enhance embedding model support and update database schema for multi-model compatibility

src/db.js

@@ -112,13 +112,6 @@ db.exec(`
   CREATE INDEX IF NOT EXISTS idx_articles_normalized_title ON articles(normalized_title);
 `);
 
-db.exec(`
-  CREATE VIRTUAL TABLE IF NOT EXISTS article_embeddings USING vec0(
-    article_id INTEGER PRIMARY KEY,
-    embedding FLOAT[1024]
-  );
-`);
-
 db.exec(`
   CREATE TABLE IF NOT EXISTS article_embedding_store (
     article_id INTEGER NOT NULL,
@@ -137,6 +130,30 @@ db.exec(`
   );
 `);
 
+// vec0 table — fixed at 8192 dims to cover any model on openrouter, shorter embeddings get zero-padded
+{
+  const existing = db.prepare(`SELECT sql FROM sqlite_master WHERE type = 'table' AND name = 'article_embeddings'`).get();
+  const currentDim = existing && existing.sql && existing.sql.match(/FLOAT\[(\d+)\]/);
+
+  if (!existing) {
+    db.exec(`
+      CREATE VIRTUAL TABLE article_embeddings USING vec0(
+        article_id INTEGER PRIMARY KEY,
+        embedding FLOAT[8192]
+      );
+    `);
+  } else if (!currentDim || parseInt(currentDim[1], 10) !== 8192) {
+    db.exec(`DROP TABLE article_embeddings`);
+    db.exec(`DELETE FROM article_embedding_meta`);
+    db.exec(`
+      CREATE VIRTUAL TABLE article_embeddings USING vec0(
+        article_id INTEGER PRIMARY KEY,
+        embedding FLOAT[8192]
+      );
+    `);
+  }
+}
+
 // migrate query_embeddings to include model in primary key
 {
   const cols = db.prepare(`PRAGMA table_info(query_embeddings)`).all();

src/embeddings.js

@@ -124,7 +124,10 @@ function rebuildVec0IfModelChanged() {
     const insertMeta = db.prepare(`INSERT INTO article_embedding_meta (article_id, model) VALUES (?, ?)`);
 
     for (const row of rows) {
-      insertVec.run(BigInt(row.article_id), row.embedding);
+      const vals = new Float32Array(row.embedding.buffer, row.embedding.byteOffset, row.embedding.byteLength / 4);
+      const padded = new Float32Array(VEC0_DIM);
+      padded.set(vals);
+      insertVec.run(BigInt(row.article_id), Buffer.from(padded.buffer));
       insertMeta.run(row.article_id, EMBEDDING_MODEL);
     }
 
@@ -153,10 +156,19 @@ function buildEmbeddingInput(article) {
   return [title, description, content].join('\n\n');
 }
 
+const VEC0_DIM = 8192;
+
 function serializeEmbedding(values) {
   return Buffer.from(new Float32Array(values).buffer);
 }
 
+function padEmbeddingForVec0(values) {
+  if (values.length === VEC0_DIM) return serializeEmbedding(values);
+  const padded = new Float32Array(VEC0_DIM);
+  padded.set(values);
+  return Buffer.from(padded.buffer);
+}
+
 function normalizeQuery(input) {
   return String(input || '')
     .trim()
@@ -197,8 +209,8 @@ async function requestEmbedding(input) {
 
   const payload = await response.json();
   const embedding = payload && payload.data && payload.data[0] && payload.data[0].embedding;
-  if (!Array.isArray(embedding) || embedding.length !== 1024) {
-    throw new Error(`unexpected embedding length: ${Array.isArray(embedding) ? embedding.length : 'missing'}`);
+  if (!Array.isArray(embedding) || embedding.length === 0) {
+    throw new Error(`invalid embedding in response: ${Array.isArray(embedding) ? 'empty' : 'missing'}`);
   }
 
   return embedding;
@@ -221,8 +233,9 @@ async function generateAndStoreEmbedding(id) {
     // already in store — make sure vec0 is also up to date
     if (!selectEmbeddingBuffer.get(id)) {
       const row = selectEmbeddingFromStore.get(id, EMBEDDING_MODEL);
+      const vals = new Float32Array(row.embedding.buffer, row.embedding.byteOffset, row.embedding.byteLength / 4);
       deleteEmbedding.run(BigInt(id));
-      insertEmbedding.run(BigInt(id), row.embedding);
+      insertEmbedding.run(BigInt(id), padEmbeddingForVec0(vals));
       upsertEmbeddingMeta.run(id, EMBEDDING_MODEL);
     }
 
@@ -251,7 +264,7 @@ async function generateAndStoreEmbedding(id) {
 
     upsertEmbeddingStore.run(id, EMBEDDING_MODEL, buffer);
     deleteEmbedding.run(BigInt(id));
-    insertEmbedding.run(BigInt(id), buffer);
+    insertEmbedding.run(BigInt(id), padEmbeddingForVec0(embedding));
     upsertEmbeddingMeta.run(id, EMBEDDING_MODEL);
 
     return { stored: true, shouldPauseBatch: false };
@@ -289,7 +302,9 @@ async function backfillMissingEmbeddings(limit = 100) {
 
 function getEmbeddingBuffer(articleId) {
   const row = selectEmbeddingFromStore.get(articleId, EMBEDDING_MODEL);
-  return row ? row.embedding : null;
+  if (!row) return null;
+  const vals = new Float32Array(row.embedding.buffer, row.embedding.byteOffset, row.embedding.byteLength / 4);
+  return padEmbeddingForVec0(vals);
 }
 
 function findArticlesByEmbedding(embedding, limit) {
@@ -312,13 +327,14 @@ async function getOrCreateQueryEmbedding(query) {
 
   const cached = selectQueryEmbedding.get(normalizedQuery, EMBEDDING_MODEL);
   if (cached) {
-    return cached.embedding;
+    const vals = new Float32Array(cached.embedding.buffer, cached.embedding.byteOffset, cached.embedding.byteLength / 4);
+    return padEmbeddingForVec0(vals);
   }
 
   const embedding = await requestEmbedding(normalizedQuery);
   const buffer = serializeEmbedding(embedding);
   upsertQueryEmbedding.run(normalizedQuery, EMBEDDING_MODEL, buffer);
-  return buffer;
+  return padEmbeddingForVec0(embedding);
 }
 
 function findSimilarArticles(articleId, limit) {