ralph-hero-knowledge-index 0.1.23 → 0.1.24

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "ralph-knowledge",
-  "version": "0.1.23",
+  "version": "0.1.24",
   "description": "Knowledge graph for ralph-hero: semantic search, relationship traversal, and document indexing across thoughts/ documents. Optional companion to ralph-hero.",
   "author": {
     "name": "Chad Dubiel",

package/.mcp.json

@@ -2,7 +2,7 @@
   "mcpServers": {
     "ralph-knowledge": {
       "command": "npx",
-      "args": ["-y", "ralph-hero-knowledge-index@0.1.23"]
+      "args": ["-y", "ralph-hero-knowledge-index@0.1.24"]
     }
   }
 }

package/dist/embedder.d.ts

@@ -1,4 +1,31 @@
 import { type FeatureExtractionPipeline } from "@huggingface/transformers";
+import { type Chunk, type ChunkerOptions } from "./chunker.js";
 export declare function getEmbedder(): Promise<FeatureExtractionPipeline>;
 export declare function embed(text: string): Promise<Float32Array>;
+/**
+ * A chunk paired with the embedding of its (contextualized) content.
+ * Extends the base Chunk from the chunker module with an embedding vector
+ * and an optional contextPrefix (populated by Phase 6 — contextual retrieval).
+ */
+export interface DocumentChunk extends Chunk {
+    embedding: Float32Array;
+    contextPrefix?: string;
+}
+/**
+ * Embed a document by splitting it into chunks and emitting one embedding
+ * per chunk. The embedded text for each chunk is
+ * `${title}\n${tagLine}\n${chunk.content}` so the semantic anchors (title +
+ * tags) travel with every chunk embedding — matching the shape of the legacy
+ * `prepareTextForEmbedding()` but without the 500-char truncation.
+ *
+ * Short documents (<= chunkSize) produce exactly one chunk covering the whole
+ * content. Empty content yields a single chunk with empty content (so callers
+ * still get a title/tag-only embedding for stub documents).
+ */
+export declare function embedDocument(title: string, tags: string[], content: string, opts?: ChunkerOptions): Promise<DocumentChunk[]>;
+/**
+ * Back-compat shim: kept so callers outside the reindex path can still build
+ * a title/tags/first-paragraph string. No longer used by `embedDocument` (the
+ * per-chunk flow prepends title + tags directly).
+ */
 export declare function prepareTextForEmbedding(title: string, tags: string[], content: string): string;

package/dist/embedder.js

@@ -1,6 +1,6 @@
 import { pipeline, } from "@huggingface/transformers";
+import { chunkText } from "./chunker.js";
 const MODEL_ID = "Xenova/all-MiniLM-L6-v2";
-const MAX_CHARS = 500;
 let embedderInstance = null;
 export async function getEmbedder() {
     if (!embedderInstance) {
@@ -11,19 +11,58 @@ export async function getEmbedder() {
 }
 export async function embed(text) {
     const embedder = await getEmbedder();
-    const truncated = text.slice(0, MAX_CHARS);
-    const output = await embedder(truncated, {
+    // Pass text directly — the transformer's own 512-token window handles overflow.
+    const output = await embedder(text, {
         pooling: "mean",
         normalize: true,
     });
     return new Float32Array(output.data);
 }
+/**
+ * Embed a document by splitting it into chunks and emitting one embedding
+ * per chunk. The embedded text for each chunk is
+ * `${title}\n${tagLine}\n${chunk.content}` so the semantic anchors (title +
+ * tags) travel with every chunk embedding — matching the shape of the legacy
+ * `prepareTextForEmbedding()` but without the 500-char truncation.
+ *
+ * Short documents (<= chunkSize) produce exactly one chunk covering the whole
+ * content. Empty content yields a single chunk with empty content (so callers
+ * still get a title/tag-only embedding for stub documents).
+ */
+export async function embedDocument(title, tags, content, opts) {
+    const tagLine = tags.length > 0 ? tags.join(", ") : "";
+    // If content is empty, still emit one chunk so the document has a searchable
+    // embedding anchored on title + tags (preserves legacy behavior for
+    // frontmatter-only / stub documents).
+    const chunks = content.length === 0
+        ? [{ index: 0, content: "", charStart: 0, charEnd: 0 }]
+        : chunkText(content, opts);
+    const out = [];
+    for (const chunk of chunks) {
+        const parts = [title, tagLine, chunk.content].filter(p => p.length > 0);
+        const embedText = parts.join("\n");
+        const embedding = await embed(embedText);
+        out.push({
+            index: chunk.index,
+            content: chunk.content,
+            charStart: chunk.charStart,
+            charEnd: chunk.charEnd,
+            embedding,
+        });
+    }
+    return out;
+}
+/**
+ * Back-compat shim: kept so callers outside the reindex path can still build
+ * a title/tags/first-paragraph string. No longer used by `embedDocument` (the
+ * per-chunk flow prepends title + tags directly).
+ */
 export function prepareTextForEmbedding(title, tags, content) {
     const tagLine = tags.length > 0 ? tags.join(", ") : "";
     // Extract first paragraph: split on blank lines, take first non-empty segment
     const paragraphs = content.split(/\n\n+/);
     const firstParagraph = paragraphs.find(p => p.trim().length > 0)?.trim() ?? "";
     const parts = [title, tagLine, firstParagraph].filter(p => p.length > 0);
-    return parts.join("\n").slice(0, MAX_CHARS);
+    return parts.join("\n");
 }
 //# sourceMappingURL=embedder.js.map

package/dist/embedder.js.map

@@ -1 +1 @@
-{"version":3,"file":"embedder.js","sourceRoot":"","sources":["../src/embedder.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,QAAQ,GAET,MAAM,2BAA2B,CAAC;AAEnC,MAAM,QAAQ,GAAG,yBAAyB,CAAC;AAC3C,MAAM,SAAS,GAAG,GAAG,CAAC;AAEtB,IAAI,gBAAgB,GAAqC,IAAI,CAAC;AAE9D,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,IAAI,CAAC,gBAAgB,EAAE,CAAC;QACtB,mEAAmE;QACnE,gBAAgB,GAAG,CAAC,MAAM,QAAQ,CAChC,oBAAoB,EACpB,QAAQ,CACT,CAA8B,CAAC;IAClC,CAAC;IACD,OAAO,gBAAgB,CAAC;AAC1B,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,KAAK,CAAC,IAAY;IACtC,MAAM,QAAQ,GAAG,MAAM,WAAW,EAAE,CAAC;IACrC,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC;IAC3C,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,SAAS,EAAE;QACvC,OAAO,EAAE,MAAM;QACf,SAAS,EAAE,IAAI;KAChB,CAAC,CAAC;IACH,OAAO,IAAI,YAAY,CAAC,MAAM,CAAC,IAAyB,CAAC,CAAC;AAC5D,CAAC;AAED,MAAM,UAAU,uBAAuB,CACrC,KAAa,EACb,IAAc,EACd,OAAe;IAEf,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IACvD,8EAA8E;IAC9E,MAAM,UAAU,GAAG,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;IAC1C,MAAM,cAAc,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;IAC/E,MAAM,KAAK,GAAG,CAAC,KAAK,EAAE,OAAO,EAAE,cAAc,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IACzE,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC;AAC9C,CAAC"}
+{"version":3,"file":"embedder.js","sourceRoot":"","sources":["../src/embedder.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,QAAQ,GAET,MAAM,2BAA2B,CAAC;AACnC,OAAO,EAAE,SAAS,EAAmC,MAAM,cAAc,CAAC;AAE1E,MAAM,QAAQ,GAAG,yBAAyB,CAAC;AAE3C,IAAI,gBAAgB,GAAqC,IAAI,CAAC;AAE9D,MAAM,CAAC,KAAK,UAAU,WAAW;IAC/B,IAAI,CAAC,gBAAgB,EAAE,CAAC;QACtB,mEAAmE;QACnE,gBAAgB,GAAG,CAAC,MAAM,QAAQ,CAChC,oBAAoB,EACpB,QAAQ,CACT,CAA8B,CAAC;IAClC,CAAC;IACD,OAAO,gBAAgB,CAAC;AAC1B,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,KAAK,CAAC,IAAY;IACtC,MAAM,QAAQ,GAAG,MAAM,WAAW,EAAE,CAAC;IACrC,gFAAgF;IAChF,MAAM,MAAM,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE;QAClC,OAAO,EAAE,MAAM;QACf,SAAS,EAAE,IAAI;KAChB,CAAC,CAAC;IACH,OAAO,IAAI,YAAY,CAAC,MAAM,CAAC,IAAyB,CAAC,CAAC;AAC5D,CAAC;AAYD;;;;;;;;;;GAUG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CACjC,KAAa,EACb,IAAc,EACd,OAAe,EACf,IAAqB;IAErB,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAEvD,6EAA6E;IAC7E,oEAAoE;IACpE,sCAAsC;IACtC,MAAM,MAAM,GAAY,OAAO,CAAC,MAAM,KAAK,CAAC;QAC1C,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,OAAO,EAAE,EAAE,EAAE,SAAS,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC;QACvD,CAAC,CAAC,SAAS,CAAC,OAAO,EAAE,IAAI,CAAC,CAAC;IAE7B,MAAM,GAAG,GAAoB,EAAE,CAAC;IAChC,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QAC3B,MAAM,KAAK,GAAG,CAAC,KAAK,EAAE,OAAO,EAAE,KAAK,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;QACxE,MAAM,SAAS,GAAG,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACnC,MAAM,SAAS,GAAG,MAAM,KAAK,CAAC,SAAS,CAAC,CAAC;QACzC,GAAG,CAAC,IAAI,CAAC;YACP,KAAK,EAAE,KAAK,CAAC,KAAK;YAClB,OAAO,EAAE,KAAK,CAAC,OAAO;YACtB,SAAS,EAAE,KAAK,CAAC,SAAS;YAC1B,OAAO,EAAE,KAAK,CAAC,OAAO;YACtB,SAAS;SACV,CAAC,CAAC;IACL,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,uBAAuB,CACrC,KAAa,EACb,IAAc,EACd,OAAe;IAEf,MAAM,OAAO,GAAG,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IACvD,8EAA8E;IAC9E,MAAM,UAAU,GAAG,OAAO,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;IAC1C,MAAM,cAAc,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC;IAC/E,MAAM,KAAK,GAAG,CAAC,KAAK,EAAE,OAAO,EAAE,cAAc,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IACzE,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1B,CAAC"}

package/dist/reindex.js

@@ -4,7 +4,7 @@ import { homedir } from "node:os";
 import { KnowledgeDB } from "./db.js";
 import { FtsSearch } from "./search.js";
 import { VectorSearch } from "./vector-search.js";
-import { embed, prepareTextForEmbedding } from "./embedder.js";
+import { embedDocument } from "./embedder.js";
 import { parseDocument } from "./parser.js";
 import { findMarkdownFiles } from "./file-scanner.js";
 import { generateIndexes } from "./generate-indexes.js";
@@ -37,7 +37,10 @@ export async function reindex(dirs, dbPath, generate = false, ignorePatterns) {
     }
     console.log(`Found ${filesOnDisk.length} total markdown files`);
     const filesOnDiskSet = new Set(filesOnDisk.map(f => resolve(f)));
-    // Phase 1: Delete stale entries for files no longer on disk
+    // Phase 1: Delete stale entries for files no longer on disk.
+    // Chunk rows cascade from documents via ON DELETE CASCADE on chunks.document_id,
+    // but the vec0 virtual table does not participate in FK cascades — we must
+    // explicitly delete chunk-level vec rows via GLOB pattern.
     const syncedPaths = db.getAllSyncPaths();
     let deleted = 0;
     for (const syncedPath of syncedPaths) {
@@ -45,6 +48,8 @@ export async function reindex(dirs, dbPath, generate = false, ignorePatterns) {
             const id = basename(syncedPath, ".md");
             fts.deleteFtsEntry(id);
             db.deleteDocument(id);
+            vec.deleteChunkVecsByDoc(id);
+            // Also delete any legacy doc-level vec row (pre-chunks schema).
             vec.deleteEmbedding(id);
             db.deleteSyncRecord(syncedPath);
             deleted++;
@@ -117,10 +122,25 @@ export async function reindex(dirs, dbPath, generate = false, ignorePatterns) {
             db.upsertStubDocument(edge.targetId);
             db.addRelationship(edge.sourceId, edge.targetId, "untyped", edge.context);
         }
-        const text = prepareTextForEmbedding(parsed.title, parsed.tags, parsed.content);
+        // Chunk-aware embedding: emit one embedding per chunk, persist to both
+        // the `chunks` table and the `documents_vec` virtual table with chunk ids
+        // of the form `${doc.id}#c${index}`.
+        //
+        // We first clear any stale chunk rows for this doc_id (the document
+        // body may have shrunk across re-indexes) and stale chunk vec rows (which
+        // don't cascade from the `chunks` table because vec0 is a virtual table).
+        db.db.prepare("DELETE FROM chunks WHERE document_id = ?").run(parsed.id);
+        vec.deleteChunkVecsByDoc(parsed.id);
+        // Drop any pre-chunks schema vec row that used the bare doc id.
+        vec.deleteEmbedding(parsed.id);
         try {
-            const embedding = await embed(text);
-            vec.upsertEmbedding(parsed.id, embedding);
+            const chunks = await embedDocument(parsed.title, parsed.tags, parsed.content);
+            const insertChunk = db.db.prepare("INSERT INTO chunks (id, document_id, chunk_index, content, char_start, char_end) VALUES (?, ?, ?, ?, ?, ?)");
+            for (const chunk of chunks) {
+                const chunkId = `${parsed.id}#c${chunk.index}`;
+                insertChunk.run(chunkId, parsed.id, chunk.index, chunk.content, chunk.charStart, chunk.charEnd);
+                vec.upsertEmbedding(chunkId, chunk.embedding);
+            }
         }
         catch (e) {
             console.warn(`Failed to embed ${id}: ${e.message}`);

package/dist/reindex.js.map

@@ -1 +1 @@
-{"version":3,"file":"reindex.js","sourceRoot":"","sources":["../src/reindex.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,QAAQ,EAAE,MAAM,SAAS,CAAC;AACjD,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,WAAW,CAAC;AAC9D,OAAO,EAAE,OAAO,EAAE,MAAM,SAAS,CAAC;AAClC,OAAO,EAAE,WAAW,EAAE,MAAM,SAAS,CAAC;AACtC,OAAO,EAAE,SAAS,EAAE,MAAM,aAAa,CAAC;AACxC,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAE,KAAK,EAAE,uBAAuB,EAAE,MAAM,eAAe,CAAC;AAC/D,OAAO,EAAE,aAAa,EAAuB,MAAM,aAAa,CAAC;AACjE,OAAO,EAAE,iBAAiB,EAAE,MAAM,mBAAmB,CAAC;AACtD,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AACxD,OAAO,EAAE,UAAU,EAAwB,MAAM,aAAa,CAAC;AAC/D,OAAO,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAEhD,MAAM,CAAC,KAAK,UAAU,OAAO,CAC3B,IAAc,EACd,MAAc,EACd,WAAoB,KAAK,EACzB,cAAyB;IAEzB,OAAO,CAAC,GAAG,CAAC,YAAY,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,MAAM,EAAE,CAAC,CAAC;IAExD,MAAM,EAAE,GAAG,IAAI,WAAW,CAAC,MAAM,CAAC,CAAC;IACnC,MAAM,GAAG,GAAG,IAAI,SAAS,CAAC,EAAE,CAAC,CAAC;IAC9B,GAAG,CAAC,WAAW,EAAE,CAAC;IAClB,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,EAAE,CAAC,CAAC;IACjC,GAAG,CAAC,WAAW,EAAE,CAAC;IAElB,8EAA8E;IAC9E,MAAM,cAAc,GAAG,GAAG,CAAC;IAC3B,MAAM,cAAc,GAAG,EAAE,CAAC,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACpD,IAAI,mBAAmB,GAAG,KAAK,CAAC;IAChC,IAAI,cAAc,KAAK,cAAc,EAAE,CAAC;QACtC,OAAO,CAAC,GAAG,CAAC,uEAAuE,CAAC,CAAC;QACrF,EAAE,CAAC,gBAAgB,EAAE,CAAC;QACtB,EAAE,CAAC,OAAO,CAAC,gBAAgB,EAAE,cAAc,CAAC,CAAC;QAC7C,mBAAmB,GAAG,IAAI,CAAC;IAC7B,CAAC;IAED,kCAAkC;IAClC,MAAM,WAAW,GAAa,EAAE,CAAC;IACjC,KAAK,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;QACvB,MAAM,OAAO,GAAG,iBAAiB,CAAC,GAAG,EAAE,cAAc,CAAC,CAAC;QACvD,MAAM,KAAK,GAAG,iBAAiB,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC;QAC9C,OAAO,CAAC,GAAG,CAAC,KAAK,GAAG,KAAK,KAAK,CAAC,MAAM,QAAQ,CAAC,CAAC;QAC/C,WAAW,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,CAAC;IAC7B,CAAC;IACD,OAAO,CAAC,GAAG,CAAC,SAAS,WAAW,CAAC,MAAM,uBAAuB,CAAC,CAAC;IAEhE,MAAM,cAAc,GAAG,IAAI,GAAG,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEjE,4DAA4D;IAC5D,MAAM,WAAW,GAAG,EAAE,CAAC,eAAe,EAAE,CAAC;IACzC,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,KAAK,MAAM,UAAU,IAAI,WAAW,EAAE,CAAC;QACrC,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,UAAU,CAAC,EAAE,CAAC;YACpC,MAAM,EAAE,GAAG,QAAQ,CAAC,UAAU,EAAE,KAAK,CAAC,CAAC;YACvC,GAAG,CAAC,cAAc,CAAC,EAAE,CAAC,CAAC;YACvB,EAAE,CAAC,cAAc,CAAC,EAAE,CAAC,CAAC;YACtB,GAAG,CAAC,eAAe,CAAC,EAAE,CAAC,CAAC;YACxB,EAAE,CAAC,gBAAgB,CAAC,UAAU,CAAC,CAAC;YAChC,OAAO,EAAE,CAAC;QACZ,CAAC;IACH,CAAC;IACD,IAAI,OAAO,GAAG,CAAC,EAAE,CAAC;QAChB,OAAO,CAAC,GAAG,CAAC,aAAa,OAAO,gBAAgB,CAAC,CAAC;IACpD,CAAC;IAED,yCAAyC;IACzC,MAAM,UAAU,GAAqB,EAAE,CAAC;IACxC,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,KAAK,MAAM,QAAQ,IAAI,WAAW,EAAE,CAAC;QACnC,MAAM,OAAO,GAAG,OAAO,CAAC,QAAQ,CAAC,CAAC;QAClC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,OAAO,CAAC,CAAC;QAEpD,8CAA8C;QAC9C,MAAM,UAAU,GAAG,EAAE,CAAC,aAAa,CAAC,OAAO,CAAC,CAAC;QAC7C,IAAI,UAAU,IAAI,UAAU,CAAC,KAAK,KAAK,KAAK,EAAE,CAAC;YAC7C,OAAO,EAAE,CAAC;YACV,SAAS;QACX,CAAC;QAED,MAAM,GAAG,GAAG,YAAY,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;QAC5C,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACjE,MAAM,OAAO,GAAG,SAAS;YACvB,CAAC,CAAC,QAAQ,CAAC,OAAO,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,OAAO,CAAC;YAC7C,CAAC,CAAC,QAAQ,CAAC;QACb,MAAM,EAAE,GAAG,QAAQ,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;QAErC,MAAM,MAAM,GAAG,aAAa,CAAC,EAAE,EAAE,OAAO,EAAE,GAAG,CAAC,CAAC;QAC/C,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAExB,MAAM,OAAO,GAAa,EAAE,CAAC;QAC7B,IAAI,CAAC,MAAM,CAAC,IAAI;YAAE,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACvC,IAAI,CAAC,MAAM,CAAC,IAAI;YAAE,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACvC,IAAI,CAAC,MAAM,CAAC,MAAM;YAAE,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QAC3C,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACvB,OAAO,CAAC,IAAI,CAAC,cAAc,EAAE,yBAAyB,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAC9E,CAAC;QAED,uEAAuE;QACvE,IAAI,EAAE,CAAC,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC;YACjC,GAAG,CAAC,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAChC,CAAC;QAED,EAAE,CAAC,cAAc,CAAC;YAChB,EAAE,EAAE,MAAM,CAAC,EAAE;YACb,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,KAAK,EAAE,MAAM,CAAC,KAAK;YACnB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,MAAM,EAAE,MAAM,CAAC,MAAM;YACrB,WAAW,EAAE,MAAM,CAAC,WAAW;YAC/B,OAAO,EAAE,MAAM,CAAC,OAAO;SACxB,CAAC,CAAC;QAEH,oCAAoC;QACpC,GAAG,CAAC,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAE9B,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAC3B,EAAE,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;QACrC,CAAC;QAED,4EAA4E;QAC5E,EAAE,CAAC,EAAE,CAAC,OAAO,CAAC,+CAA+C,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAE9E,6EAA6E;QAC7E,gFAAgF;QAChF,8EAA8E;QAC9E,8CAA8C;QAC9C,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,aAAa,EAAE,CAAC;YACvC,EAAE,CAAC,kBAAkB,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;YACpC,EAAE,CAAC,eAAe,CAAC,GAAG,CAAC,QAAQ,EAAE,GAAG,CAAC,QAAQ,EAAE,GAAG,CAAC,IAAI,CAAC,CAAC;QAC3D,CAAC;QAED,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,YAAY,EAAE,CAAC;YACvC,EAAE,CAAC,kBAAkB,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YACrC,EAAE,CAAC,eAAe,CAAC,IAAI,CAAC,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,SAAS,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC;QAC5E,CAAC;QAED,MAAM,IAAI,GAAG,uBAAuB,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,OAAO,CAAC,CAAC;QAChF,IAAI,CAAC;YACH,MAAM,SAAS,GAAG,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC;YACpC,GAAG,CAAC,eAAe,CAAC,MAAM,CAAC,EAAE,EAAE,SAAS,CAAC,CAAC;QAC5C,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,OAAO,CAAC,IAAI,CAAC,mBAAmB,EAAE,KAAM,CAAW,CAAC,OAAO,EAAE,CAAC,CAAC;QACjE,CAAC;QAED,EAAE,CAAC,gBAAgB,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC;QAEpC,OAAO,EAAE,CAAC;QACV,IAAI,OAAO,GAAG,EAAE,KAAK,CAAC,EAAE,CAAC;YACvB,OAAO,CAAC,GAAG,CAAC,KAAK,OAAO,IAAI,WAAW,CAAC,MAAM,UAAU,CAAC,CAAC;QAC5D,CAAC;IACH,CAAC;IAED,qFAAqF;IACrF,8FAA8F;IAC9F,IAAI,mBAAmB,EAAE,CAAC;QACxB,GAAG,CAAC,YAAY,EAAE,CAAC;IACrB,CAAC;IAED,gGAAgG;IAChG,MAAM,YAAY,GAAG,IAAI,GAAG,CACzB,EAAE,CAAC,EAAE,CAAC,OAAO,CAAC,8CAA8C,CAAC,CAAC,GAAG,EAAmC;SAClG,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CACzB,CAAC;IAEF,uEAAuE;IACvE,IAAI,SAAS,GAAG,CAAC,CAAC;IAClB,KAAK,MAAM,QAAQ,IAAI,YAAY,EAAE,CAAC;QACpC,IAAI,CAAC,EAAE,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,CAAC;YACjC,EAAE,CAAC,kBAAkB,CAAC,QAAQ,CAAC,CAAC;YAChC,SAAS,EAAE,CAAC;QACd,CAAC;IACH,CAAC;IACD,OAAO,CAAC,GAAG,CAAC,aAAa,SAAS,sCAAsC,CAAC,CAAC;IAE1E,IAAI,CAAC;QACH,IAAI,QAAQ,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAChC,OAAO,CAAC,GAAG,CAAC,2BAA2B,CAAC,CAAC;YACzC,eAAe,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC;YACrC,OAAO,CAAC,GAAG,CAAC,wBAAwB,CAAC,CAAC;QACxC,CAAC;IACH,CAAC;YAAS,CAAC;QACT,OAAO,CAAC,GAAG,CAAC,SAAS,OAAO,uBAAuB,OAAO,uBAAuB,CAAC,CAAC;QACnF,EAAE,CAAC,KAAK,EAAE,CAAC;IACb,CAAC;AACH,CAAC;AAED,MAAM,eAAe,GAAG,IAAI,CAAC,OAAO,EAAE,EAAE,aAAa,EAAE,cAAc,CAAC,CAAC;AAYvE;;;;;;;;;;;;;GAaG;AACH,MAAM,UAAU,WAAW;IACzB,MAAM,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACtC,MAAM,UAAU,GAAG,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC;IACrD,MAAM,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC;IAC5D,MAAM,KAAK,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC;IACtD,MAAM,OAAO,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC;IAE3D,MAAM,MAAM,GAAG,UAAU,EAAE,CAAC;IAE5B,MAAM,aAAa,GAAG,GAAW,EAAE,CACjC,KAAK;QACL,OAAO,CAAC,GAAG,CAAC,kBAAkB;QAC9B,MAAM,CAAC,MAAM;QACb,eAAe,CAAC;IAElB,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACvB,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC,CAAC;QACrC,OAAO;YACL,IAAI,EAAE,OAAO;YACb,MAAM,EAAE,aAAa,EAAE;YACvB,QAAQ,EAAE,CAAC,UAAU;YACrB,MAAM,EAAE,KAAK;YACb,MAAM;SACP,CAAC;IACJ,CAAC;IAED,MAAM,OAAO,GAAG,OAAO,CAAC,GAAG,CAAC,oBAAoB,CAAC;IACjD,IAAI,OAAO,EAAE,CAAC;QACZ,MAAM,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;QACrE,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACtB,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC,CAAC;YACrC,OAAO;gBACL,IAAI,EAAE,MAAM;gBACZ,MAAM,EAAE,aAAa,EAAE;gBACvB,QAAQ,EAAE,CAAC,UAAU;gBACrB,MAAM,EAAE,KAAK;gBACb,MAAM;aACP,CAAC;QACJ,CAAC;IACH,CAAC;IAED,IAAI,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC5C,OAAO,CAAC,GAAG,CAAC,0BAA0B,CAAC,CAAC;QACxC,OAAO;YACL,IAAI,EAAE,MAAM,CAAC,KAAK;YAClB,MAAM,EAAE,aAAa,EAAE;YACvB,QAAQ,EAAE,CAAC,UAAU;YACrB,MAAM,EAAE,QAAQ;YAChB,MAAM;SACP,CAAC;IACJ,CAAC;IAED,OAAO,CAAC,GAAG,CAAC,4BAA4B,CAAC,CAAC;IAC1C,OAAO;QACL,IAAI,EAAE,CAAC,gBAAgB,CAAC;QACxB,MAAM,EAAE,aAAa,EAAE;QACvB,QAAQ,EAAE,CAAC,UAAU;QACrB,MAAM,EAAE,UAAU;QAClB,MAAM;KACP,CAAC;AACJ,CAAC;AAED,MAAM,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;AACvD,IAAI,MAAM,EAAE,CAAC;IACX,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,WAAW,EAAE,CAAC;IACzD,OAAO,CAAC,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,cAAc,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC9E,CAAC"}
+{"version":3,"file":"reindex.js","sourceRoot":"","sources":["../src/reindex.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,QAAQ,EAAE,MAAM,SAAS,CAAC;AACjD,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,WAAW,CAAC;AAC9D,OAAO,EAAE,OAAO,EAAE,MAAM,SAAS,CAAC;AAClC,OAAO,EAAE,WAAW,EAAE,MAAM,SAAS,CAAC;AACtC,OAAO,EAAE,SAAS,EAAE,MAAM,aAAa,CAAC;AACxC,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAC;AAClD,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AAC9C,OAAO,EAAE,aAAa,EAAuB,MAAM,aAAa,CAAC;AACjE,OAAO,EAAE,iBAAiB,EAAE,MAAM,mBAAmB,CAAC;AACtD,OAAO,EAAE,eAAe,EAAE,MAAM,uBAAuB,CAAC;AACxD,OAAO,EAAE,UAAU,EAAwB,MAAM,aAAa,CAAC;AAC/D,OAAO,EAAE,iBAAiB,EAAE,MAAM,aAAa,CAAC;AAEhD,MAAM,CAAC,KAAK,UAAU,OAAO,CAC3B,IAAc,EACd,MAAc,EACd,WAAoB,KAAK,EACzB,cAAyB;IAEzB,OAAO,CAAC,GAAG,CAAC,YAAY,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,MAAM,EAAE,CAAC,CAAC;IAExD,MAAM,EAAE,GAAG,IAAI,WAAW,CAAC,MAAM,CAAC,CAAC;IACnC,MAAM,GAAG,GAAG,IAAI,SAAS,CAAC,EAAE,CAAC,CAAC;IAC9B,GAAG,CAAC,WAAW,EAAE,CAAC;IAClB,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,EAAE,CAAC,CAAC;IACjC,GAAG,CAAC,WAAW,EAAE,CAAC;IAElB,8EAA8E;IAC9E,MAAM,cAAc,GAAG,GAAG,CAAC;IAC3B,MAAM,cAAc,GAAG,EAAE,CAAC,OAAO,CAAC,gBAAgB,CAAC,CAAC;IACpD,IAAI,mBAAmB,GAAG,KAAK,CAAC;IAChC,IAAI,cAAc,KAAK,cAAc,EAAE,CAAC;QACtC,OAAO,CAAC,GAAG,CAAC,uEAAuE,CAAC,CAAC;QACrF,EAAE,CAAC,gBAAgB,EAAE,CAAC;QACtB,EAAE,CAAC,OAAO,CAAC,gBAAgB,EAAE,cAAc,CAAC,CAAC;QAC7C,mBAAmB,GAAG,IAAI,CAAC;IAC7B,CAAC;IAED,kCAAkC;IAClC,MAAM,WAAW,GAAa,EAAE,CAAC;IACjC,KAAK,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;QACvB,MAAM,OAAO,GAAG,iBAAiB,CAAC,GAAG,EAAE,cAAc,CAAC,CAAC;QACvD,MAAM,KAAK,GAAG,iBAAiB,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC;QAC9C,OAAO,CAAC,GAAG,CAAC,KAAK,GAAG,KAAK,KAAK,CAAC,MAAM,QAAQ,CAAC,CAAC;QAC/C,WAAW,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,CAAC;IAC7B,CAAC;IACD,OAAO,CAAC,GAAG,CAAC,SAAS,WAAW,CAAC,MAAM,uBAAuB,CAAC,CAAC;IAEhE,MAAM,cAAc,GAAG,IAAI,GAAG,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAEjE,6DAA6D;IAC7D,iFAAiF;IACjF,2EAA2E;IAC3E,2DAA2D;IAC3D,MAAM,WAAW,GAAG,EAAE,CAAC,eAAe,EAAE,CAAC;IACzC,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,KAAK,MAAM,UAAU,IAAI,WAAW,EAAE,CAAC;QACrC,IAAI,CAAC,cAAc,CAAC,GAAG,CAAC,UAAU,CAAC,EAAE,CAAC;YACpC,MAAM,EAAE,GAAG,QAAQ,CAAC,UAAU,EAAE,KAAK,CAAC,CAAC;YACvC,GAAG,CAAC,cAAc,CAAC,EAAE,CAAC,CAAC;YACvB,EAAE,CAAC,cAAc,CAAC,EAAE,CAAC,CAAC;YACtB,GAAG,CAAC,oBAAoB,CAAC,EAAE,CAAC,CAAC;YAC7B,gEAAgE;YAChE,GAAG,CAAC,eAAe,CAAC,EAAE,CAAC,CAAC;YACxB,EAAE,CAAC,gBAAgB,CAAC,UAAU,CAAC,CAAC;YAChC,OAAO,EAAE,CAAC;QACZ,CAAC;IACH,CAAC;IACD,IAAI,OAAO,GAAG,CAAC,EAAE,CAAC;QAChB,OAAO,CAAC,GAAG,CAAC,aAAa,OAAO,gBAAgB,CAAC,CAAC;IACpD,CAAC;IAED,yCAAyC;IACzC,MAAM,UAAU,GAAqB,EAAE,CAAC;IACxC,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,KAAK,MAAM,QAAQ,IAAI,WAAW,EAAE,CAAC;QACnC,MAAM,OAAO,GAAG,OAAO,CAAC,QAAQ,CAAC,CAAC;QAClC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,OAAO,CAAC,CAAC;QAEpD,8CAA8C;QAC9C,MAAM,UAAU,GAAG,EAAE,CAAC,aAAa,CAAC,OAAO,CAAC,CAAC;QAC7C,IAAI,UAAU,IAAI,UAAU,CAAC,KAAK,KAAK,KAAK,EAAE,CAAC;YAC7C,OAAO,EAAE,CAAC;YACV,SAAS;QACX,CAAC;QAED,MAAM,GAAG,GAAG,YAAY,CAAC,QAAQ,EAAE,OAAO,CAAC,CAAC;QAC5C,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACjE,MAAM,OAAO,GAAG,SAAS;YACvB,CAAC,CAAC,QAAQ,CAAC,OAAO,CAAC,SAAS,EAAE,IAAI,CAAC,EAAE,OAAO,CAAC;YAC7C,CAAC,CAAC,QAAQ,CAAC;QACb,MAAM,EAAE,GAAG,QAAQ,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;QAErC,MAAM,MAAM,GAAG,aAAa,CAAC,EAAE,EAAE,OAAO,EAAE,GAAG,CAAC,CAAC;QAC/C,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAExB,MAAM,OAAO,GAAa,EAAE,CAAC;QAC7B,IAAI,CAAC,MAAM,CAAC,IAAI;YAAE,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACvC,IAAI,CAAC,MAAM,CAAC,IAAI;YAAE,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QACvC,IAAI,CAAC,MAAM,CAAC,MAAM;YAAE,OAAO,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QAC3C,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACvB,OAAO,CAAC,IAAI,CAAC,cAAc,EAAE,yBAAyB,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QAC9E,CAAC;QAED,uEAAuE;QACvE,IAAI,EAAE,CAAC,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC,EAAE,CAAC;YACjC,GAAG,CAAC,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAChC,CAAC;QAED,EAAE,CAAC,cAAc,CAAC;YAChB,EAAE,EAAE,MAAM,CAAC,EAAE;YACb,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,KAAK,EAAE,MAAM,CAAC,KAAK;YACnB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,MAAM,EAAE,MAAM,CAAC,MAAM;YACrB,WAAW,EAAE,MAAM,CAAC,WAAW;YAC/B,OAAO,EAAE,MAAM,CAAC,OAAO;SACxB,CAAC,CAAC;QAEH,oCAAoC;QACpC,GAAG,CAAC,cAAc,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAE9B,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAC3B,EAAE,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;QACrC,CAAC;QAED,4EAA4E;QAC5E,EAAE,CAAC,EAAE,CAAC,OAAO,CAAC,+CAA+C,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAE9E,6EAA6E;QAC7E,gFAAgF;QAChF,8EAA8E;QAC9E,8CAA8C;QAC9C,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,aAAa,EAAE,CAAC;YACvC,EAAE,CAAC,kBAAkB,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;YACpC,EAAE,CAAC,eAAe,CAAC,GAAG,CAAC,QAAQ,EAAE,GAAG,CAAC,QAAQ,EAAE,GAAG,CAAC,IAAI,CAAC,CAAC;QAC3D,CAAC;QAED,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,YAAY,EAAE,CAAC;YACvC,EAAE,CAAC,kBAAkB,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;YACrC,EAAE,CAAC,eAAe,CAAC,IAAI,CAAC,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,SAAS,EAAE,IAAI,CAAC,OAAO,CAAC,CAAC;QAC5E,CAAC;QAED,uEAAuE;QACvE,0EAA0E;QAC1E,qCAAqC;QACrC,EAAE;QACF,oEAAoE;QACpE,0EAA0E;QAC1E,0EAA0E;QAC1E,EAAE,CAAC,EAAE,CAAC,OAAO,CAAC,0CAA0C,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QACzE,GAAG,CAAC,oBAAoB,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QACpC,gEAAgE;QAChE,GAAG,CAAC,eAAe,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;QAE/B,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,MAAM,aAAa,CAAC,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,OAAO,CAAC,CAAC;YAC9E,MAAM,WAAW,GAAG,EAAE,CAAC,EAAE,CAAC,OAAO,CAC/B,4GAA4G,CAC7G,CAAC;YACF,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;gBAC3B,MAAM,OAAO,GAAG,GAAG,MAAM,CAAC,EAAE,KAAK,KAAK,CAAC,KAAK,EAAE,CAAC;gBAC/C,WAAW,CAAC,GAAG,CACb,OAAO,EACP,MAAM,CAAC,EAAE,EACT,KAAK,CAAC,KAAK,EACX,KAAK,CAAC,OAAO,EACb,KAAK,CAAC,SAAS,EACf,KAAK,CAAC,OAAO,CACd,CAAC;gBACF,GAAG,CAAC,eAAe,CAAC,OAAO,EAAE,KAAK,CAAC,SAAS,CAAC,CAAC;YAChD,CAAC;QACH,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,OAAO,CAAC,IAAI,CAAC,mBAAmB,EAAE,KAAM,CAAW,CAAC,OAAO,EAAE,CAAC,CAAC;QACjE,CAAC;QAED,EAAE,CAAC,gBAAgB,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC;QAEpC,OAAO,EAAE,CAAC;QACV,IAAI,OAAO,GAAG,EAAE,KAAK,CAAC,EAAE,CAAC;YACvB,OAAO,CAAC,GAAG,CAAC,KAAK,OAAO,IAAI,WAAW,CAAC,MAAM,UAAU,CAAC,CAAC;QAC5D,CAAC;IACH,CAAC;IAED,qFAAqF;IACrF,8FAA8F;IAC9F,IAAI,mBAAmB,EAAE,CAAC;QACxB,GAAG,CAAC,YAAY,EAAE,CAAC;IACrB,CAAC;IAED,gGAAgG;IAChG,MAAM,YAAY,GAAG,IAAI,GAAG,CACzB,EAAE,CAAC,EAAE,CAAC,OAAO,CAAC,8CAA8C,CAAC,CAAC,GAAG,EAAmC;SAClG,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CACzB,CAAC;IAEF,uEAAuE;IACvE,IAAI,SAAS,GAAG,CAAC,CAAC;IAClB,KAAK,MAAM,QAAQ,IAAI,YAAY,EAAE,CAAC;QACpC,IAAI,CAAC,EAAE,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,CAAC;YACjC,EAAE,CAAC,kBAAkB,CAAC,QAAQ,CAAC,CAAC;YAChC,SAAS,EAAE,CAAC;QACd,CAAC;IACH,CAAC;IACD,OAAO,CAAC,GAAG,CAAC,aAAa,SAAS,sCAAsC,CAAC,CAAC;IAE1E,IAAI,CAAC;QACH,IAAI,QAAQ,IAAI,IAAI,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAChC,OAAO,CAAC,GAAG,CAAC,2BAA2B,CAAC,CAAC;YACzC,eAAe,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,UAAU,CAAC,CAAC;YACrC,OAAO,CAAC,GAAG,CAAC,wBAAwB,CAAC,CAAC;QACxC,CAAC;IACH,CAAC;YAAS,CAAC;QACT,OAAO,CAAC,GAAG,CAAC,SAAS,OAAO,uBAAuB,OAAO,uBAAuB,CAAC,CAAC;QACnF,EAAE,CAAC,KAAK,EAAE,CAAC;IACb,CAAC;AACH,CAAC;AAED,MAAM,eAAe,GAAG,IAAI,CAAC,OAAO,EAAE,EAAE,aAAa,EAAE,cAAc,CAAC,CAAC;AAYvE;;;;;;;;;;;;;GAaG;AACH,MAAM,UAAU,WAAW;IACzB,MAAM,OAAO,GAAG,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IACtC,MAAM,UAAU,GAAG,OAAO,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC;IACrD,MAAM,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC;IAC5D,MAAM,KAAK,GAAG,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC;IACtD,MAAM,OAAO,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC;IAE3D,MAAM,MAAM,GAAG,UAAU,EAAE,CAAC;IAE5B,MAAM,aAAa,GAAG,GAAW,EAAE,CACjC,KAAK;QACL,OAAO,CAAC,GAAG,CAAC,kBAAkB;QAC9B,MAAM,CAAC,MAAM;QACb,eAAe,CAAC;IAElB,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACvB,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC,CAAC;QACrC,OAAO;YACL,IAAI,EAAE,OAAO;YACb,MAAM,EAAE,aAAa,EAAE;YACvB,QAAQ,EAAE,CAAC,UAAU;YACrB,MAAM,EAAE,KAAK;YACb,MAAM;SACP,CAAC;IACJ,CAAC;IAED,MAAM,OAAO,GAAG,OAAO,CAAC,GAAG,CAAC,oBAAoB,CAAC;IACjD,IAAI,OAAO,EAAE,CAAC;QACZ,MAAM,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;QACrE,IAAI,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACtB,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC,CAAC;YACrC,OAAO;gBACL,IAAI,EAAE,MAAM;gBACZ,MAAM,EAAE,aAAa,EAAE;gBACvB,QAAQ,EAAE,CAAC,UAAU;gBACrB,MAAM,EAAE,KAAK;gBACb,MAAM;aACP,CAAC;QACJ,CAAC;IACH,CAAC;IAED,IAAI,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC5C,OAAO,CAAC,GAAG,CAAC,0BAA0B,CAAC,CAAC;QACxC,OAAO;YACL,IAAI,EAAE,MAAM,CAAC,KAAK;YAClB,MAAM,EAAE,aAAa,EAAE;YACvB,QAAQ,EAAE,CAAC,UAAU;YACrB,MAAM,EAAE,QAAQ;YAChB,MAAM;SACP,CAAC;IACJ,CAAC;IAED,OAAO,CAAC,GAAG,CAAC,4BAA4B,CAAC,CAAC;IAC1C,OAAO;QACL,IAAI,EAAE,CAAC,gBAAgB,CAAC;QACxB,MAAM,EAAE,aAAa,EAAE;QACvB,QAAQ,EAAE,CAAC,UAAU;QACrB,MAAM,EAAE,UAAU;QAClB,MAAM;KACP,CAAC;AACJ,CAAC;AAED,MAAM,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,QAAQ,CAAC,YAAY,CAAC,CAAC;AACvD,IAAI,MAAM,EAAE,CAAC;IACX,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,WAAW,EAAE,CAAC;IACzD,OAAO,CAAC,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,CAAC,cAAc,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;AAC9E,CAAC"}

package/dist/vector-search.d.ts

@@ -12,5 +12,15 @@ export declare class VectorSearch {
     dropIndex(): void;
     upsertEmbedding(id: string, embedding: Float32Array): void;
     deleteEmbedding(id: string): void;
+    /**
+     * Delete all chunk-level vec rows for a document. Chunk ids follow the
+     * pattern `${docId}#c${index}` so we match via a SQLite GLOB.
+     *
+     * This is used by reindex to drop stale chunks when a source markdown file
+     * has been deleted or modified. Complements `ON DELETE CASCADE` on the
+     * `chunks` table (which deletes chunk rows but not their vec counterparts,
+     * because the vec0 virtual table does not participate in FK cascades).
+     */
+    deleteChunkVecsByDoc(docId: string): void;
     search(queryEmbedding: Float32Array, limit?: number): VectorResult[];
 }

package/dist/vector-search.js

@@ -42,6 +42,21 @@ export class VectorSearch {
             .prepare("DELETE FROM documents_vec WHERE id = ?")
             .run(id);
     }
+    /**
+     * Delete all chunk-level vec rows for a document. Chunk ids follow the
+     * pattern `${docId}#c${index}` so we match via a SQLite GLOB.
+     *
+     * This is used by reindex to drop stale chunks when a source markdown file
+     * has been deleted or modified. Complements `ON DELETE CASCADE` on the
+     * `chunks` table (which deletes chunk rows but not their vec counterparts,
+     * because the vec0 virtual table does not participate in FK cascades).
+     */
+    deleteChunkVecsByDoc(docId) {
+        this.ensureVecLoaded();
+        this.knowledgeDb.db
+            .prepare("DELETE FROM documents_vec WHERE id GLOB ?")
+            .run(`${docId}#c*`);
+    }
     search(queryEmbedding, limit = 10) {
         this.ensureVecLoaded();
         const buf = float32ToBuffer(queryEmbedding);

package/dist/vector-search.js.map

@@ -1 +1 @@
-{"version":3,"file":"vector-search.js","sourceRoot":"","sources":["../src/vector-search.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,SAAS,MAAM,YAAY,CAAC;AAQxC,SAAS,eAAe,CAAC,GAAiB;IACxC,OAAO,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,GAAG,CAAC,UAAU,EAAE,GAAG,CAAC,UAAU,CAAC,CAAC;AACjE,CAAC;AAED,MAAM,OAAO,YAAY;IAGH;IAFZ,SAAS,GAAG,KAAK,CAAC;IAE1B,YAAoB,WAAwB;QAAxB,gBAAW,GAAX,WAAW,CAAa;IAAG,CAAC;IAExC,eAAe;QACrB,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,CAAC;YACpB,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;YACpC,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC;QACxB,CAAC;IACH,CAAC;IAED,WAAW;QACT,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,IAAI,CAAC;;;;;KAKxB,CAAC,CAAC;IACL,CAAC;IAED,SAAS;QACP,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,IAAI,CAAC,oCAAoC,CAAC,CAAC;IACjE,CAAC;IAED,eAAe,CAAC,EAAU,EAAE,SAAuB;QACjD,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,MAAM,GAAG,GAAG,eAAe,CAAC,SAAS,CAAC,CAAC;QACvC,IAAI,CAAC,WAAW,CAAC,EAAE;aAChB,OAAO,CAAC,wCAAwC,CAAC;aACjD,GAAG,CAAC,EAAE,CAAC,CAAC;QACX,IAAI,CAAC,WAAW,CAAC,EAAE;aAChB,OAAO,CAAC,yDAAyD,CAAC;aAClE,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,CAAC;IAClB,CAAC;IAED,eAAe,CAAC,EAAU;QACxB,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,IAAI,CAAC,WAAW,CAAC,EAAE;aAChB,OAAO,CAAC,wCAAwC,CAAC;aACjD,GAAG,CAAC,EAAE,CAAC,CAAC;IACb,CAAC;IAED,MAAM,CAAC,cAA4B,EAAE,QAAgB,EAAE;QACrD,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,MAAM,GAAG,GAAG,eAAe,CAAC,cAAc,CAAC,CAAC;QAC5C,OAAO,IAAI,CAAC,WAAW,CAAC,EAAE;aACvB,OAAO,CACN;;;;;KAKH,CACE;aACA,GAAG,CAAC,GAAG,EAAE,KAAK,CAAmB,CAAC;IACvC,CAAC;CACF"}
+{"version":3,"file":"vector-search.js","sourceRoot":"","sources":["../src/vector-search.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,SAAS,MAAM,YAAY,CAAC;AAQxC,SAAS,eAAe,CAAC,GAAiB;IACxC,OAAO,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,EAAE,GAAG,CAAC,UAAU,EAAE,GAAG,CAAC,UAAU,CAAC,CAAC;AACjE,CAAC;AAED,MAAM,OAAO,YAAY;IAGH;IAFZ,SAAS,GAAG,KAAK,CAAC;IAE1B,YAAoB,WAAwB;QAAxB,gBAAW,GAAX,WAAW,CAAa;IAAG,CAAC;IAExC,eAAe;QACrB,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,CAAC;YACpB,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;YACpC,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC;QACxB,CAAC;IACH,CAAC;IAED,WAAW;QACT,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,IAAI,CAAC;;;;;KAKxB,CAAC,CAAC;IACL,CAAC;IAED,SAAS;QACP,IAAI,CAAC,WAAW,CAAC,EAAE,CAAC,IAAI,CAAC,oCAAoC,CAAC,CAAC;IACjE,CAAC;IAED,eAAe,CAAC,EAAU,EAAE,SAAuB;QACjD,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,MAAM,GAAG,GAAG,eAAe,CAAC,SAAS,CAAC,CAAC;QACvC,IAAI,CAAC,WAAW,CAAC,EAAE;aAChB,OAAO,CAAC,wCAAwC,CAAC;aACjD,GAAG,CAAC,EAAE,CAAC,CAAC;QACX,IAAI,CAAC,WAAW,CAAC,EAAE;aAChB,OAAO,CAAC,yDAAyD,CAAC;aAClE,GAAG,CAAC,EAAE,EAAE,GAAG,CAAC,CAAC;IAClB,CAAC;IAED,eAAe,CAAC,EAAU;QACxB,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,IAAI,CAAC,WAAW,CAAC,EAAE;aAChB,OAAO,CAAC,wCAAwC,CAAC;aACjD,GAAG,CAAC,EAAE,CAAC,CAAC;IACb,CAAC;IAED;;;;;;;;OAQG;IACH,oBAAoB,CAAC,KAAa;QAChC,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,IAAI,CAAC,WAAW,CAAC,EAAE;aAChB,OAAO,CAAC,2CAA2C,CAAC;aACpD,GAAG,CAAC,GAAG,KAAK,KAAK,CAAC,CAAC;IACxB,CAAC;IAED,MAAM,CAAC,cAA4B,EAAE,QAAgB,EAAE;QACrD,IAAI,CAAC,eAAe,EAAE,CAAC;QACvB,MAAM,GAAG,GAAG,eAAe,CAAC,cAAc,CAAC,CAAC;QAC5C,OAAO,IAAI,CAAC,WAAW,CAAC,EAAE;aACvB,OAAO,CACN;;;;;KAKH,CACE;aACA,GAAG,CAAC,GAAG,EAAE,KAAK,CAAmB,CAAC;IACvC,CAAC;CACF"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ralph-hero-knowledge-index",
-  "version": "0.1.23",
+  "version": "0.1.24",
   "type": "module",
   "main": "dist/index.js",
   "bin": {

package/src/__tests__/embedder.test.ts

@@ -1,5 +1,20 @@
-import { describe, it, expect } from "vitest";
-import { prepareTextForEmbedding } from "../embedder.js";
+import { describe, it, expect, vi, beforeEach } from "vitest";
+
+// Mock @huggingface/transformers so we don't need to load the real ONNX
+// model during unit tests. The fake pipeline returns a constant 384-dim
+// vector; we track call count via the `embedCalls` array below.
+const embedCalls: string[] = [];
+vi.mock("@huggingface/transformers", () => {
+  const fakePipeline = async (text: string, _opts: unknown) => {
+    embedCalls.push(text);
+    return { data: new Float32Array(384) };
+  };
+  return {
+    pipeline: vi.fn(async () => fakePipeline),
+  };
+});
+
+import { prepareTextForEmbedding, embedDocument } from "../embedder.js";
 
 describe("prepareTextForEmbedding", () => {
   it("includes title, tags, and first paragraph", () => {
@@ -40,14 +55,15 @@ describe("prepareTextForEmbedding", () => {
     expect(result).not.toContain("\n\n");
   });
 
-  it("truncates at MAX_CHARS (500) total", () => {
+  it("no longer truncates at 500 chars (MAX_CHARS removed)", () => {
     const longParagraph = "A".repeat(600);
     const result = prepareTextForEmbedding(
       "Title",
       ["tag1", "tag2"],
       longParagraph,
     );
-    expect(result.length).toBe(500);
+    // Title (5) + \n + tag1, tag2 (10) + \n + 600 A's = 617 chars
+    expect(result.length).toBe(617);
     expect(result.startsWith("Title\ntag1, tag2\n")).toBe(true);
   });
 
@@ -98,3 +114,86 @@ describe("prepareTextForEmbedding", () => {
     expect(result).toBe("My Title\ngraphology, search\nFirst paragraph.");
   });
 });
+
+describe("embedDocument", () => {
+  beforeEach(() => {
+    embedCalls.length = 0;
+  });
+
+  it("returns exactly one chunk for short content", async () => {
+    const result = await embedDocument("Title", ["tag"], "short content");
+    expect(result).toHaveLength(1);
+    expect(result[0]!.index).toBe(0);
+    expect(result[0]!.content).toBe("short content");
+    expect(result[0]!.charStart).toBe(0);
+    expect(result[0]!.charEnd).toBe("short content".length);
+    expect(result[0]!.embedding).toBeInstanceOf(Float32Array);
+  });
+
+  it("embeds with title + tagLine + chunk.content prepended", async () => {
+    await embedDocument("My Title", ["graphology", "search"], "body text");
+    expect(embedCalls).toHaveLength(1);
+    expect(embedCalls[0]).toBe("My Title\ngraphology, search\nbody text");
+  });
+
+  it("omits empty title/tags/content from the embed input", async () => {
+    await embedDocument("", [], "only content here");
+    expect(embedCalls).toContain("only content here");
+
+    embedCalls.length = 0;
+    await embedDocument("Just Title", [], "");
+    // Empty content -> one chunk with empty string, only title is non-empty.
+    expect(embedCalls).toContain("Just Title");
+  });
+
+  it("yields >= 4 chunks for an 8K-char document", async () => {
+    const longContent = "A".repeat(8000);
+    const result = await embedDocument("Title", [], longContent);
+    expect(result.length).toBeGreaterThanOrEqual(4);
+    // Each chunk gets its own embedding.
+    expect(embedCalls).toHaveLength(result.length);
+  });
+
+  it("produces Float32Array embeddings of length 384", async () => {
+    const result = await embedDocument("T", [], "hello world");
+    expect(result[0]!.embedding).toBeInstanceOf(Float32Array);
+    expect(result[0]!.embedding.length).toBe(384);
+  });
+
+  it("chunk indexes are monotonically increasing from 0", async () => {
+    const longContent = "word ".repeat(3000); // ~15K chars, many chunks
+    const result = await embedDocument("T", [], longContent);
+    expect(result.length).toBeGreaterThan(1);
+    for (let i = 0; i < result.length; i++) {
+      expect(result[i]!.index).toBe(i);
+    }
+  });
+
+  it("chunk offsets reconstruct the original content", async () => {
+    const content = "A".repeat(5000);
+    const result = await embedDocument("T", [], content);
+    for (const chunk of result) {
+      expect(content.slice(chunk.charStart, chunk.charEnd)).toBe(chunk.content);
+    }
+  });
+
+  it("empty content yields one chunk with empty content (anchors on title/tags)", async () => {
+    const result = await embedDocument("Just Title", ["some-tag"], "");
+    expect(result).toHaveLength(1);
+    expect(result[0]!.content).toBe("");
+    expect(result[0]!.charStart).toBe(0);
+    expect(result[0]!.charEnd).toBe(0);
+    // Still got embedded using title + tag.
+    expect(embedCalls).toContain("Just Title\nsome-tag");
+  });
+
+  it("respects custom chunker options", async () => {
+    const content = "A".repeat(500);
+    const result = await embedDocument("T", [], content, {
+      chunkSize: 100,
+      chunkOverlap: 10,
+    });
+    // With chunkSize=100 over 500 chars, we expect multiple chunks.
+    expect(result.length).toBeGreaterThan(1);
+  });
+});

package/src/__tests__/reindex.test.ts

@@ -1,24 +1,44 @@
-import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { describe, it, expect, vi, beforeEach } from "vitest";
 import { mkdtempSync, writeFileSync, mkdirSync, unlinkSync, utimesSync } from "node:fs";
 import { join, resolve } from "node:path";
 import { tmpdir } from "node:os";
 import { findMarkdownFiles } from "../file-scanner.js";
 import { FtsSearch } from "../search.js";
+import { VectorSearch } from "../vector-search.js";
+
+// Mock embedder so we don't load the real transformer model during tests.
+// embedDocument returns one DocumentChunk per call with a constant 384-dim
+// embedding; this matches the new chunk-aware reindex flow.
+vi.mock("../embedder.js", async () => {
+  // Import the real chunker so the mock chunks content the same way as prod.
+  const { chunkText } = await import("../chunker.js");
+  return {
+    embed: vi.fn(async () => new Float32Array(384)),
+    embedDocument: vi.fn(async (_title: string, _tags: string[], content: string) => {
+      const chunks = content.length === 0
+        ? [{ index: 0, content: "", charStart: 0, charEnd: 0 }]
+        : chunkText(content);
+      return chunks.map(c => ({
+        index: c.index,
+        content: c.content,
+        charStart: c.charStart,
+        charEnd: c.charEnd,
+        embedding: new Float32Array(384),
+      }));
+    }),
+    prepareTextForEmbedding: vi.fn((title: string, tags: string[], content: string) => {
+      const tagLine = tags.length > 0 ? tags.join(", ") : "";
+      const parts = [title, tagLine, content].filter(p => p.length > 0);
+      return parts.join("\n");
+    }),
+  };
+});
 
-vi.mock("../embedder.js", () => ({
-  embed: vi.fn(async () => new Float32Array(384)),
-  prepareTextForEmbedding: vi.fn((title: string, tags: string[], content: string) => {
-    const tagLine = tags.length > 0 ? tags.join(", ") : "";
-    const parts = [title, tagLine, content].filter(p => p.length > 0);
-    return parts.join("\n").slice(0, 500);
-  }),
-}));
-
-import { embed } from "../embedder.js";
-import { reindex, resolveDirs } from "../reindex.js";
+import { embedDocument } from "../embedder.js";
+import { reindex } from "../reindex.js";
 import { KnowledgeDB } from "../db.js";
 
-const mockedEmbed = vi.mocked(embed);
+const mockedEmbed = vi.mocked(embedDocument);
 
 function makeDoc(title: string): string {
   return `---\ndate: 2026-03-24\ntype: research\nstatus: draft\n---\n\n# ${title}\n\nContent for ${title}.`;
@@ -354,149 +374,159 @@ describe("incremental reindex", () => {
     db1.close();
   });
 
-  it("scenario 13: reindex honors .ralphignore for file discovery", async () => {
-    writeFileSync(join(dir, "kept.md"), makeDoc("Kept"));
-    writeFileSync(join(dir, "skipped.md"), makeDoc("Skipped"));
-    writeFileSync(join(dir, ".ralphignore"), "skipped.md\n");
+  it("scenario 13: 8K-char document produces >= 4 chunk rows", async () => {
+    const longBody = "A".repeat(8000);
+    writeFileSync(
+      join(dir, "long-doc.md"),
+      `---\ndate: 2026-03-24\ntype: research\nstatus: draft\n---\n\n# Long Doc\n\n${longBody}`,
+    );
 
     await reindex([dir], dbPath);
-    expect(mockedEmbed).toHaveBeenCalledTimes(1);
 
     const db = new KnowledgeDB(dbPath);
-    expect(db.getDocument("kept")).toBeTruthy();
-    expect(db.getDocument("skipped")).toBeUndefined();
+    const row = db.db
+      .prepare("SELECT COUNT(*) as n FROM chunks WHERE document_id = ?")
+      .get("long-doc") as { n: number };
+    expect(row.n).toBeGreaterThanOrEqual(4);
     db.close();
   });
 
-  it("scenario 14: reindex honors caller-supplied ignorePatterns arg", async () => {
-    writeFileSync(join(dir, "kept.md"), makeDoc("Kept"));
-    mkdirSync(join(dir, "drafts"));
-    writeFileSync(join(dir, "drafts", "wip.md"), makeDoc("WIP"));
+  it("scenario 14: documents_vec row count equals total chunk count", async () => {
+    writeFileSync(join(dir, "doc-a.md"), makeDoc("Doc A"));
+    writeFileSync(join(dir, "doc-b.md"), makeDoc("Doc B"));
+    const longBody = "A".repeat(6000);
+    writeFileSync(
+      join(dir, "long-doc.md"),
+      `---\ndate: 2026-03-24\ntype: research\nstatus: draft\n---\n\n# Long Doc\n\n${longBody}`,
+    );
 
-    await reindex([dir], dbPath, false, ["drafts/**"]);
-    // Only kept.md should have been embedded.
-    expect(mockedEmbed).toHaveBeenCalledTimes(1);
+    await reindex([dir], dbPath);
 
     const db = new KnowledgeDB(dbPath);
-    expect(db.getDocument("kept")).toBeTruthy();
-    expect(db.getDocument("wip")).toBeUndefined();
+    // Instantiating VectorSearch loads sqlite-vec so documents_vec is queryable.
+    new VectorSearch(db).createIndex();
+    const chunksRow = db.db.prepare("SELECT COUNT(*) as n FROM chunks").get() as {
+      n: number;
+    };
+    const vecRow = db.db
+      .prepare("SELECT COUNT(*) as n FROM documents_vec")
+      .get() as { n: number };
+    expect(vecRow.n).toBe(chunksRow.n);
+    expect(chunksRow.n).toBeGreaterThanOrEqual(3); // at least one per doc
     db.close();
   });
-});
 
-describe("resolveDirs precedence", () => {
-  const ORIGINAL_ARGV = process.argv;
-  const ORIGINAL_ENV = {
-    RALPH_KNOWLEDGE_DIRS: process.env.RALPH_KNOWLEDGE_DIRS,
-    RALPH_KNOWLEDGE_DB: process.env.RALPH_KNOWLEDGE_DB,
-    RALPH_KNOWLEDGE_CONFIG: process.env.RALPH_KNOWLEDGE_CONFIG,
-  };
-  let tmpHome: string;
-  let configDir: string;
+  it("scenario 15: chunk ids follow pattern {docId}#c{index}", async () => {
+    const longBody = "A".repeat(6000);
+    writeFileSync(
+      join(dir, "long-doc.md"),
+      `---\ndate: 2026-03-24\ntype: research\nstatus: draft\n---\n\n# Long Doc\n\n${longBody}`,
+    );
 
-  beforeEach(() => {
-    process.argv = ["node", "reindex.js"];
-    delete process.env.RALPH_KNOWLEDGE_DIRS;
-    delete process.env.RALPH_KNOWLEDGE_DB;
-    configDir = mkdtempSync(join(tmpdir(), "resolve-dirs-"));
-    tmpHome = configDir;
-    process.env.RALPH_KNOWLEDGE_CONFIG = join(configDir, "knowledge.config.json");
-  });
+    await reindex([dir], dbPath);
 
-  afterEach(() => {
-    process.argv = ORIGINAL_ARGV;
-    for (const key of Object.keys(ORIGINAL_ENV) as (keyof typeof ORIGINAL_ENV)[]) {
-      const orig = ORIGINAL_ENV[key];
-      if (orig === undefined) {
-        delete process.env[key];
-      } else {
-        process.env[key] = orig;
-      }
+    const db = new KnowledgeDB(dbPath);
+    new VectorSearch(db).createIndex();
+    const rows = db.db
+      .prepare("SELECT id, chunk_index FROM chunks WHERE document_id = ? ORDER BY chunk_index")
+      .all("long-doc") as Array<{ id: string; chunk_index: number }>;
+    expect(rows.length).toBeGreaterThan(1);
+    const idPattern = /^long-doc#c\d+$/;
+    for (const r of rows) {
+      expect(r.id).toMatch(idPattern);
+      expect(r.id).toBe(`long-doc#c${r.chunk_index}`);
     }
+    // Verify documents_vec ids also follow the pattern for this doc.
+    const vecRows = db.db
+      .prepare("SELECT id FROM documents_vec WHERE id GLOB ?")
+      .all("long-doc#c*") as Array<{ id: string }>;
+    expect(vecRows.length).toBe(rows.length);
+    for (const v of vecRows) {
+      expect(v.id).toMatch(idPattern);
+    }
+    db.close();
   });
 
-  it("CLI positional args beat env var even when both are set", () => {
+  it("scenario 16: deleting source file removes its chunks and vec rows", async () => {
+    const filePath = join(dir, "disposable.md");
+    const longBody = "A".repeat(6000);
     writeFileSync(
-      process.env.RALPH_KNOWLEDGE_CONFIG!,
-      JSON.stringify({ roots: ["/from/config"] }),
+      filePath,
+      `---\ndate: 2026-03-24\ntype: research\nstatus: draft\n---\n\n# Disposable\n\n${longBody}`,
     );
-    process.argv = ["node", "reindex.js", "/from/cli"];
-    process.env.RALPH_KNOWLEDGE_DIRS = "/from/env";
-    const r = resolveDirs();
-    expect(r.source).toBe("cli");
-    expect(r.dirs).toEqual(["/from/cli"]);
-  });
+    writeFileSync(join(dir, "keeper.md"), makeDoc("Keeper"));
 
-  it("env var beats config file roots when CLI is empty", () => {
-    writeFileSync(
-      process.env.RALPH_KNOWLEDGE_CONFIG!,
-      JSON.stringify({ roots: ["/from/config"] }),
-    );
-    process.env.RALPH_KNOWLEDGE_DIRS = "/from/env-a,/from/env-b";
-    const r = resolveDirs();
-    expect(r.source).toBe("env");
-    expect(r.dirs).toEqual(["/from/env-a", "/from/env-b"]);
-  });
+    await reindex([dir], dbPath);
 
-  it("config file roots beat fallback when CLI and env are absent", () => {
-    writeFileSync(
-      process.env.RALPH_KNOWLEDGE_CONFIG!,
-      JSON.stringify({ roots: ["/from/config-a", "/from/config-b"] }),
-    );
-    const r = resolveDirs();
-    expect(r.source).toBe("config");
-    expect(r.dirs).toEqual(["/from/config-a", "/from/config-b"]);
-  });
+    const db1 = new KnowledgeDB(dbPath);
+    new VectorSearch(db1).createIndex();
+    const chunksBefore = db1.db
+      .prepare("SELECT COUNT(*) as n FROM chunks WHERE document_id = ?")
+      .get("disposable") as { n: number };
+    expect(chunksBefore.n).toBeGreaterThan(1);
+    const vecsBefore = db1.db
+      .prepare("SELECT COUNT(*) as n FROM documents_vec WHERE id GLOB ?")
+      .get("disposable#c*") as { n: number };
+    expect(vecsBefore.n).toBe(chunksBefore.n);
+    db1.close();
 
-  it("falls back to ../../thoughts when no source is configured", () => {
-    // Point env var at a nonexistent config path so loadConfig returns {}.
-    process.env.RALPH_KNOWLEDGE_CONFIG = join(configDir, "missing.json");
-    const r = resolveDirs();
-    expect(r.source).toBe("fallback");
-    expect(r.dirs).toEqual(["../../thoughts"]);
-  });
+    unlinkSync(filePath);
+    await reindex([dir], dbPath);
 
-  it("dbPath precedence: CLI arg > env var > config > default", () => {
-    writeFileSync(
-      process.env.RALPH_KNOWLEDGE_CONFIG!,
-      JSON.stringify({ roots: ["/x"], dbPath: "/from/config.db" }),
-    );
-    // CLI wins
-    process.argv = ["node", "reindex.js", "/cli/root", "/cli/override.db"];
-    process.env.RALPH_KNOWLEDGE_DB = "/from/env.db";
-    expect(resolveDirs().dbPath).toBe("/cli/override.db");
-
-    // Env wins over config when CLI is absent
-    process.argv = ["node", "reindex.js"];
-    process.env.RALPH_KNOWLEDGE_DIRS = "/env/root";
-    process.env.RALPH_KNOWLEDGE_DB = "/from/env.db";
-    expect(resolveDirs().dbPath).toBe("/from/env.db");
-
-    // Config wins when neither CLI nor env set dbPath
-    delete process.env.RALPH_KNOWLEDGE_DB;
-    expect(resolveDirs().dbPath).toBe("/from/config.db");
+    const db2 = new KnowledgeDB(dbPath);
+    new VectorSearch(db2).createIndex();
+    // Document gone -> chunks cascaded.
+    expect(db2.getDocument("disposable")).toBeUndefined();
+    const chunksAfter = db2.db
+      .prepare("SELECT COUNT(*) as n FROM chunks WHERE document_id = ?")
+      .get("disposable") as { n: number };
+    expect(chunksAfter.n).toBe(0);
+    // Vec rows for the deleted doc are gone (GLOB-based cleanup).
+    const vecsAfter = db2.db
+      .prepare("SELECT COUNT(*) as n FROM documents_vec WHERE id GLOB ?")
+      .get("disposable#c*") as { n: number };
+    expect(vecsAfter.n).toBe(0);
+    // The kept doc still has its chunks.
+    const keeperChunks = db2.db
+      .prepare("SELECT COUNT(*) as n FROM chunks WHERE document_id = ?")
+      .get("keeper") as { n: number };
+    expect(keeperChunks.n).toBeGreaterThanOrEqual(1);
+    db2.close();
   });
 
-  it("forwards config.ignorePatterns on the returned config object", () => {
+  it("scenario 17: re-indexing same file does not duplicate chunks", async () => {
+    const filePath = join(dir, "stable.md");
+    const body = "A".repeat(6000);
     writeFileSync(
-      process.env.RALPH_KNOWLEDGE_CONFIG!,
-      JSON.stringify({
-        roots: ["/r1"],
-        ignorePatterns: ["draft/**", "*.bak"],
-      }),
+      filePath,
+      `---\ndate: 2026-03-24\ntype: research\nstatus: draft\n---\n\n# Stable\n\n${body}`,
     );
-    const r = resolveDirs();
-    expect(r.config.ignorePatterns).toEqual(["draft/**", "*.bak"]);
-  });
 
-  it("treats an empty RALPH_KNOWLEDGE_DIRS as unset and falls through", () => {
-    writeFileSync(
-      process.env.RALPH_KNOWLEDGE_CONFIG!,
-      JSON.stringify({ roots: ["/from/config"] }),
-    );
-    process.env.RALPH_KNOWLEDGE_DIRS = "  ,  ";
-    const r = resolveDirs();
-    expect(r.source).toBe("config");
-    expect(r.dirs).toEqual(["/from/config"]);
+    await reindex([dir], dbPath);
+    const db1 = new KnowledgeDB(dbPath);
+    const firstCount = (db1.db
+      .prepare("SELECT COUNT(*) as n FROM chunks WHERE document_id = ?")
+      .get("stable") as { n: number }).n;
+    db1.close();
+    expect(firstCount).toBeGreaterThan(1);
+
+    // Bump mtime to force re-embed.
+    const future = Date.now() / 1000 + 2;
+    utimesSync(filePath, future, future);
+
+    await reindex([dir], dbPath);
+    const db2 = new KnowledgeDB(dbPath);
+    new VectorSearch(db2).createIndex();
+    const secondCount = (db2.db
+      .prepare("SELECT COUNT(*) as n FROM chunks WHERE document_id = ?")
+      .get("stable") as { n: number }).n;
+    // Stale deletion before insert means chunk count stays the same, not 2x.
+    expect(secondCount).toBe(firstCount);
+    // And vec rows should match.
+    const vecCount = (db2.db
+      .prepare("SELECT COUNT(*) as n FROM documents_vec WHERE id GLOB ?")
+      .get("stable#c*") as { n: number }).n;
+    expect(vecCount).toBe(secondCount);
+    db2.close();
   });
 });

package/src/embedder.ts

@@ -2,9 +2,9 @@ import {
   pipeline,
   type FeatureExtractionPipeline,
 } from "@huggingface/transformers";
+import { chunkText, type Chunk, type ChunkerOptions } from "./chunker.js";
 
 const MODEL_ID = "Xenova/all-MiniLM-L6-v2";
-const MAX_CHARS = 500;
 
 let embedderInstance: FeatureExtractionPipeline | null = null;
 
@@ -21,14 +21,71 @@ export async function getEmbedder(): Promise<FeatureExtractionPipeline> {
 
 export async function embed(text: string): Promise<Float32Array> {
   const embedder = await getEmbedder();
-  const truncated = text.slice(0, MAX_CHARS);
-  const output = await embedder(truncated, {
+  // Pass text directly — the transformer's own 512-token window handles overflow.
+  const output = await embedder(text, {
     pooling: "mean",
     normalize: true,
   });
   return new Float32Array(output.data as ArrayLike<number>);
 }
 
+/**
+ * A chunk paired with the embedding of its (contextualized) content.
+ * Extends the base Chunk from the chunker module with an embedding vector
+ * and an optional contextPrefix (populated by Phase 6 — contextual retrieval).
+ */
+export interface DocumentChunk extends Chunk {
+  embedding: Float32Array;
+  contextPrefix?: string;
+}
+
+/**
+ * Embed a document by splitting it into chunks and emitting one embedding
+ * per chunk. The embedded text for each chunk is
+ * `${title}\n${tagLine}\n${chunk.content}` so the semantic anchors (title +
+ * tags) travel with every chunk embedding — matching the shape of the legacy
+ * `prepareTextForEmbedding()` but without the 500-char truncation.
+ *
+ * Short documents (<= chunkSize) produce exactly one chunk covering the whole
+ * content. Empty content yields a single chunk with empty content (so callers
+ * still get a title/tag-only embedding for stub documents).
+ */
+export async function embedDocument(
+  title: string,
+  tags: string[],
+  content: string,
+  opts?: ChunkerOptions,
+): Promise<DocumentChunk[]> {
+  const tagLine = tags.length > 0 ? tags.join(", ") : "";
+
+  // If content is empty, still emit one chunk so the document has a searchable
+  // embedding anchored on title + tags (preserves legacy behavior for
+  // frontmatter-only / stub documents).
+  const chunks: Chunk[] = content.length === 0
+    ? [{ index: 0, content: "", charStart: 0, charEnd: 0 }]
+    : chunkText(content, opts);
+
+  const out: DocumentChunk[] = [];
+  for (const chunk of chunks) {
+    const parts = [title, tagLine, chunk.content].filter(p => p.length > 0);
+    const embedText = parts.join("\n");
+    const embedding = await embed(embedText);
+    out.push({
+      index: chunk.index,
+      content: chunk.content,
+      charStart: chunk.charStart,
+      charEnd: chunk.charEnd,
+      embedding,
+    });
+  }
+  return out;
+}
+
+/**
+ * Back-compat shim: kept so callers outside the reindex path can still build
+ * a title/tags/first-paragraph string. No longer used by `embedDocument` (the
+ * per-chunk flow prepends title + tags directly).
+ */
 export function prepareTextForEmbedding(
   title: string,
   tags: string[],
@@ -39,5 +96,5 @@ export function prepareTextForEmbedding(
   const paragraphs = content.split(/\n\n+/);
   const firstParagraph = paragraphs.find(p => p.trim().length > 0)?.trim() ?? "";
   const parts = [title, tagLine, firstParagraph].filter(p => p.length > 0);
-  return parts.join("\n").slice(0, MAX_CHARS);
+  return parts.join("\n");
 }

package/src/reindex.ts

@@ -4,7 +4,7 @@ import { homedir } from "node:os";
 import { KnowledgeDB } from "./db.js";
 import { FtsSearch } from "./search.js";
 import { VectorSearch } from "./vector-search.js";
-import { embed, prepareTextForEmbedding } from "./embedder.js";
+import { embedDocument } from "./embedder.js";
 import { parseDocument, type ParsedDocument } from "./parser.js";
 import { findMarkdownFiles } from "./file-scanner.js";
 import { generateIndexes } from "./generate-indexes.js";
@@ -48,7 +48,10 @@ export async function reindex(
 
   const filesOnDiskSet = new Set(filesOnDisk.map(f => resolve(f)));
 
-  // Phase 1: Delete stale entries for files no longer on disk
+  // Phase 1: Delete stale entries for files no longer on disk.
+  // Chunk rows cascade from documents via ON DELETE CASCADE on chunks.document_id,
+  // but the vec0 virtual table does not participate in FK cascades — we must
+  // explicitly delete chunk-level vec rows via GLOB pattern.
   const syncedPaths = db.getAllSyncPaths();
   let deleted = 0;
   for (const syncedPath of syncedPaths) {
@@ -56,6 +59,8 @@ export async function reindex(
       const id = basename(syncedPath, ".md");
       fts.deleteFtsEntry(id);
       db.deleteDocument(id);
+      vec.deleteChunkVecsByDoc(id);
+      // Also delete any legacy doc-level vec row (pre-chunks schema).
       vec.deleteEmbedding(id);
       db.deleteSyncRecord(syncedPath);
       deleted++;
@@ -138,10 +143,35 @@ export async function reindex(
       db.addRelationship(edge.sourceId, edge.targetId, "untyped", edge.context);
     }
 
-    const text = prepareTextForEmbedding(parsed.title, parsed.tags, parsed.content);
+    // Chunk-aware embedding: emit one embedding per chunk, persist to both
+    // the `chunks` table and the `documents_vec` virtual table with chunk ids
+    // of the form `${doc.id}#c${index}`.
+    //
+    // We first clear any stale chunk rows for this doc_id (the document
+    // body may have shrunk across re-indexes) and stale chunk vec rows (which
+    // don't cascade from the `chunks` table because vec0 is a virtual table).
+    db.db.prepare("DELETE FROM chunks WHERE document_id = ?").run(parsed.id);
+    vec.deleteChunkVecsByDoc(parsed.id);
+    // Drop any pre-chunks schema vec row that used the bare doc id.
+    vec.deleteEmbedding(parsed.id);
+
     try {
-      const embedding = await embed(text);
-      vec.upsertEmbedding(parsed.id, embedding);
+      const chunks = await embedDocument(parsed.title, parsed.tags, parsed.content);
+      const insertChunk = db.db.prepare(
+        "INSERT INTO chunks (id, document_id, chunk_index, content, char_start, char_end) VALUES (?, ?, ?, ?, ?, ?)"
+      );
+      for (const chunk of chunks) {
+        const chunkId = `${parsed.id}#c${chunk.index}`;
+        insertChunk.run(
+          chunkId,
+          parsed.id,
+          chunk.index,
+          chunk.content,
+          chunk.charStart,
+          chunk.charEnd,
+        );
+        vec.upsertEmbedding(chunkId, chunk.embedding);
+      }
     } catch (e) {
       console.warn(`Failed to embed ${id}: ${(e as Error).message}`);
     }

package/src/vector-search.ts

@@ -54,6 +54,22 @@ export class VectorSearch {
       .run(id);
   }
 
+  /**
+   * Delete all chunk-level vec rows for a document. Chunk ids follow the
+   * pattern `${docId}#c${index}` so we match via a SQLite GLOB.
+   *
+   * This is used by reindex to drop stale chunks when a source markdown file
+   * has been deleted or modified. Complements `ON DELETE CASCADE` on the
+   * `chunks` table (which deletes chunk rows but not their vec counterparts,
+   * because the vec0 virtual table does not participate in FK cascades).
+   */
+  deleteChunkVecsByDoc(docId: string): void {
+    this.ensureVecLoaded();
+    this.knowledgeDb.db
+      .prepare("DELETE FROM documents_vec WHERE id GLOB ?")
+      .run(`${docId}#c*`);
+  }
+
   search(queryEmbedding: Float32Array, limit: number = 10): VectorResult[] {
     this.ensureVecLoaded();
     const buf = float32ToBuffer(queryEmbedding);