fix: drop voy-search and the WASM module it distributed (#228)

.gitignore

@@ -2,6 +2,7 @@
 node_modules/
 vendor/
 wp-agentic-admin.zip
+agentic-admin.zip
 *.pdf
 tests/e2e/screenshots/
 tests/e2e/RESULTS.md

docs/RAG-CODEBASE-SEARCH.md

@@ -8,7 +8,7 @@ The RAG (Retrieval-Augmented Generation) system lets the LLM answer questions ab
 
 1. **Extracting** code from your active theme and plugins (PHP backend)
 2. **Embedding** code chunks into vectors using Transformers.js (CPU/WASM)
-3. **Storing** vectors in a Voy search index, persisted in IndexedDB
+3. **Storing** vectors in IndexedDB, as plain `Float32Array`s
 4. **Searching** with semantic similarity when users ask about code
 
 All processing happens locally — no code leaves the browser.
@@ -22,7 +22,7 @@ User: "find the login function"
   ↓
 [Transformers.js] → embed query on CPU/WASM
   ↓
-[Voy index] → nearest neighbor search across 1000+ code chunks
+[vector index] → cosine similarity scan across 1000+ code chunks
   ↓
 [Results] → top 3 matching code snippets with file paths + line numbers
   ↓
@@ -43,9 +43,16 @@ User: "find the login function"
 | Dependency | How loaded | Size | Purpose |
 |------------|-----------|------|---------|
 | [Transformers.js v3](https://huggingface.co/docs/transformers.js) | CDN (lazy, on first use) | ~100MB + 23MB model | Text embeddings |
-| [voy-search](https://github.com/tantaraio/voy) | Bundled via npm | ~168KB WASM | Vector nearest-neighbor search |
 | IndexedDB | Browser native | — | Persist index across sessions |
 
+**Why no vector-search library?** Search is an exhaustive cosine scan written in
+plain JavaScript in `vector-store.js`. The embedding model emits L2-normalised
+vectors, so cosine similarity is a dot product; at 384 dimensions a few thousand
+chunks score in single-digit milliseconds, far below the cost of embedding the
+query itself. An ANN index buys nothing at this scale, and dropping it removed
+the only WebAssembly binary the plugin distributed — every shipped file now has
+readable source in this repository.
+
 **Why CDN for Transformers.js?** At ~100MB it would 17x the current 5.8MB bundle. Lazy-loading from CDN means zero cost until RAG is actually used, and the model is cached by the browser after first download.
 
 **Why CPU for embeddings?** The LLM already uses ~1.5GB VRAM via WebGPU. Running embeddings on GPU too would risk OOM. WASM/CPU is slower but avoids contention entirely.
@@ -68,14 +75,14 @@ User: "index codebase"
 
 ```
 User: "search code for authentication"
-→ Embeds query, searches Voy index
+→ Embeds query, scores it against every indexed vector
 → Returns top 3 matching code snippets
 → LLM summarizes the results
 ```
 
 ### The index persists
 
-After indexing once, the Voy index is restored from IndexedDB on page reload. No need to re-index unless your code changes.
+After indexing once, the index is restored from IndexedDB on page reload. No need to re-index unless your code changes.
 
 To rebuild: say **"reindex the codebase"**.
 
@@ -120,7 +127,7 @@ Returns 50 files per page. The JS ability paginates automatically until `has_mor
 
 ### Persistence
 
-- Serialized Voy index stored in IndexedDB (`wp-agentic-rag-db`)
+- Vectors stored as `Float32Array`s in IndexedDB (`wp-agentic-rag-db`, store `embedding-index`)
 - Chunk metadata (path, lines, content, type) stored alongside
 - Restored automatically on `vectorStore.init()`
 

package.json

@@ -8,8 +8,8 @@
     "test:abilities": "node tests/abilities/runner.js",
     "test:e2e": "node tests/abilities/e2e-runner.js",
     "watch": "wp-scripts start src/extensions/index.js --output-path=build-extensions",
-    "clean": "del-cli build-extensions wp-agentic-admin.zip",
-    "dist": "npm run clean && npm run build && rm -f wp-agentic-admin.zip && zip -r wp-agentic-admin.zip -@ < .distpackage",
+    "clean": "del-cli build-extensions agentic-admin.zip wp-agentic-admin.zip",
+    "dist": "npm run clean && npm run build && zip -r agentic-admin.zip -@ < .distpackage",
     "test": "wp-scripts test-unit-js --testMatch='**/__tests__/**/*.test.js'",
     "test:watch": "npm run test -- --watch",
     "lint:js": "wp-scripts lint-js src/",
@@ -33,8 +33,7 @@
     "@huggingface/transformers": "^3.8.1",
     "@mlc-ai/web-llm": "^0.2.82",
     "@wordpress/icons": "^12.0.0",
-    "fastest-levenshtein": "^1.0.16",
-    "voy-search": "^0.6.3"
+    "fastest-levenshtein": "^1.0.16"
   },
   "keywords": [
     "wordpress",

readme.txt

@@ -83,7 +83,7 @@ Agentic Admin is fully open source under GPL-2.0-or-later. The complete, human-r
 
 https://github.com/pluginslab/wp-agentic-admin
 
-The files under `build-extensions/` are generated from the sources in `src/` with @wordpress/scripts (webpack). They include the JavaScript bundles (`index.js`, `sw.js`, `indexing-worker.js`, and code-split chunks), the CSS, and the voy-search vector-index WebAssembly module (`*.module.wasm`, built from its npm package; source: https://github.com/tantaraio/voy). To regenerate them from a checkout:
+The files under `build-extensions/` are generated from the sources in `src/` with @wordpress/scripts (webpack). They contain only JavaScript and CSS: the bundles (`index.js`, `sw.js`, `indexing-worker.js`, and code-split chunks) and the stylesheets. No WebAssembly, binaries, or other compiled artifacts are distributed with the plugin. To regenerate them from a checkout:
 
 1. `npm install`
 2. `npm run build`
@@ -107,6 +107,9 @@ There is no build step for the PHP. The WebLLM engine is bundled into the plugin
 * Improved: ChatInput keyboard handling simplified (Space inserts a space, no push-to-talk hijacking).
 * Improved: KB embedding moved to a Web Worker with persistent progress across tab switches.
 * Pinned: Transformers.js CDN URL to @3.8.1 (was floating @3 range), privacy-first plugin shouldn't depend on a CDN range that can ship new code without a deliberate bump.
+* Removed: the voy-search dependency and the WebAssembly module it distributed. Vector search is now plain JavaScript (exhaustive cosine over L2-normalised embeddings), so the plugin ships no compiled binaries at all and every distributed file has readable source in the repository.
+* Fixed: re-running the knowledge base index no longer leaves vectors and chunk metadata misaligned, which could return the wrong code chunk for a query.
+* Fixed: `.well-known` scanning resolves via get_home_path() instead of ABSPATH, so subdirectory installs scan the real site root.
 * Removed: 7 stale tab references and 6+ stale docs files (FEEDBACK-DEV.md).
 * Tests: 96 unit tests passing, plus the new manifest test suite (7 cases), index test suite (6 cases), knowledge-base test suite (16 cases), and react-agent regression tests (3 cases for the per-call state cleanup fix).
 

src/extensions/abilities/code-search.js

@@ -96,7 +96,7 @@ export function registerCodeSearch() {
 
 		execute: async ( params ) => {
 			try {
-				// Initialize vector store (loads Voy, restores from IndexedDB).
+				// Initialize vector store (restores index from IndexedDB).
 				await vectorStore.init();
 
 				if ( ! vectorStore.isReady() ) {

src/extensions/services/indexing-worker.js

@@ -3,7 +3,7 @@
  *
  * Runs Transformers.js embedding inside a Web Worker to avoid blocking
  * the main thread during knowledge base builds. Only handles the slow
- * part (neural network inference); the main thread builds the Voy index
+ * part (neural network inference); the main thread builds the index
  * and persists to IndexedDB.
  *
  * Message protocol:

src/extensions/services/knowledge-base.js

@@ -106,7 +106,7 @@ function clearKBStatus() {
 /* ── Worker helper ───────────────────────────────────────────────────── */
 
 /**
- * Embed chunks in a Web Worker, then build Voy index on main thread.
+ * Embed chunks in a Web Worker, then build the index on the main thread.
  *
  * @param {Object[]} chunks     Chunks to index.
  * @param {Function} onProgress Callback: (done, total, message) => void.
@@ -133,7 +133,7 @@ function indexInWorker( chunks, onProgress ) {
 			} else if ( msg.type === 'complete' ) {
 				worker.terminate();
 
-				// Build Voy index on main thread (fast, < 1s).
+				// Build the index on the main thread (fast, < 1s).
 				try {
 					onProgress(
 						msg.chunkMetadata.length,