@modusensus/dsh-mneme 0.6.7 → 0.6.9

package/{lib → dsh-mneme/lib}/vector-index.js

@@ -1,116 +1,116 @@
-// Thin vector-index layer over the store's embedding column. Keeps the
-// current model fingerprint + stats in a dedicated `vector_meta` table so the
-// index can detect drift (embeddings generated by a different model) and the
-// Web panel can report index health. The heavy lifting (cosine search) lives
-// in store.searchVector; this module owns the index lifecycle.
-
-const META_SCHEMA = `
-CREATE TABLE IF NOT EXISTS vector_meta (
-  key        TEXT PRIMARY KEY,
-  value      TEXT NOT NULL
-);
-`;
-
-export function createVectorIndex({ store, logger }) {
-  const db = store.db;
-  db.exec(META_SCHEMA);
-
-  function getMeta(key) {
-    const row = db.prepare("SELECT value FROM vector_meta WHERE key = ?").get(key);
-    return row?.value ?? undefined;
-  }
-
-  function setMeta(key, value) {
-    db.prepare(
-      `INSERT INTO vector_meta (key, value) VALUES (?, ?)
-       ON CONFLICT(key) DO UPDATE SET value = excluded.value`
-    ).run(key, String(value));
-  }
-
-  return {
-    /** Record the model that generated the current vectors. */
-    markModel(modelHash, dimension) {
-      setMeta("model_hash", modelHash);
-      setMeta("dimension", String(dimension ?? ""));
-      setMeta("updated_at", new Date().toISOString());
-    },
-
-    modelHash() {
-      return getMeta("model_hash");
-    },
-
-    dimension() {
-      const raw = getMeta("dimension");
-      return raw && Number.isInteger(Number(raw)) ? Number(raw) : undefined;
-    },
-
-    /** Store (or clear with null vector) one memory's embedding. */
-    saveEmbedding(id, vector) {
-      store.setEmbedding(id, vector);
-    },
-
-    /** Drop one memory's embedding. */
-    deleteEmbedding(id) {
-      store.setEmbedding(id, null);
-    },
-
-    /** Read one memory's cached vector; undefined when none/parse failure. */
-    getEmbedding(id) {
-      const row = db.prepare("SELECT embedding FROM memories WHERE id = ?").get(id);
-      if (!row?.embedding) return undefined;
-      try {
-        const v = JSON.parse(row.embedding);
-        return Array.isArray(v) && v.length ? v : undefined;
-      } catch {
-        return undefined;
-      }
-    },
-
-    /** Cosine search over embedded rows; results carry a 0..1 score. */
-    search(queryVector, { limit = 20, threshold = 0.65 } = {}) {
-      return store.searchVector(queryVector, { limit, threshold });
-    },
-
-    /** Re-embed every row missing an embedding. Returns indexed count. */
-    async rebuildIndex(embedder, { limit = 1000 } = {}) {
-      // The loop needs a single-text embedder. Native embedSingle is preferred;
-      // embed-only OpenAI-compatible clients (issue #10) are accepted too via
-      // their `embed` single-text interface, so /vector-reindex no longer
-      // silently returns 0 for them. An embedder exposing neither is ignored.
-      let embedOne = null;
-      if (embedder && typeof embedder.embedSingle === "function") {
-        embedOne = (text) => embedder.embedSingle(text);
-      } else if (embedder && typeof embedder.embed === "function") {
-        embedOne = (text) => Promise.resolve(embedder.embed(text));
-      }
-      if (!embedOne) return { indexed: 0, skipped: 0 };
-      const rows = store.needsEmbedding(limit);
-      let indexed = 0;
-      for (const row of rows) {
-        try {
-          const text = [row.title, row.content].filter(Boolean).join("\n");
-          const vector = await embedOne(text);
-          if (vector && vector.length) {
-            store.setEmbedding(row.id, vector);
-            indexed++;
-          }
-        } catch (error) {
-          logger?.warn?.(`[dsh-mneme] rebuild embed failed for ${row.id}: ${String(error)}`);
-        }
-      }
-      if (indexed > 0 && embedder.modelHash) this.markModel(embedder.modelHash, embedder.dimension);
-      return { indexed, skipped: rows.length - indexed };
-    },
-
-    /** Index stats for the Web panel / debugging. */
-    getStats() {
-      return {
-        modelHash: getMeta("model_hash"),
-        dimension: this.dimension(),
-        updatedAt: getMeta("updated_at"),
-        embeddedCount: store.embeddedCount(),
-        totalCount: store.count()
-      };
-    }
-  };
-}
+// Thin vector-index layer over the store's embedding column. Keeps the
+// current model fingerprint + stats in a dedicated `vector_meta` table so the
+// index can detect drift (embeddings generated by a different model) and the
+// Web panel can report index health. The heavy lifting (cosine search) lives
+// in store.searchVector; this module owns the index lifecycle.
+
+const META_SCHEMA = `
+CREATE TABLE IF NOT EXISTS vector_meta (
+  key        TEXT PRIMARY KEY,
+  value      TEXT NOT NULL
+);
+`;
+
+export function createVectorIndex({ store, logger }) {
+  const db = store.db;
+  db.exec(META_SCHEMA);
+
+  function getMeta(key) {
+    const row = db.prepare("SELECT value FROM vector_meta WHERE key = ?").get(key);
+    return row?.value ?? undefined;
+  }
+
+  function setMeta(key, value) {
+    db.prepare(
+      `INSERT INTO vector_meta (key, value) VALUES (?, ?)
+       ON CONFLICT(key) DO UPDATE SET value = excluded.value`
+    ).run(key, String(value));
+  }
+
+  return {
+    /** Record the model that generated the current vectors. */
+    markModel(modelHash, dimension) {
+      setMeta("model_hash", modelHash);
+      setMeta("dimension", String(dimension ?? ""));
+      setMeta("updated_at", new Date().toISOString());
+    },
+
+    modelHash() {
+      return getMeta("model_hash");
+    },
+
+    dimension() {
+      const raw = getMeta("dimension");
+      return raw && Number.isInteger(Number(raw)) ? Number(raw) : undefined;
+    },
+
+    /** Store (or clear with null vector) one memory's embedding. */
+    saveEmbedding(id, vector) {
+      store.setEmbedding(id, vector);
+    },
+
+    /** Drop one memory's embedding. */
+    deleteEmbedding(id) {
+      store.setEmbedding(id, null);
+    },
+
+    /** Read one memory's cached vector; undefined when none/parse failure. */
+    getEmbedding(id) {
+      const row = db.prepare("SELECT embedding FROM memories WHERE id = ?").get(id);
+      if (!row?.embedding) return undefined;
+      try {
+        const v = JSON.parse(row.embedding);
+        return Array.isArray(v) && v.length ? v : undefined;
+      } catch {
+        return undefined;
+      }
+    },
+
+    /** Cosine search over embedded rows; results carry a 0..1 score. */
+    search(queryVector, { limit = 20, threshold = 0.65 } = {}) {
+      return store.searchVector(queryVector, { limit, threshold });
+    },
+
+    /** Re-embed every row missing an embedding. Returns indexed count. */
+    async rebuildIndex(embedder, { limit = 1000 } = {}) {
+      // The loop needs a single-text embedder. Native embedSingle is preferred;
+      // embed-only OpenAI-compatible clients (issue #10) are accepted too via
+      // their `embed` single-text interface, so /vector-reindex no longer
+      // silently returns 0 for them. An embedder exposing neither is ignored.
+      let embedOne = null;
+      if (embedder && typeof embedder.embedSingle === "function") {
+        embedOne = (text) => embedder.embedSingle(text);
+      } else if (embedder && typeof embedder.embed === "function") {
+        embedOne = (text) => Promise.resolve(embedder.embed(text));
+      }
+      if (!embedOne) return { indexed: 0, skipped: 0 };
+      const rows = store.needsEmbedding(limit);
+      let indexed = 0;
+      for (const row of rows) {
+        try {
+          const text = [row.title, row.content].filter(Boolean).join("\n");
+          const vector = await embedOne(text);
+          if (vector && vector.length) {
+            store.setEmbedding(row.id, vector);
+            indexed++;
+          }
+        } catch (error) {
+          logger?.warn?.(`[dsh-mneme] rebuild embed failed for ${row.id}: ${String(error)}`);
+        }
+      }
+      if (indexed > 0 && embedder.modelHash) this.markModel(embedder.modelHash, embedder.dimension);
+      return { indexed, skipped: rows.length - indexed };
+    },
+
+    /** Index stats for the Web panel / debugging. */
+    getStats() {
+      return {
+        modelHash: getMeta("model_hash"),
+        dimension: this.dimension(),
+        updatedAt: getMeta("updated_at"),
+        embeddedCount: store.embeddedCount(),
+        totalCount: store.count()
+      };
+    }
+  };
+}