipeline"+import {+ captionMarkdownImages,+ loadCaptionCache,+ stripMarkdownImageReferences,+ __test,+} from "./image-caption-pipeline"import type { LlmConfig } from "@/stores/wiki-store"const cfg: LlmConfig = {@@ -57,8 +62,8 @@ describe("findImageReferences (helper)", () => {it("captures markdown image syntax with position info", () => {const refs = __test.findImageReferences("text\n\n more")expect(refs).toEqual([- { full: "", alt: "", url: "a.png", index: 5, length: 10 },- { full: "", alt: "label", url: "b.jpg", index: 16, length: 15 },+ { full: "", alt: "", url: "a.png", destination: "a.png", index: 5, length: 10 },+ { full: "", alt: "label", url: "b.jpg", destination: "b.jpg", index: 16, length: 15 },])})@@ -67,6 +72,76 @@ describe("findImageReferences (helper)", () => {expect(refs).toHaveLength(1)expect(refs[0].url).toBe("z.png")})++ it("captures local image destinations containing spaces", () => {+ const refs = __test.findImageReferences(+ " ",+ )++ expect(refs.map((ref) => ref.url)).toEqual([+ "assets/figure one.png",+ "assets/figure two.jpg",+ ])+ })++ it("handles parentheses, optional titles, encoded spaces, and empty destinations", () => {+ const refs = __test.findImageReferences([+ ".png)",+ ".png> \"Screenshot\")",+ "",+ "",+ ].join("\n"))++ expect(refs.map((ref) => ref.url)).toEqual([+ "media/Shot (1).png",+ "media/Shot (2).png",+ "media/figure three.png",+ ])+ })++ it("does not mistake a trailing parenthesized filename segment for a title", () => {+ const refs = __test.findImageReferences(")")+ expect(refs[0]?.url).toBe("media/Shot (1)")+ })++ it("stops malformed image candidates at the current line", () => {+ const markdown = ""+ const refs = __test.findImageReferences(markdown)+ expect(refs).toHaveLength(1)+ expect(refs[0].url).toBe("real image.png")+ expect(refs[0].index).toBe(markdown.indexOf(")+ })++ it("does not join embeds or reference images to a later Markdown link", () => {+ const samples = [+ "See ![[a.png]] and [docs](https://x.test)",+ "![logo][1] then [link](https://x.test)",+ "x![citation needed] see [here](https://x.test)",+ ]+ for (const sample of samples) {+ expect(__test.findImageReferences(sample)).toEqual([])+ expect(stripMarkdownImageReferences(sample)).toBe(sample)+ }+ })++ it("normalizes encoded spaces and escaped parentheses for filesystem access", () => {+ const refs = __test.findImageReferences(+ " two.png)",+ )+ expect(refs.map((ref) => ref.url)).toEqual([+ "media/figure one.png",+ "media/figure)two.png",+ ])+ })+})++describe("stripMarkdownImageReferences", () => {+ it("strips titled and adjacent image references without touching malformed input", () => {+ expect(stripMarkdownImageReferences("ab"))+ .toBe("a b")+ expect(stripMarkdownImageReferences(")+ .toBe("+ })})describe("captionMarkdownImages", () => {@@ -116,6 +191,38 @@ describe("captionMarkdownImages", () => {})})+ it("captions and preserves an image destination containing spaces", async () => {+ mockReadBase64.mockResolvedValue({ base64: "AAAA", mimeType: "image/png" })+ mockCaption.mockResolvedValue("a process diagram")++ const out = await captionMarkdownImages(+ "/proj",+ "",+ cfg,+ )++ expect(mockReadBase64).toHaveBeenCalledWith("/proj/wiki/assets/process flow.png")+ expect(out.enrichedMarkdown).toBe(+ "",+ )+ })++ it("deduplicates equivalent plain and angle-bracket destinations", async () => {+ mockReadBase64.mockResolvedValue({ base64: "AAAA", mimeType: "image/png" })+ mockCaption.mockResolvedValue("same image")++ const out = await captionMarkdownImages(+ "/proj",+ "\n",+ cfg,+ )++ expect(mockCaption).toHaveBeenCalledTimes(1)+ expect(out.enrichedMarkdown).toBe(+ "\n",+ )+ })+it("forwards outputLanguage to captionImage", async () => {mockReadBase64.mockResolvedValue({ base64: "AAAA", mimeType: "image/png" })mockCaption.mockResolvedValue("ein rotes Quadrat")diff --git a/src/lib/image-caption-pipeline.ts b/src/lib/image-caption-pipeline.tsindex 8f63b3cd..e9ac54aa 100644--- a/src/lib/image-caption-pipeline.ts+++ b/src/lib/image-caption-pipeline.ts@@ -161,19 +161,41 @@ async function writeCache(projectPath: string, cache: CaptionCache): Promise<voi* captured but ignored when deciding whether to caption (we* re-caption even non-empty alt because user-typed alt text* usually says "Figure 3" — useless to retrieval).- * - HTML `<img src="...">`: NOT captured by this regex; we+ * - HTML `<img src="...">`: NOT captured by this parser; we* don't generate those, and re-captioning hand-typed HTML* would surprise the user.* - Reference-style images (`![alt][ref]` + `[ref]: url`): NOT* handled — we don't generate them either. Add support if* it matters.*/-const MD_IMAGE_RE = /(!\[)([^\]]*)(\]\()([^)\s]+)(\))/g+const TRAILING_IMAGE_TITLE_RE = /\s+(?:"[^"]*"|'[^']*')\s*$/++function normalizeImageDestination(destination: string): string {+ let value = destination.trim()+ if (value.startsWith("<")) {+ const close = value.indexOf(">")+ if (close < 0) return ""+ value = value.slice(1, close).trim()+ } else {+ value = value.replace(TRAILING_IMAGE_TITLE_RE, "").trim()+ }+ if (!value) return ""+ value = value.replace(/\\([\\() ])/g, "$1")+ try {+ // decodeURI resolves normal Markdown URL encoding such as %20 while+ // deliberately preserving encoded path separators such as %2F.+ return decodeURI(value)+ } catch {+ return value+ }+}interface ImageRef {full: string // the entire `` substringalt: stringurl: string+ /** Destination exactly as written between parentheses, including a title. */+ destination: string/** 0-based offset of the match in the source markdown — used to* slice the surrounding text for context-aware captioning. */index: number@@ -207,23 +229,103 @@ const CONTEXT_CHARS = 150function findImageReferences(markdown: string): ImageRef[] {const out: ImageRef[] = []- // Use exec() in a loop instead of matchAll() so we capture- // `m.index` (the position in the source) — needed for the- // context slicer below.- const re = new RegExp(MD_IMAGE_RE.source, MD_IMAGE_RE.flags)- let m: RegExpExecArray | null- while ((m = re.exec(markdown)) !== null) {+ let cursor = 0+ while (cursor < markdown.length) {+ const start = markdown.indexOf("![", cursor)+ if (start < 0) break+ let altEnd = start + 2+ let escapedAlt = false+ for (; altEnd < markdown.length; altEnd++) {+ const char = markdown[altEnd]+ if (char === "\r" || char === "\n") break+ if (escapedAlt) {+ escapedAlt = false+ continue+ }+ if (char === "\\") {+ escapedAlt = true+ continue+ }+ if (char === "]") break+ }+ if (!markdown.startsWith("](", altEnd)) {+ cursor = Math.max(start + 2, altEnd + 1)+ continue+ }+ const destinationStart = altEnd + 2+ let end = destinationStart+ if (markdown[destinationStart] === "<") {+ let angleEnd = destinationStart + 1+ while (+ angleEnd < markdown.length &&+ markdown[angleEnd] !== ">" &&+ markdown[angleEnd] !== "\r" &&+ markdown[angleEnd] !== "\n"+ ) angleEnd+++ if (markdown[angleEnd] !== ">") {+ cursor = Math.max(destinationStart, angleEnd + 1)+ continue+ }+ end = angleEnd + 1+ while (end < markdown.length && markdown[end] !== ")" && !/[\r\n]/.test(markdown[end])) end+++ } else {+ let depth = 0+ let escaped = false+ for (; end < markdown.length; end++) {+ const char = markdown[end]+ if (char === "\r" || char === "\n") break+ if (escaped) {+ escaped = false+ continue+ }+ if (char === "\\") {+ escaped = true+ continue+ }+ if (char === "(") depth+++ else if (char === ")") {+ if (depth === 0) break+ depth--+ }+ }+ }+ if (markdown[end] !== ")") {+ cursor = Math.max(destinationStart, end + 1)+ continue+ }+ const destination = markdown.slice(destinationStart, end)+ const url = normalizeImageDestination(destination)+ if (!url) {+ cursor = end + 1+ continue+ }+ const full = markdown.slice(start, end + 1)out.push({- full: m[0],- alt: m[2],- url: m[4],- index: m.index,- length: m[0].length,+ full,+ alt: markdown.slice(start + 2, altEnd),+ url,+ destination,+ index: start,+ length: full.length,})+ cursor = end + 1}return out}+export function stripMarkdownImageReferences(markdown: string): string {+ const refs = findImageReferences(markdown)+ if (refs.length === 0) return markdown+ const pieces: string[] = []+ let cursor = 0+ for (const ref of refs) {+ pieces.push(markdown.slice(cursor, ref.index), " ")+ cursor = ref.index + ref.length+ }+ pieces.push(markdown.slice(cursor))+ return pieces.join("")+}+/*** Slice the chars BEFORE and AFTER an image match in the source* markdown. Bounds-safe (window clamps to document edges) and@@ -494,23 +596,28 @@ export async function captionMarkdownImages(}}- // Rewrite the markdown — replace every captured image ref's alt- // text with its caption. We use a callback-style replace rather- // than running findImageReferences twice so we can sanitize the- // alt text inline (no `]` characters or newlines, both of which- // would break the markdown).- const enrichedMarkdown = markdown.replace(- MD_IMAGE_RE,- (whole, openBang, _alt, closeBracket, url, closeParen) => {- const caption = captionByUrl.get(url)- if (!caption) return whole- const safe = caption- .replace(/[\r\n]+/g, " ") // collapse newlines- .replace(/]/g, ")") // ] would close the alt early- .trim()- return `${openBang}${safe}${closeBracket}${url}${closeParen}`- },- )+ // Rewrite all refs in one pass so large image-heavy documents do not copy+ // the full Markdown string once per image.+ const pieces: string[] = []+ let rewriteCursor = 0+ for (const ref of refs) {+ const caption = captionByUrl.get(ref.url)+ pieces.push(markdown.slice(rewriteCursor, ref.index))+ rewriteCursor = ref.index + ref.length+ if (!caption) {+ pieces.push(ref.full)+ continue+ }+ const safe = caption+ .replace(/[\r\n]+/g, " ")+ .replace(/]/g, ")")+ .replace(/\\+$/g, "")+ .trim()+ const replacement = ``+ pieces.push(replacement)+ }+ pieces.push(markdown.slice(rewriteCursor))+ const enrichedMarkdown = pieces.join("")return { enrichedMarkdown, freshCaptions, cachedCaptions, failed }}@@ -522,5 +629,4 @@ export const __test = {findImageReferences,normalizeCaptionLanguage,sha256OfBase64,- MD_IMAGE_RE,}diff --git a/src/lib/ingest-queue.test.ts b/src/lib/ingest-queue.test.tsindex 238171fe..633c1f12 100644--- a/src/lib/ingest-queue.test.ts+++ b/src/lib/ingest-queue.test.ts@@ -20,16 +20,13 @@ vi.mock("./sweep-reviews", () => ({sweepResolvedReviews: vi.fn().mockResolvedValue(0),}))-// Mock embedding so cleanupWrittenFiles' cascade-delete to LanceDB is-// observable. The real module is over in `./embedding` but-// cleanupWrittenFiles dynamically imports it via `@/lib/embedding`,-// hence the absolute mock target.-const removePageEmbeddingMock = vi.fn<(projectPath: string, slug: string) => Promise<void>>(+// Mock the centralized page cascade used by cleanupWrittenFiles.+const cascadeDeleteWikiPageMock = vi.fn<(projectPath: string, filePath: string) => Promise<void>>(async () => {},)-vi.mock("@/lib/embedding", () => ({- removePageEmbedding: (projectPath: string, slug: string) =>- removePageEmbeddingMock(projectPath, slug),+vi.mock("@/lib/wiki-page-delete", () => ({+ cascadeDeleteWikiPage: (projectPath: string, filePath: string) =>+ cascadeDeleteWikiPageMock(projectPath, filePath),}))// Mock project-identity — tests don't hit Tauri plugin-store. Maps the@@ -103,7 +100,11 @@ beforeEach(async () => {mockDeleteFile.mockResolvedValue(undefined as unknown as void)mockSweep.mockReset()mockSweep.mockResolvedValue(0)- removePageEmbeddingMock.mockReset()+ cascadeDeleteWikiPageMock.mockReset()+ cascadeDeleteWikiPageMock.mockImplementation(async (_projectPath, filePath) => {+ const { deleteFile } = await import("@/commands/fs")+ await deleteFile(filePath)+ })// Default: persisted queue file doesn't existmockReadFile.mockRejectedValue(new Error("ENOENT"))@@ -1344,8 +1345,8 @@ describe("ingest-queue — pause/resume processing", () => {})})-// ── cleanupWrittenFiles — file delete + LanceDB chunk cascade ──────-describe("cleanupWrittenFiles — embedding cascade", () => {+// ── cleanupWrittenFiles — centralized page cascade ─────────────────+describe("cleanupWrittenFiles — page cascade", () => {it("deletes each file AND drops its embedding chunks (relative paths)", async () => {const { deleteFile } = await import("@/commands/fs")const mockDeleteFile = vi.mocked(deleteFile)@@ -1362,10 +1363,9 @@ describe("cleanupWrittenFiles — embedding cascade", () => {expect(mockDeleteFile).toHaveBeenNthCalledWith(1, "/proj/wiki/concepts/rope.md")expect(mockDeleteFile).toHaveBeenNthCalledWith(2, "/proj/wiki/entities/transformer.md")- // Embedding cascade uses page slugs (basename minus .md).- expect(removePageEmbeddingMock).toHaveBeenCalledTimes(2)- expect(removePageEmbeddingMock).toHaveBeenNthCalledWith(1, "/proj", "rope")- expect(removePageEmbeddingMock).toHaveBeenNthCalledWith(2, "/proj", "transformer")+ expect(cascadeDeleteWikiPageMock).toHaveBeenCalledTimes(2)+ expect(cascadeDeleteWikiPageMock).toHaveBeenNthCalledWith(1, "/proj", "/proj/wiki/concepts/rope.md")+ expect(cascadeDeleteWikiPageMock).toHaveBeenNthCalledWith(2, "/proj", "/proj/wiki/entities/transformer.md")})it("uses absolute paths verbatim (doesn't double-prefix the project path)", async () => {@@ -1377,8 +1377,10 @@ describe("cleanupWrittenFiles — embedding cascade", () => {await cleanupWrittenFiles("/proj", ["/abs/elsewhere/wiki/concepts/foo.md"])expect(mockDeleteFile).toHaveBeenCalledWith("/abs/elsewhere/wiki/concepts/foo.md")- // Slug derivation still works on absolute paths.- expect(removePageEmbeddingMock).toHaveBeenCalledWith("/proj", "foo")+ expect(cascadeDeleteWikiPageMock).toHaveBeenCalledWith(+ "/proj",+ "/abs/elsewhere/wiki/concepts/foo.md",+ )})it("continues to subsequent files when one delete throws", async () => {@@ -1397,45 +1399,52 @@ describe("cleanupWrittenFiles — embedding cascade", () => {// Both deleteFile attempts happened — the helper kept going.expect(mockDeleteFile).toHaveBeenCalledTimes(2)- // First file's embedding cascade was skipped (deleteFile threw),- // second file's cascade still ran.- expect(removePageEmbeddingMock).toHaveBeenCalledTimes(1)- expect(removePageEmbeddingMock).toHaveBeenCalledWith("/proj", "present")+ // Both cascades were attempted even though the first delete failed.+ expect(cascadeDeleteWikiPageMock).toHaveBeenCalledTimes(2)+ expect(cascadeDeleteWikiPageMock).toHaveBeenNthCalledWith(+ 2,+ "/proj",+ "/proj/wiki/concepts/present.md",+ )})- it("swallows removePageEmbedding errors so a LanceDB issue doesn't abort cleanup", async () => {+ it("swallows page cascade errors so a LanceDB issue doesn't abort cleanup", async () => {const { deleteFile } = await import("@/commands/fs")const mockDeleteFile = vi.mocked(deleteFile)mockDeleteFile.mockReset()mockDeleteFile.mockResolvedValue(undefined)- // First page's embedding cascade throws; second succeeds.- removePageEmbeddingMock- .mockRejectedValueOnce(new Error("lancedb unavailable"))- .mockResolvedValueOnce(undefined)+ // First page cascade throws; second succeeds.+ cascadeDeleteWikiPageMock+ .mockImplementationOnce(async (_projectPath, filePath) => {+ await mockDeleteFile(filePath)+ throw new Error("lancedb unavailable")+ })+ .mockImplementationOnce(async (_projectPath, filePath) => {+ await mockDeleteFile(filePath)+ })await cleanupWrittenFiles("/proj", ["wiki/concepts/a.md","wiki/concepts/b.md",])- // Both file deletes still happened.- expect(mockDeleteFile).toHaveBeenCalledTimes(2)// Second cascade still attempted despite first throwing.- expect(removePageEmbeddingMock).toHaveBeenCalledTimes(2)+ expect(cascadeDeleteWikiPageMock).toHaveBeenCalledTimes(2)})- it("handles Windows backslash paths via getFileStem", async () => {+ it("passes Windows backslash paths to the centralized cascade", async () => {const { deleteFile } = await import("@/commands/fs")const mockDeleteFile = vi.mocked(deleteFile)mockDeleteFile.mockReset()mockDeleteFile.mockResolvedValue(undefined)- // A path that's been rewritten with backslashes (Windows ingest- // pipeline output before normalize). getFileStem must still- // pull "rope" out cleanly so the cascade hits the right page.+ // The centralized cascade owns path normalization and page-id derivation.await cleanupWrittenFiles("C:/proj", ["wiki\\concepts\\rope.md"])- expect(removePageEmbeddingMock).toHaveBeenCalledWith("C:/proj", "rope")+ expect(cascadeDeleteWikiPageMock).toHaveBeenCalledWith(+ "C:/proj",+ "C:/proj/wiki\\concepts\\rope.md",+ )})})diff --git a/src/lib/ingest.ts b/src/lib/ingest.tsindex 360b5af9..8495044b 100644--- a/src/lib/ingest.ts+++ b/src/lib/ingest.ts@@ -37,7 +37,11 @@ import {buildImageMarkdownSection,type SavedImage,} from "@/lib/extract-source-images"-import { captionMarkdownImages, loadCaptionCache } from "@/lib/image-caption-pipeline"+import {+ captionMarkdownImages,+ loadCaptionCache,+ stripMarkdownImageReferences,+} from "@/lib/image-caption-pipeline"import type { MultimodalConfig } from "@/stores/wiki-store"import { GENERATION_WIKI_TYPES } from "@/lib/wiki-page-types"import { computeContextBudget } from "@/lib/context-budget"@@ -982,13 +986,9 @@ async function autoIngestImpl(const mmCfg = useWikiStore.getState().multimodalConfigconst captionLlm = resolveCaptionConfig(mmCfg, llmConfig)if (!mmCfg.enabled && savedImages.length > 0) {- // Strip `` references — match the same regex shape- // we use elsewhere for image refs. Preserve a single space+ // Strip recognized `` references with the shared parser. Preserve a single space// where the ref used to sit so adjacent words don't fuse.- enrichedSourceContent = sourceContent.replace(- /!\[[^\]]*\]\([^)\s]+\)/g,- " ",- )+ enrichedSourceContent = stripMarkdownImageReferences(sourceContent)console.log(`[ingest:caption] disabled — stripped image refs from sourceContent (${savedImages.length} image(s) won't appear in wiki pages)`,)