feat: replace Ollama with node-llama-cpp for embeddings

package.json

@@ -1,6 +1,6 @@
 {
   "name": "traul",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "description": "Personal Intelligence Engine — watches communication streams, identifies patterns, surfaces actionable insights",
   "license": "AGPL-3.0-only",
   "repository": {
@@ -18,13 +18,17 @@
   },
   "dependencies": {
     "@slack/web-api": "^7.9.1",
-"commander": "^13.1.0",
+    "commander": "^13.1.0",
     "googleapis": "^171.4.0",
     "html-to-text": "^9.0.5",
+    "node-llama-cpp": "^3.18.1",
     "sqlite-vec": "^0.1.7-alpha.2"
   },
   "devDependencies": {
     "@types/bun": "latest",
     "@types/html-to-text": "^9.0.4"
-  }
+  },
+  "trustedDependencies": [
+    "node-llama-cpp"
+  ]
 }

skill.md

@@ -14,7 +14,7 @@ allowed-tools:
 
 CLI tool that watches communication streams (Slack, Telegram, Discord, Linear, Gmail, Claude Code sessions, Markdown files, WhatsApp), indexes messages, detects patterns via signals, and surfaces actionable insights.
 
-**Runtime:** Bun + TypeScript | **DB:** SQLite (WAL mode, FTS5, sqlite-vec) | **Embeddings:** Ollama + nomic-embed-text | **Version:** 0.1.0
+**Runtime:** Bun + TypeScript | **DB:** SQLite (WAL mode, FTS5, sqlite-vec) | **Embeddings:** node-llama-cpp (Qwen3-Embedding-0.6B), Ollama fallback | **Version:** 0.2.0
 
 **Project:** `/Users/dandaka/projects/traul`
 
@@ -39,10 +39,10 @@ Sync messages from communication sources incrementally.
 
 ### `traul search <query>`
 
-Hybrid search combining vector similarity (semantic) and FTS5 keyword matching with Reciprocal Rank Fusion. Falls back to FTS-only if Ollama is unavailable.
+Hybrid search combining vector similarity (semantic) and FTS5 keyword matching with Reciprocal Rank Fusion. Falls back to FTS-only if embedding is unavailable.
 
 **Search modes:**
-- **Hybrid (default)** — best for multi-word and exploratory queries. Finds semantically related messages even when exact keywords don't appear. Requires Ollama running with `snowflake-arctic-embed2`. Prints coverage ratio to stderr (e.g. `88% vector, 12% FTS`). Falls back to FTS-only with a warning if Ollama is unavailable.
+- **Hybrid (default)** — best for multi-word and exploratory queries. Finds semantically related messages even when exact keywords don't appear. Uses node-llama-cpp with Qwen3-Embedding-0.6B (auto-downloads ~639MB model on first use). Falls back to Ollama, then FTS-only. Prints coverage ratio to stderr (e.g. `88% vector, 12% FTS`).
 - **FTS-only (`--fts`)** — keyword matching with BM25 ranking. Faster, but requires ALL terms to match (implicit AND). Brittle with multi-word queries, especially combined with source/channel filters.
 - **OR mode (`--or`)** — joins search terms with OR instead of AND. Works with both `--fts` and hybrid. Use for broad exploratory queries where any term is relevant.
 - **Substring (`--like`)** — bypasses FTS entirely, uses SQL LIKE. Useful for exact phrases that FTS tokenization breaks (e.g. "how do I").
@@ -163,6 +163,19 @@ Structured overview with three sections:
 2. **Stats** — total messages, channels, contacts, active signals
 3. **Volume** — last 7 days message bar chart
 
+### `traul reset`
+
+Reset a data layer to force regeneration. Useful when you need to re-sync, re-chunk, or re-embed data.
+
+| Subcommand | Description |
+|------------|-------------|
+| `traul reset sync [--source <source>]` | Clear sync cursors; full refetch on next sync. Optional `--source` flag filters to a specific connector (e.g., `markdown`, `slack`). |
+| `traul reset chunks` | Delete all chunks and embeddings; rechunk on next sync. |
+| `traul reset embed` | Drop and recreate vector tables; re-embed with `traul embed`. |
+| `traul reset all` | Reset everything: sync cursors + chunks + embeddings. |
+
+**Auto-migration:** Traul automatically detects version changes on startup. If the chunking algorithm or embedding model/dimensions change between versions, affected data layers are reset automatically. No manual action needed after upgrading.
+
 ### Global Options
 
 | Option | Description |
@@ -286,7 +299,7 @@ SQL-based pattern detection engine.
 |-------|---------|
 | `messages` | Primary message store (source, channel, author, content, sent_at, metadata JSON) |
 | `messages_fts` | FTS5 virtual table (content, author_name, channel_name) with porter tokenizer |
-| `vec_messages` | sqlite-vec virtual table for vector embeddings (float[768]) |
+| `vec_messages` | sqlite-vec virtual table for vector embeddings (float[1024]) |
 | `contacts` | Unified contact directory (display_name unique) |
 | `contact_identities` | Multi-source user mapping (source + source_user_id unique) |
 | `sync_cursors` | Incremental sync state per source+key |

src/commands/reset.ts

@@ -0,0 +1,36 @@
+import type { TraulDB } from "../db/database";
+import { EMBED_DIMS } from "../lib/embeddings";
+
+type Layer = "sync" | "chunks" | "embed" | "all";
+
+const VALID_LAYERS: Layer[] = ["sync", "chunks", "embed", "all"];
+
+export function runReset(
+  db: TraulDB,
+  layer: string,
+  options: { source?: string }
+): void {
+  if (!VALID_LAYERS.includes(layer as Layer)) {
+    throw new Error(`Unknown layer: ${layer}. Valid layers: ${VALID_LAYERS.join(", ")}`);
+  }
+
+  const doSync = layer === "sync" || layer === "all";
+  const doChunks = layer === "chunks" || layer === "all";
+  const doEmbed = layer === "embed" || layer === "all" || layer === "chunks";
+
+  if (doSync) {
+    db.resetSyncCursors(options.source);
+    const scope = options.source ? `${options.source} sync cursors` : "all sync cursors";
+    console.log(`Reset ${scope}. Run 'traul sync' to refetch.`);
+  }
+
+  if (doChunks) {
+    db.resetChunks();
+    console.log("Reset all chunks. They will be regenerated on next 'traul sync' or 'traul embed'.");
+  }
+
+  if (doEmbed) {
+    db.resetEmbeddings(EMBED_DIMS);
+    console.log("Reset all embeddings. Run 'traul embed' to regenerate.");
+  }
+}

src/commands/search.ts

@@ -1,6 +1,6 @@
 import type { TraulDB } from "../db/database";
 import { formatMessage, writeJSON } from "../lib/formatter";
-import { embed, vecToBytes } from "../lib/embeddings";
+import { embedQuery, vecToBytes } from "../lib/embeddings";
 
 export async function runSearch(
   db: TraulDB,
@@ -45,15 +45,15 @@ export async function runSearch(
     results = db.ftsSearchAll(ftsQuery, searchOpts);
   } else {
     try {
-      const vec = await embed(query);
+      const vec = await embedQuery(query);
       results = db.hybridSearchAll(vecToBytes(vec), ftsQuery, searchOpts);
       const { total_messages, embedded_messages } = db.getEmbeddingStats();
       const pct = total_messages > 0 ? Math.round((embedded_messages / total_messages) * 100) : 0;
       if (pct < 100) {
         console.warn(`search: hybrid mode — ${pct}% vector, ${100 - pct}% FTS`);
       }
     } catch {
-      console.warn("search: Ollama unavailable, falling back to FTS-only");
+      console.warn("search: embedding unavailable, falling back to FTS-only");
       results = db.ftsSearchAll(ftsQuery, searchOpts);
     }
   }

src/db/database.ts

@@ -817,6 +817,35 @@ export class TraulDB {
     this.db.run("DELETE FROM sync_cursors WHERE source = ? AND key = ?", [source, key]);
   }
 
+  resetSyncCursors(source?: string): void {
+    if (source) {
+      this.db.run("DELETE FROM sync_cursors WHERE source = ?", [source]);
+    } else {
+      this.db.run("DELETE FROM sync_cursors");
+    }
+  }
+
+  resetChunks(): void {
+    this.db.run("DELETE FROM vec_chunks");
+    this.db.run("DELETE FROM chunks");
+  }
+
+  getMeta(key: string): string | null {
+    const row = this.db
+      .query<{ value: string }, [string]>(
+        "SELECT value FROM traul_meta WHERE key = ?"
+      )
+      .get(key);
+    return row?.value ?? null;
+  }
+
+  setMeta(key: string, value: string): void {
+    this.db.run(
+      "INSERT INTO traul_meta (key, value) VALUES (?, ?) ON CONFLICT(key) DO UPDATE SET value = excluded.value",
+      [key, value]
+    );
+  }
+
   close(): void {
     this.db.close();
   }

src/db/migrations.ts

@@ -0,0 +1,63 @@
+import type { TraulDB } from "./database";
+import { CHUNKER_VERSION } from "../lib/chunker";
+import { EMBED_MODEL, EMBED_DIMS } from "../lib/embeddings";
+import * as log from "../lib/logger";
+
+export interface MigrationResult {
+  chunksReset: boolean;
+  embeddingsReset: boolean;
+  syncCursorsReset: boolean;
+}
+
+export function runMigrations(db: TraulDB): MigrationResult {
+  const result: MigrationResult = {
+    chunksReset: false,
+    embeddingsReset: false,
+    syncCursorsReset: false,
+  };
+
+  const storedChunkerVersion = db.getMeta("chunker_version");
+  const storedEmbedModel = db.getMeta("embed_model");
+  const storedEmbedDims = db.getMeta("embed_dims");
+
+  const currentDims = String(EMBED_DIMS);
+
+  // Chunker version change → reset chunks + embeddings + markdown cursors
+  if (storedChunkerVersion !== null && storedChunkerVersion !== CHUNKER_VERSION) {
+    log.info(`Chunker updated (v${storedChunkerVersion} → v${CHUNKER_VERSION}), rechunking on next sync...`);
+    db.resetChunks();
+    db.resetEmbeddings(EMBED_DIMS);
+    db.resetSyncCursors("markdown");
+    result.chunksReset = true;
+    result.embeddingsReset = true;
+    result.syncCursorsReset = true;
+  }
+
+  // Embed model or dims change → reset embeddings only
+  if (
+    !result.embeddingsReset &&
+    storedEmbedModel !== null &&
+    (storedEmbedModel !== EMBED_MODEL || storedEmbedDims !== currentDims)
+  ) {
+    const reason =
+      storedEmbedModel !== EMBED_MODEL
+        ? `model changed (${storedEmbedModel} → ${EMBED_MODEL})`
+        : `dimensions changed (${storedEmbedDims} → ${currentDims})`;
+    log.info(`Embedding ${reason}, re-embed with 'traul embed'...`);
+    db.resetEmbeddings(EMBED_DIMS);
+    result.embeddingsReset = true;
+  }
+
+  // Update stored values only if changed (avoid unnecessary writes that cause SQLITE_BUSY)
+  if (storedChunkerVersion !== CHUNKER_VERSION) {
+    db.setMeta("chunker_version", CHUNKER_VERSION);
+  }
+  if (storedEmbedModel !== EMBED_MODEL) {
+    db.setMeta("embed_model", EMBED_MODEL);
+  }
+  if (storedEmbedDims !== currentDims) {
+    db.setMeta("embed_dims", currentDims);
+  }
+
+  return result;
+}

src/db/schema.ts

@@ -117,6 +117,11 @@ const SCHEMA_SQL = `
     INSERT INTO chunks_fts(chunks_fts, rowid, content) VALUES ('delete', old.id, old.content);
     INSERT INTO chunks_fts(rowid, content) VALUES (new.id, new.content);
   END;
+
+  CREATE TABLE IF NOT EXISTS traul_meta (
+    key TEXT PRIMARY KEY,
+    value TEXT NOT NULL
+  );
 `;
 
 export function initializeDatabase(path: string): Database {

src/index.ts

@@ -13,17 +13,20 @@ import { runWhatsAppAuth } from "./commands/whatsapp-auth";
 import { runDaemonStart, runDaemonStop, runDaemonStatus } from "./commands/daemon";
 import { runSql, runSchema } from "./commands/sql";
 import { runGet } from "./commands/get";
+import { runReset } from "./commands/reset";
+import { runMigrations } from "./db/migrations";
 
 const config = loadConfig();
 ensureDbDir(config.database.path);
 const db = new TraulDB(config.database.path);
+runMigrations(db);
 
 const program = new Command();
 
 program
   .name("traul")
   .description("Traul — Personal Intelligence Engine")
-  .version("0.1.0")
+  .version("0.2.0")
   .option("-v, --verbose", "enable verbose output")
   .hook("preAction", () => {
     if (program.opts().verbose) {
@@ -137,14 +140,22 @@ program
     db.close();
   });
 
+program
+  .command("reset")
+  .description("Reset a data layer (sync, chunks, embed, all)")
+  .argument("<layer>", "layer to reset: sync, chunks, embed, all")
+  .option("-s, --source <source>", "filter by source (for sync layer)")
+  .action(async (layer: string, options) => {
+    runReset(db, layer, options);
+    db.close();
+  });
+
 program
   .command("reset-embed")
-  .description("Drop all embeddings and recreate vec tables (run 'embed' after to regenerate)")
+  .description("(deprecated: use 'traul reset embed') Drop all embeddings")
   .action(async () => {
-    const { EMBED_DIMS } = await import("./lib/embeddings");
-    console.log(`Resetting vec tables to ${EMBED_DIMS} dimensions...`);
-    db.resetEmbeddings(EMBED_DIMS);
-    console.log("Done. Run 'traul embed' to regenerate embeddings.");
+    console.log("Note: 'reset-embed' is deprecated, use 'traul reset embed' instead.");
+    runReset(db, "embed", {});
     db.close();
   });
 

src/lib/chunker.ts

@@ -13,6 +13,7 @@ export interface Chunk {
 const DEFAULT_CHUNK_SIZE = 1500;
 const DEFAULT_OVERLAP = 200;
 export const CHUNK_THRESHOLD = 2000;
+export const CHUNKER_VERSION = "2";
 
 export function shouldChunk(text: string, threshold: number = CHUNK_THRESHOLD): boolean {
   return text.length > threshold;