@nicia-ai/typegraph 0.34.0 → 0.36.0

package/dist/{chunk-GQANNXXT.cjs → chunk-5A6FNS6N.cjs}

@@ -1,7 +1,8 @@
 'use strict';
 
-var chunk2JXB3NYJ_cjs = require('./chunk-2JXB3NYJ.cjs');
-var chunkYAZE7QZR_cjs = require('./chunk-YAZE7QZR.cjs');
+var chunkCUHM5CWJ_cjs = require('./chunk-CUHM5CWJ.cjs');
+var chunkD7ZEOS56_cjs = require('./chunk-D7ZEOS56.cjs');
+var chunk7YSP4V4R_cjs = require('./chunk-7YSP4V4R.cjs');
 var drizzleOrm = require('drizzle-orm');
 
 var TABLE_PREFIX = "tg_vec";
@@ -13,7 +14,15 @@ var SQLITE_VEC_CAPABILITIES = {
   // `embedding()` default) is honored by real KNN; "none" stays brute-force.
   // Both map to the same vec0 table — only `buildSearch` differs.
   indexTypes: ["hnsw", "none"],
-  maxDimensions: SQLITE_VEC_MAX_DIMENSIONS
+  maxDimensions: SQLITE_VEC_MAX_DIMENSIONS,
+  // vec0's KNN accepts primary-key `IN (SELECT …)` pushdown (verified on
+  // sqlite-vec v0.1.9), so the candidate filter constrains the KNN itself and
+  // a filtered search returns `k` live rows — no over-fetch, no scan bound, no
+  // under-fill. The only bundled engine that can promise a full page.
+  filteredApproximateSearch: {
+    mode: "filter-pushdown",
+    guaranteesFullPage: true
+  }
 };
 function usesAnnIndex(slot) {
   return slot.indexType === "hnsw";
@@ -38,7 +47,7 @@ function sqliteVecMetricToken(metric) {
   }
 }
 function vecF32Literal(embedding, name) {
-  chunk2JXB3NYJ_cjs.assertFiniteEmbedding(embedding, name);
+  chunkD7ZEOS56_cjs.assertFiniteEmbedding(embedding, name);
   return drizzleOrm.sql`vec_f32(${JSON.stringify(embedding)})`;
 }
 function distanceExpression(embeddingColumn, queryEmbedding, metric) {
@@ -65,15 +74,15 @@ var sqliteVecStrategy = {
   name: "sqlite-vec",
   capabilities: SQLITE_VEC_CAPABILITIES,
   tableName(graphId, nodeKind, fieldPath) {
-    return chunk2JXB3NYJ_cjs.vectorPhysicalName(TABLE_PREFIX, graphId, nodeKind, fieldPath);
+    return chunkD7ZEOS56_cjs.vectorPhysicalName(TABLE_PREFIX, graphId, nodeKind, fieldPath);
   },
   ownedTables(slot) {
     const table = this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath);
-    const name = chunk2JXB3NYJ_cjs.quoteIdentifier(table);
+    const name = chunkD7ZEOS56_cjs.quoteIdentifier(table);
     const metric = sqliteVecMetricToken(slot.metric);
     return [
       {
-        logicalName: `${chunk2JXB3NYJ_cjs.VECTOR_CONTRIBUTION_PREFIX}:${slot.nodeKind}.${slot.fieldPath}`,
+        logicalName: `${chunkD7ZEOS56_cjs.VECTOR_CONTRIBUTION_PREFIX}:${slot.nodeKind}.${slot.fieldPath}`,
         owner: "sqlite-vec",
         tableName: table,
         createDdl: [
@@ -90,7 +99,7 @@ var sqliteVecStrategy = {
     ];
   },
   buildUpsert(slot, params, timestamp) {
-    const table = chunkYAZE7QZR_cjs.quotedTableName(
+    const table = chunk7YSP4V4R_cjs.quotedTableName(
       this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
     );
     const value = vecF32Literal(params.embedding, "embedding");
@@ -105,8 +114,33 @@ var sqliteVecStrategy = {
       `
     ];
   },
+  buildUpsertBatch(slot, params, timestamp) {
+    const table = chunk7YSP4V4R_cjs.quotedTableName(
+      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
+    );
+    const nodeIds = drizzleOrm.sql.join(
+      params.rows.map((row) => drizzleOrm.sql`${row.nodeId}`),
+      drizzleOrm.sql`, `
+    );
+    const valueRows = drizzleOrm.sql.join(
+      params.rows.map(
+        (row) => drizzleOrm.sql`(${row.nodeId}, ${params.graphId}, ${timestamp}, ${timestamp}, ${vecF32Literal(row.embedding, "embedding")})`
+      ),
+      drizzleOrm.sql`, `
+    );
+    return [
+      drizzleOrm.sql`
+        DELETE FROM ${table}
+        WHERE "graph_id" = ${params.graphId} AND "node_id" IN (${nodeIds})
+      `,
+      drizzleOrm.sql`
+        INSERT INTO ${table} ("node_id", "graph_id", "created_at", "updated_at", "embedding")
+        VALUES ${valueRows}
+      `
+    ];
+  },
   buildDelete(slot, params) {
-    const table = chunkYAZE7QZR_cjs.quotedTableName(
+    const table = chunk7YSP4V4R_cjs.quotedTableName(
       this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
     );
     return [
@@ -116,8 +150,13 @@ var sqliteVecStrategy = {
       `
     ];
   },
-  buildSearch(slot, params) {
-    const table = chunkYAZE7QZR_cjs.quotedTableName(
+  // vec0's KNN is brute force in C — exact by construction (the
+  // "index" is a partitioned scan, not a graph) — and the non-indexed
+  // fallback below is a plain SQL scan. The compiler may therefore
+  // route the NON-approximate `.similarTo()` branch through this form.
+  searchIsExact: true,
+  buildSearch(slot, params, candidates) {
+    const table = chunk7YSP4V4R_cjs.quotedTableName(
       this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
     );
     const embeddingColumn = drizzleOrm.sql`${table}."embedding"`;
@@ -126,38 +165,54 @@ var sqliteVecStrategy = {
       params.queryEmbedding,
       params.metric
     );
-    const score = chunk2JXB3NYJ_cjs.vectorScoreExpression(distance, params.metric);
+    const score = chunkD7ZEOS56_cjs.vectorScoreExpression(distance, params.metric);
     if (usesAnnIndex(slot)) {
       const query = vecF32Literal(params.queryEmbedding, "queryEmbedding");
+      const knnK = params.limit + (params.offset ?? 0);
       const conditions2 = [
         drizzleOrm.sql`${table}."embedding" MATCH ${query}`,
-        drizzleOrm.sql`k = ${params.limit}`,
+        drizzleOrm.sql`k = ${knnK}`,
         drizzleOrm.sql`${table}."graph_id" = ${params.graphId}`
       ];
       if (params.minScore !== void 0) {
         conditions2.push(
-          chunk2JXB3NYJ_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
+          chunkD7ZEOS56_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
         );
       }
-      return drizzleOrm.sql`
+      if (candidates !== void 0) {
+        conditions2.push(drizzleOrm.sql`${table}."node_id" IN (${candidates})`);
+      }
+      const knnBody = drizzleOrm.sql`
         SELECT ${table}."node_id" AS node_id, ${score} AS score
         FROM ${table}
         WHERE ${drizzleOrm.sql.join(conditions2, drizzleOrm.sql` AND `)}
         ORDER BY distance ASC
       `;
+      if (params.offset === void 0 || params.offset === 0) {
+        return knnBody;
+      }
+      return drizzleOrm.sql`
+        WITH knn_page AS MATERIALIZED (${knnBody})
+        SELECT node_id, score FROM knn_page
+        LIMIT ${params.limit} OFFSET ${params.offset}
+      `;
     }
     const conditions = [drizzleOrm.sql`${table}."graph_id" = ${params.graphId}`];
+    if (candidates !== void 0) {
+      conditions.push(drizzleOrm.sql`${table}."node_id" IN (${candidates})`);
+    }
     if (params.minScore !== void 0) {
       conditions.push(
-        chunk2JXB3NYJ_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
+        chunkD7ZEOS56_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
       );
     }
+    const pageClause = chunkCUHM5CWJ_cjs.vectorPageClause(params.limit, params.offset);
     return drizzleOrm.sql`
       SELECT ${table}."node_id" AS node_id, ${score} AS score
       FROM ${table}
       WHERE ${drizzleOrm.sql.join(conditions, drizzleOrm.sql` AND `)}
       ORDER BY ${distance} ASC
-      LIMIT ${params.limit}
+      ${pageClause}
     `;
   },
   distanceExpression(embeddingColumn, queryEmbedding, metric) {
@@ -172,7 +227,7 @@ var sqliteVecStrategy = {
     return void 0;
   },
   buildDropStorage(slot) {
-    const table = chunk2JXB3NYJ_cjs.quoteIdentifier(
+    const table = chunkD7ZEOS56_cjs.quoteIdentifier(
       this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
     );
     return [`DROP TABLE IF EXISTS ${table}`];
@@ -180,5 +235,5 @@ var sqliteVecStrategy = {
 };
 
 exports.sqliteVecStrategy = sqliteVecStrategy;
-//# sourceMappingURL=chunk-GQANNXXT.cjs.map
-//# sourceMappingURL=chunk-GQANNXXT.cjs.map
+//# sourceMappingURL=chunk-5A6FNS6N.cjs.map
+//# sourceMappingURL=chunk-5A6FNS6N.cjs.map

package/dist/chunk-5A6FNS6N.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/query/dialect/vector/sqlite-vec-strategy.ts"],"names":["assertFiniteEmbedding","sql","vectorPhysicalName","quoteIdentifier","VECTOR_CONTRIBUTION_PREFIX","quotedTableName","vectorScoreExpression","conditions","vectorMinScoreCondition","vectorPageClause"],"mappings":";;;;;;;AA+EA,IAAM,YAAA,GAAe,QAAA;AAOrB,IAAM,yBAAA,GAA4B,GAAA;AAElC,IAAM,uBAAA,GAA8C;AAAA,EAClD,SAAA,EAAW,IAAA;AAAA,EACX,OAAA,EAAS,CAAC,QAAA,EAAU,IAAI,CAAA;AAAA;AAAA;AAAA;AAAA,EAIxB,UAAA,EAAY,CAAC,MAAA,EAAQ,MAAM,CAAA;AAAA,EAC3B,aAAA,EAAe,yBAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAKf,yBAAA,EAA2B;AAAA,IACzB,IAAA,EAAM,iBAAA;AAAA,IACN,kBAAA,EAAoB;AAAA;AAExB,CAAA;AAGA,SAAS,aAAa,IAAA,EAA2B;AAC/C,EAAA,OAAO,KAAK,SAAA,KAAc,MAAA;AAC5B;AAGA,SAAS,qBAAqB,MAAA,EAA8B;AAC1D,EAAA,QAAQ,MAAA;AAAQ,IACd,KAAK,QAAA,EAAU;AACb,MAAA,OAAO,QAAA;AAAA,IACT;AAAA,IACA,KAAK,IAAA,EAAM;AACT,MAAA,OAAO,IAAA;AAAA,IACT;AAAA,IACA,KAAK,eAAA,EAAiB;AACpB,MAAA,MAAM,IAAI,KAAA;AAAA,QACR;AAAA,OACF;AAAA,IACF;AAAA,IACA,SAAS;AACP,MAAA,MAAM,WAAA,GAAqB,MAAA;AAC3B,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,2BAAA,EAA8B,MAAA,CAAO,WAAW,CAAC,CAAA,CAAE,CAAA;AAAA,IACrE;AAAA;AAEJ;AAQA,SAAS,aAAA,CAAc,WAA8B,IAAA,EAAmB;AACtE,EAAAA,uCAAA,CAAsB,WAAW,IAAI,CAAA;AACrC,EAAA,OAAOC,cAAA,CAAA,QAAA,EAAc,IAAA,CAAK,SAAA,CAAU,SAAS,CAAC,CAAA,CAAA,CAAA;AAChD;AAEA,SAAS,kBAAA,CACP,eAAA,EACA,cAAA,EACA,MAAA,EACK;AACL,EAAA,MAAM,KAAA,GAAQ,aAAA,CAAc,cAAA,EAAgB,gBAAgB,CAAA;AAC5D,EAAA,QAAQ,MAAA;AAAQ,IACd,KAAK,QAAA,EAAU;AACb,MAAA,OAAOA,cAAA,CAAA,oBAAA,EAA0B,eAAe,CAAA,EAAA,EAAK,KAAK,CAAA,CAAA,CAAA;AAAA,IAC5D;AAAA,IACA,KAAK,IAAA,EAAM;AACT,MAAA,OAAOA,cAAA,CAAA,gBAAA,EAAsB,eAAe,CAAA,EAAA,EAAK,KAAK,CAAA,CAAA,CAAA;AAAA,IACxD;AAAA,IACA,KAAK,eAAA,EAAiB;AACpB,MAAA,MAAM,IAAI,KAAA;AAAA,QACR;AAAA,OACF;AAAA,IACF;AAAA,IACA,SAAS;AACP,MAAA,MAAM,WAAA,GAAqB,MAAA;AAC3B,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,2BAAA,EAA8B,MAAA,CAAO,WAAW,CAAC,CAAA,CAAE,CAAA;AAAA,IACrE;AAAA;AAEJ;AAEO,IAAM,iBAAA,GAAoC;AAAA,EAC/C,IAAA,EAAM,YAAA;AAAA,EACN,YAAA,EAAc,uBAAA;AAAA,EAEd,SAAA,CAAU,OAAA,EAAS,QAAA,EAAU,SAAA,EAAW;AACtC,IAAA,OAAOC,oCAAA,CAAmB,YAAA,EAAc,OAAA,EAAS,QAAA,EAAU,SAAS,CAAA;AAAA,EACtE,CAAA;AAAA,EAEA,YAAY,IAAA,EAA4C;AACtD,IAAA,MAAM,KAAA,GAAQ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS,CAAA;AACxE,IAAA,MAAM,IAAA,GAAOC,kCAAgB,KAAK,CAAA;AAClC,IAAA,MAAM,MAAA,GAAS,oBAAA,CAAqB,IAAA,CAAK,MAAM,CAAA;AAO/C,IAAA,OAAO;AAAA,MACL;AAAA,QACE,WAAA,EAAa,GAAGC,4CAA0B,CAAA,CAAA,EAAI,KAAK,QAAQ,CAAA,CAAA,EAAI,KAAK,SAAS,CAAA,CAAA;AAAA,QAC7E,KAAA,EAAO,YAAA;AAAA,QACP,SAAA,EAAW,KAAA;AAAA,QACX,SAAA,EAAW;AAAA,UACT,sCAAsC,IAAI,CAAA;AAAA;AAAA;AAAA;AAAA;AAAA,kBAAA,EAKhC,IAAA,CAAK,UAAU,CAAA,kBAAA,EAAqB,MAAM;AAAA,EAAA;AAAA,SAEtD;AAAA,QACA,aAAA,EAAe;AAAA;AACjB,KACF;AAAA,EACF,CAAA;AAAA,EAEA,WAAA,CAAY,IAAA,EAAM,MAAA,EAA+B,SAAA,EAA2B;AAC1E,IAAA,MAAM,KAAA,GAAQC,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AACA,IAAA,MAAM,KAAA,GAAQ,aAAA,CAAc,MAAA,CAAO,SAAA,EAAW,WAAW,CAAA;AAIzD,IAAA,OAAO;AAAA,MACLJ,cAAA;AAAA,oBAAA,EACgB,KAAK;AAAA,0BAAA,EACC,MAAA,CAAO,MAAM,CAAA,kBAAA,EAAqB,MAAA,CAAO,OAAO;AAAA,MAAA,CAAA;AAAA,MAEtEA,cAAA;AAAA,oBAAA,EACgB,KAAK,CAAA;AAAA,gBAAA,EACT,MAAA,CAAO,MAAM,CAAA,EAAA,EAAK,MAAA,CAAO,OAAO,KAAK,SAAS,CAAA,EAAA,EAAK,SAAS,CAAA,EAAA,EAAK,KAAK,CAAA;AAAA,MAAA;AAAA,KAEpF;AAAA,EACF,CAAA;AAAA,EAEA,gBAAA,CACE,IAAA,EACA,MAAA,EACA,SAAA,EACgB;AAChB,IAAA,MAAM,KAAA,GAAQI,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AAGA,IAAA,MAAM,UAAUJ,cAAA,CAAI,IAAA;AAAA,MAClB,MAAA,CAAO,KAAK,GAAA,CAAI,CAAC,QAAQA,cAAA,CAAA,EAAM,GAAA,CAAI,MAAM,CAAA,CAAE,CAAA;AAAA,MAC3CA,cAAA,CAAA,EAAA;AAAA,KACF;AACA,IAAA,MAAM,YAAYA,cAAA,CAAI,IAAA;AAAA,MACpB,OAAO,IAAA,CAAK,GAAA;AAAA,QACV,CAAC,GAAA,KACCA,cAAA,CAAA,CAAA,EAAO,GAAA,CAAI,MAAM,KAAK,MAAA,CAAO,OAAO,CAAA,EAAA,EAAK,SAAS,KAAK,SAAS,CAAA,EAAA,EAAK,cAAc,GAAA,CAAI,SAAA,EAAW,WAAW,CAAC,CAAA,CAAA;AAAA,OAClH;AAAA,MACAA,cAAA,CAAA,EAAA;AAAA,KACF;AACA,IAAA,OAAO;AAAA,MACLA,cAAA;AAAA,oBAAA,EACgB,KAAK;AAAA,2BAAA,EACE,MAAA,CAAO,OAAO,CAAA,mBAAA,EAAsB,OAAO,CAAA;AAAA,MAAA,CAAA;AAAA,MAElEA,cAAA;AAAA,oBAAA,EACgB,KAAK,CAAA;AAAA,eAAA,EACV,SAAS;AAAA,MAAA;AAAA,KAEtB;AAAA,EACF,CAAA;AAAA,EAEA,WAAA,CAAY,MAAM,MAAA,EAA+C;AAC/D,IAAA,MAAM,KAAA,GAAQI,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AACA,IAAA,OAAO;AAAA,MACLJ,cAAA;AAAA,oBAAA,EACgB,KAAK;AAAA,0BAAA,EACC,MAAA,CAAO,MAAM,CAAA,kBAAA,EAAqB,MAAA,CAAO,OAAO;AAAA,MAAA;AAAA,KAExE;AAAA,EACF,CAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,aAAA,EAAe,IAAA;AAAA,EAEf,WAAA,CAAY,IAAA,EAAM,MAAA,EAA4B,UAAA,EAAuB;AACnE,IAAA,MAAM,KAAA,GAAQI,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AACA,IAAA,MAAM,eAAA,GAAkBJ,iBAAM,KAAK,CAAA,YAAA,CAAA;AACnC,IAAA,MAAM,QAAA,GAAW,kBAAA;AAAA,MACf,eAAA;AAAA,MACA,MAAA,CAAO,cAAA;AAAA,MACP,MAAA,CAAO;AAAA,KACT;AACA,IAAA,MAAM,KAAA,GAAQK,uCAAA,CAAsB,QAAA,EAAU,MAAA,CAAO,MAAM,CAAA;AAE3D,IAAA,IAAI,YAAA,CAAa,IAAI,CAAA,EAAG;AAMtB,MAAA,MAAM,KAAA,GAAQ,aAAA,CAAc,MAAA,CAAO,cAAA,EAAgB,gBAAgB,CAAA;AAGnE,MAAA,MAAM,IAAA,GAAO,MAAA,CAAO,KAAA,IAAS,MAAA,CAAO,MAAA,IAAU,CAAA,CAAA;AAC9C,MAAA,MAAMC,WAAAA,GAAoB;AAAA,QACxBN,cAAA,CAAA,EAAM,KAAK,CAAA,mBAAA,EAAsB,KAAK,CAAA,CAAA;AAAA,QACtCA,qBAAU,IAAI,CAAA,CAAA;AAAA,QACdA,cAAA,CAAA,EAAM,KAAK,CAAA,cAAA,EAAiB,MAAA,CAAO,OAAO,CAAA;AAAA,OAC5C;AACA,MAAA,IAAI,MAAA,CAAO,aAAa,MAAA,EAAW;AACjC,QAAAM,WAAAA,CAAW,IAAA;AAAA,UACTC,yCAAA,CAAwB,QAAA,EAAU,MAAA,CAAO,MAAA,EAAQ,OAAO,QAAQ;AAAA,SAClE;AAAA,MACF;AAIA,MAAA,IAAI,eAAe,MAAA,EAAW;AAC5B,QAAAD,YAAW,IAAA,CAAKN,cAAA,CAAA,EAAM,KAAK,CAAA,eAAA,EAAkB,UAAU,CAAA,CAAA,CAAG,CAAA;AAAA,MAC5D;AACA,MAAA,MAAM,OAAA,GAAUA,cAAA;AAAA,eAAA,EACL,KAAK,0BAA0B,KAAK,CAAA;AAAA,aAAA,EACtC,KAAK;AAAA,cAAA,EACJA,cAAA,CAAI,IAAA,CAAKM,WAAAA,EAAYN,cAAA,CAAA,KAAA,CAAU,CAAC;AAAA;AAAA,MAAA,CAAA;AAG1C,MAAA,IAAI,MAAA,CAAO,MAAA,KAAW,MAAA,IAAa,MAAA,CAAO,WAAW,CAAA,EAAG;AACtD,QAAA,OAAO,OAAA;AAAA,MACT;AAIA,MAAA,OAAOA,cAAA;AAAA,uCAAA,EAC4B,OAAO,CAAA;AAAA;AAAA,cAAA,EAEhC,MAAA,CAAO,KAAK,CAAA,QAAA,EAAW,MAAA,CAAO,MAAM;AAAA,MAAA,CAAA;AAAA,IAEhD;AAGA,IAAA,MAAM,aAAoB,CAACA,cAAA,CAAA,EAAM,KAAK,CAAA,cAAA,EAAiB,MAAA,CAAO,OAAO,CAAA,CAAE,CAAA;AACvE,IAAA,IAAI,eAAe,MAAA,EAAW;AAC5B,MAAA,UAAA,CAAW,IAAA,CAAKA,cAAA,CAAA,EAAM,KAAK,CAAA,eAAA,EAAkB,UAAU,CAAA,CAAA,CAAG,CAAA;AAAA,IAC5D;AACA,IAAA,IAAI,MAAA,CAAO,aAAa,MAAA,EAAW;AACjC,MAAA,UAAA,CAAW,IAAA;AAAA,QACTO,yCAAA,CAAwB,QAAA,EAAU,MAAA,CAAO,MAAA,EAAQ,OAAO,QAAQ;AAAA,OAClE;AAAA,IACF;AACA,IAAA,MAAM,UAAA,GAAaC,kCAAA,CAAiB,MAAA,CAAO,KAAA,EAAO,OAAO,MAAM,CAAA;AAC/D,IAAA,OAAOR,cAAA;AAAA,aAAA,EACI,KAAK,0BAA0B,KAAK,CAAA;AAAA,WAAA,EACtC,KAAK;AAAA,YAAA,EACJA,cAAA,CAAI,IAAA,CAAK,UAAA,EAAYA,cAAA,CAAA,KAAA,CAAU,CAAC;AAAA,eAAA,EAC7B,QAAQ,CAAA;AAAA,MAAA,EACjB,UAAU;AAAA,IAAA,CAAA;AAAA,EAEhB,CAAA;AAAA,EAEA,kBAAA,CAAmB,eAAA,EAAiB,cAAA,EAAgB,MAAA,EAAQ;AAC1D,IAAA,OAAO,kBAAA,CAAmB,eAAA,EAAiB,cAAA,EAAgB,MAAM,CAAA;AAAA,EACnE,CAAA;AAAA;AAAA;AAAA,EAIA,gBAAA,GAAoC;AAClC,IAAA,OAAO,MAAA;AAAA,EACT,CAAA;AAAA,EAEA,cAAA,GAAkC;AAChC,IAAA,OAAO,MAAA;AAAA,EACT,CAAA;AAAA,EAEA,iBAAiB,IAAA,EAAyB;AACxC,IAAA,MAAM,KAAA,GAAQE,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AAEA,IAAA,OAAO,CAAC,CAAA,qBAAA,EAAwB,KAAK,CAAA,CAAE,CAAA;AAAA,EACzC;AACF","file":"chunk-5A6FNS6N.cjs","sourcesContent":["/**\n * sqlite-vec strategy — real `vec0` KNN for local SQLite.\n *\n * This upgrades `createLocalSqliteBackend` (better-sqlite3 + the sqlite-vec\n * extension) from the legacy brute-force-on-a-shared-table model to genuine\n * approximate-nearest-neighbor search on per-`(nodeKind, fieldPath)` storage.\n *\n * ## Storage choice: a `vec0` virtual table per field\n *\n * Each slot is backed by its own `vec0` virtual table:\n *\n * ```sql\n * CREATE VIRTUAL TABLE tg_vec_<kind>_<field> USING vec0(\n *   node_id TEXT PRIMARY KEY,\n *   graph_id TEXT PARTITION KEY,\n *   +created_at TEXT,\n *   +updated_at TEXT,\n *   embedding float[<N>] distance_metric=<cosine|l2>\n * );\n * ```\n *\n * Verified against the installed `sqlite-vec` `v0.1.9`:\n *\n * - **vec0 over a plain `vec_f32` BLOB column.** The shipped option (a) — a\n *   `vec0` virtual table — is chosen over option (b) — a regular table whose\n *   `vec_f32` BLOB column is brute-forced — because vec0 is the *only* form\n *   that gives local SQLite a real ANN index (today's code only ever\n *   brute-forces). vec0 indexes inline (the virtual table *is* the index), so\n *   there is no separate `CREATE INDEX` step.\n * - **KNN with partition-correct filtering.** `WHERE embedding MATCH vec_f32(q)\n *   AND k = ? AND graph_id = ?` returns the nearest `k`, ordered by `distance`,\n *   scoped to one graph — a near vector in another `graph_id` partition does\n *   not leak. This is strictly better than libSQL's table-global `vector_top_k`\n *   (which over-fetches for multi-graph correctness): vec0's `PARTITION KEY`\n *   makes the per-graph KNN exact.\n * - **One table serves both read paths.** The same `embedding` column answers\n *   the compiler's brute-force `distanceExpression` (`vec_distance_cosine` /\n *   `vec_distance_l2` over the column, `ORDER BY … LIMIT`, no `MATCH`), so the\n *   `.where(field.similarTo(...))` CTE path needs no second structure.\n * - **`k` and `LIMIT` are mutually exclusive** in a vec0 `MATCH` query, so the\n *   ANN path uses `k = <limit>` with no `LIMIT`, and the brute-force path uses\n *   `LIMIT <limit>` with no `MATCH`.\n * - **No `ON CONFLICT` / `INSERT OR REPLACE`** on a vec0 primary key, so upsert\n *   is `DELETE` + `INSERT` (atomic under the caller's outer transaction),\n *   mirroring the FTS5 strategy.\n *\n * Metric support is `cosine` + `l2` (no `inner_product` — sqlite-vec has no\n * `vec_distance_ip`), advertised as data on `capabilities`. The table's\n * `distance_metric` is fixed at creation from the slot's metric, alongside its\n * fixed dimension `N`.\n *\n * The libSQL and pgvector siblings implement the same `VectorStrategy`\n * contract for their engines.\n */\nimport { type SQL, sql } from \"drizzle-orm\";\n\nimport { quotedTableName } from \"../../../backend/drizzle/operations/shared\";\nimport { type StrategyTableContribution } from \"../../../backend/table-contribution\";\nimport {\n  type DeleteEmbeddingParams,\n  type UpsertEmbeddingBatchParams,\n  type UpsertEmbeddingParams,\n  type VectorCapabilities,\n  type VectorMetric,\n  type VectorSearchParams,\n} from \"../../../backend/types\";\nimport {\n  assertFiniteEmbedding,\n  quoteIdentifier,\n  VECTOR_CONTRIBUTION_PREFIX,\n  vectorMinScoreCondition,\n  vectorPhysicalName,\n  vectorScoreExpression,\n  type VectorSlot,\n  type VectorStrategy,\n} from \"../vector-strategy\";\nimport { vectorPageClause } from \"./pagination\";\n\n/** Physical-name prefix for the per-field `vec0` virtual table. */\nconst TABLE_PREFIX = \"tg_vec\";\n\n/**\n * sqlite-vec stores each dimension as a 32-bit float; `float[N]` accepts\n * large `N`, but we advertise the same conservative ceiling the legacy\n * SQLite path used so capability discovery is stable across the swap.\n */\nconst SQLITE_VEC_MAX_DIMENSIONS = 8000;\n\nconst SQLITE_VEC_CAPABILITIES: VectorCapabilities = {\n  supported: true,\n  metrics: [\"cosine\", \"l2\"],\n  // vec0 always indexes inline, so the portable ANN intent (\"hnsw\", the\n  // `embedding()` default) is honored by real KNN; \"none\" stays brute-force.\n  // Both map to the same vec0 table — only `buildSearch` differs.\n  indexTypes: [\"hnsw\", \"none\"],\n  maxDimensions: SQLITE_VEC_MAX_DIMENSIONS,\n  // vec0's KNN accepts primary-key `IN (SELECT …)` pushdown (verified on\n  // sqlite-vec v0.1.9), so the candidate filter constrains the KNN itself and\n  // a filtered search returns `k` live rows — no over-fetch, no scan bound, no\n  // under-fill. The only bundled engine that can promise a full page.\n  filteredApproximateSearch: {\n    mode: \"filter-pushdown\",\n    guaranteesFullPage: true,\n  },\n};\n\n/** Whether a slot's declared index type maps to vec0's `MATCH … k =` KNN. */\nfunction usesAnnIndex(slot: VectorSlot): boolean {\n  return slot.indexType === \"hnsw\";\n}\n\n/** vec0 `distance_metric` token accepted in the virtual-table definition. */\nfunction sqliteVecMetricToken(metric: VectorMetric): string {\n  switch (metric) {\n    case \"cosine\": {\n      return \"cosine\";\n    }\n    case \"l2\": {\n      return \"l2\";\n    }\n    case \"inner_product\": {\n      throw new Error(\n        \"sqlite-vec does not support inner_product. Use 'cosine' or 'l2'.\",\n      );\n    }\n    default: {\n      const _exhaustive: never = metric;\n      throw new Error(`Unsupported vector metric: ${String(_exhaustive)}`);\n    }\n  }\n}\n\n/**\n * Formats an embedding as the bound JSON argument to `vec_f32(...)`. The\n * value rides as a parameter (not interpolated) — `vec_f32` parses the JSON\n * array into the packed float blob. Finiteness is validated first so a\n * NaN/Infinity names its index before JSON.stringify masks it as `null`.\n */\nfunction vecF32Literal(embedding: readonly number[], name: string): SQL {\n  assertFiniteEmbedding(embedding, name);\n  return sql`vec_f32(${JSON.stringify(embedding)})`;\n}\n\nfunction distanceExpression(\n  embeddingColumn: SQL,\n  queryEmbedding: readonly number[],\n  metric: VectorMetric,\n): SQL {\n  const query = vecF32Literal(queryEmbedding, \"queryEmbedding\");\n  switch (metric) {\n    case \"cosine\": {\n      return sql`vec_distance_cosine(${embeddingColumn}, ${query})`;\n    }\n    case \"l2\": {\n      return sql`vec_distance_l2(${embeddingColumn}, ${query})`;\n    }\n    case \"inner_product\": {\n      throw new Error(\n        \"sqlite-vec does not support inner_product. Use 'cosine' or 'l2'.\",\n      );\n    }\n    default: {\n      const _exhaustive: never = metric;\n      throw new Error(`Unsupported vector metric: ${String(_exhaustive)}`);\n    }\n  }\n}\n\nexport const sqliteVecStrategy: VectorStrategy = {\n  name: \"sqlite-vec\",\n  capabilities: SQLITE_VEC_CAPABILITIES,\n\n  tableName(graphId, nodeKind, fieldPath) {\n    return vectorPhysicalName(TABLE_PREFIX, graphId, nodeKind, fieldPath);\n  },\n\n  ownedTables(slot): readonly StrategyTableContribution[] {\n    const table = this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath);\n    const name = quoteIdentifier(table);\n    const metric = sqliteVecMetricToken(slot.metric);\n\n    // vec0 virtual tables cannot be modeled as a Drizzle table, so this is a\n    // raw-ddl contribution — emitted verbatim and invisible to drizzle-kit.\n    // `graph_id` is a PARTITION KEY so per-graph KNN is exact; `created_at` /\n    // `updated_at` are auxiliary (`+`) columns (stored, not filtered on). The\n    // virtual table *is* the index — no separate CREATE INDEX is emitted.\n    return [\n      {\n        logicalName: `${VECTOR_CONTRIBUTION_PREFIX}:${slot.nodeKind}.${slot.fieldPath}`,\n        owner: \"sqlite-vec\",\n        tableName: table,\n        createDdl: [\n          `CREATE VIRTUAL TABLE IF NOT EXISTS ${name} USING vec0(\n  node_id TEXT PRIMARY KEY,\n  graph_id TEXT PARTITION KEY,\n  +created_at TEXT,\n  +updated_at TEXT,\n  embedding float[${slot.dimensions}] distance_metric=${metric}\n);`,\n        ],\n        runtimeEnsure: true,\n      },\n    ];\n  },\n\n  buildUpsert(slot, params: UpsertEmbeddingParams, timestamp): readonly SQL[] {\n    const table = quotedTableName(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    const value = vecF32Literal(params.embedding, \"embedding\");\n    // vec0 rejects ON CONFLICT / INSERT OR REPLACE on its primary key —\n    // emulate upsert with DELETE + INSERT, atomic under the caller's\n    // outer transaction (mirrors the FTS5 fulltext strategy).\n    return [\n      sql`\n        DELETE FROM ${table}\n        WHERE \"node_id\" = ${params.nodeId} AND \"graph_id\" = ${params.graphId}\n      `,\n      sql`\n        INSERT INTO ${table} (\"node_id\", \"graph_id\", \"created_at\", \"updated_at\", \"embedding\")\n        VALUES (${params.nodeId}, ${params.graphId}, ${timestamp}, ${timestamp}, ${value})\n      `,\n    ];\n  },\n\n  buildUpsertBatch(\n    slot,\n    params: UpsertEmbeddingBatchParams,\n    timestamp,\n  ): readonly SQL[] {\n    const table = quotedTableName(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    // Same DELETE + INSERT upsert emulation as buildUpsert, in multi-row\n    // form: one IN-list delete, one multi-row insert.\n    const nodeIds = sql.join(\n      params.rows.map((row) => sql`${row.nodeId}`),\n      sql`, `,\n    );\n    const valueRows = sql.join(\n      params.rows.map(\n        (row) =>\n          sql`(${row.nodeId}, ${params.graphId}, ${timestamp}, ${timestamp}, ${vecF32Literal(row.embedding, \"embedding\")})`,\n      ),\n      sql`, `,\n    );\n    return [\n      sql`\n        DELETE FROM ${table}\n        WHERE \"graph_id\" = ${params.graphId} AND \"node_id\" IN (${nodeIds})\n      `,\n      sql`\n        INSERT INTO ${table} (\"node_id\", \"graph_id\", \"created_at\", \"updated_at\", \"embedding\")\n        VALUES ${valueRows}\n      `,\n    ];\n  },\n\n  buildDelete(slot, params: DeleteEmbeddingParams): readonly SQL[] {\n    const table = quotedTableName(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    return [\n      sql`\n        DELETE FROM ${table}\n        WHERE \"node_id\" = ${params.nodeId} AND \"graph_id\" = ${params.graphId}\n      `,\n    ];\n  },\n\n  // vec0's KNN is brute force in C — exact by construction (the\n  // \"index\" is a partitioned scan, not a graph) — and the non-indexed\n  // fallback below is a plain SQL scan. The compiler may therefore\n  // route the NON-approximate `.similarTo()` branch through this form.\n  searchIsExact: true,\n\n  buildSearch(slot, params: VectorSearchParams, candidates?: SQL): SQL {\n    const table = quotedTableName(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    const embeddingColumn = sql`${table}.\"embedding\"`;\n    const distance = distanceExpression(\n      embeddingColumn,\n      params.queryEmbedding,\n      params.metric,\n    );\n    const score = vectorScoreExpression(distance, params.metric);\n\n    if (usesAnnIndex(slot)) {\n      // vec0 KNN. `k` and `LIMIT` are mutually exclusive in a MATCH query, so\n      // bound `k = limit` and omit LIMIT. `graph_id` filters by the partition\n      // key (exact, no recall loss). Order by vec0's own `distance` column —\n      // the recomputed `distance` expression is reused only for score/minScore\n      // so the math matches the brute-force path and the shared helpers.\n      const query = vecF32Literal(params.queryEmbedding, \"queryEmbedding\");\n      // `k` covers the requested page: vec0 has no OFFSET inside a MATCH\n      // query, so fetch `limit + offset` neighbors and page in a wrapper.\n      const knnK = params.limit + (params.offset ?? 0);\n      const conditions: SQL[] = [\n        sql`${table}.\"embedding\" MATCH ${query}`,\n        sql`k = ${knnK}`,\n        sql`${table}.\"graph_id\" = ${params.graphId}`,\n      ];\n      if (params.minScore !== undefined) {\n        conditions.push(\n          vectorMinScoreCondition(distance, params.metric, params.minScore),\n        );\n      }\n      // vec0 KNN accepts primary-key `IN (SELECT ...)` pushdown (verified on\n      // sqlite-vec v0.1.9): the filter constrains the KNN candidate set\n      // itself, so `k` live results come back — exact, no over-fetch.\n      if (candidates !== undefined) {\n        conditions.push(sql`${table}.\"node_id\" IN (${candidates})`);\n      }\n      const knnBody = sql`\n        SELECT ${table}.\"node_id\" AS node_id, ${score} AS score\n        FROM ${table}\n        WHERE ${sql.join(conditions, sql` AND `)}\n        ORDER BY distance ASC\n      `;\n      if (params.offset === undefined || params.offset === 0) {\n        return knnBody;\n      }\n      // MATERIALIZED fences the page wrapper: SQLite would otherwise\n      // flatten the subquery and push the outer LIMIT into the vec0 MATCH\n      // query, which rejects `k = ?` and LIMIT together.\n      return sql`\n        WITH knn_page AS MATERIALIZED (${knnBody})\n        SELECT node_id, score FROM knn_page\n        LIMIT ${params.limit} OFFSET ${params.offset}\n      `;\n    }\n\n    // Brute-force scan: no MATCH, so LIMIT is allowed (and required).\n    const conditions: SQL[] = [sql`${table}.\"graph_id\" = ${params.graphId}`];\n    if (candidates !== undefined) {\n      conditions.push(sql`${table}.\"node_id\" IN (${candidates})`);\n    }\n    if (params.minScore !== undefined) {\n      conditions.push(\n        vectorMinScoreCondition(distance, params.metric, params.minScore),\n      );\n    }\n    const pageClause = vectorPageClause(params.limit, params.offset);\n    return sql`\n      SELECT ${table}.\"node_id\" AS node_id, ${score} AS score\n      FROM ${table}\n      WHERE ${sql.join(conditions, sql` AND `)}\n      ORDER BY ${distance} ASC\n      ${pageClause}\n    `;\n  },\n\n  distanceExpression(embeddingColumn, queryEmbedding, metric) {\n    return distanceExpression(embeddingColumn, queryEmbedding, metric);\n  },\n\n  // vec0 indexes inline (the virtual table is the index), so there is no\n  // standalone ANN index to create or drop — both return undefined.\n  buildCreateIndex(): SQL | undefined {\n    return undefined;\n  },\n\n  buildDropIndex(): SQL | undefined {\n    return undefined;\n  },\n\n  buildDropStorage(slot): readonly string[] {\n    const table = quoteIdentifier(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    // Dropping the vec0 virtual table removes its backing shadow tables too.\n    return [`DROP TABLE IF EXISTS ${table}`];\n  },\n};\n"]}

package/dist/{chunk-3Z6TW3CT.cjs → chunk-5AQIKK56.cjs}

@@ -1,7 +1,8 @@
 'use strict';
 
-var chunk2JXB3NYJ_cjs = require('./chunk-2JXB3NYJ.cjs');
-var chunkYAZE7QZR_cjs = require('./chunk-YAZE7QZR.cjs');
+var chunkCUHM5CWJ_cjs = require('./chunk-CUHM5CWJ.cjs');
+var chunkD7ZEOS56_cjs = require('./chunk-D7ZEOS56.cjs');
+var chunk7YSP4V4R_cjs = require('./chunk-7YSP4V4R.cjs');
 var drizzleOrm = require('drizzle-orm');
 
 var TABLE_PREFIX = "tg_vec";
@@ -12,7 +13,16 @@ var LIBSQL_CAPABILITIES = {
   metrics: ["cosine", "l2"],
   // "hnsw" = portable ANN intent, realized as DiskANN. "none" = brute-force.
   indexTypes: ["hnsw", "none"],
-  maxDimensions: LIBSQL_MAX_DIMENSIONS
+  maxDimensions: LIBSQL_MAX_DIMENSIONS,
+  // `vector_top_k` is a table function: no filter pushdown, no way to re-enter
+  // the index. A filtered DiskANN search over-fetches by
+  // CANDIDATE_FILTER_OVERFETCH and filters afterwards, so it under-fills once
+  // more than the headroom is filtered out. Unlike sqlite-vec (pushdown) and
+  // pgvector >= 0.8 (iterative scan), libSQL has no recovery at all.
+  filteredApproximateSearch: {
+    mode: "post-filter",
+    guaranteesFullPage: false
+  }
 };
 function usesAnnIndex(slot) {
   return slot.indexType === "hnsw";
@@ -37,7 +47,7 @@ function libsqlMetricToken(metric) {
   }
 }
 function vector32Literal(embedding, name) {
-  chunk2JXB3NYJ_cjs.assertFiniteEmbedding(embedding, name);
+  chunkD7ZEOS56_cjs.assertFiniteEmbedding(embedding, name);
   return drizzleOrm.sql`vector32(${`[${embedding.join(",")}]`})`;
 }
 function distanceExpression(embeddingColumn, queryEmbedding, metric) {
@@ -64,11 +74,11 @@ var libsqlVectorStrategy = {
   name: "libsql-native",
   capabilities: LIBSQL_CAPABILITIES,
   tableName(graphId, nodeKind, fieldPath) {
-    return chunk2JXB3NYJ_cjs.vectorPhysicalName(TABLE_PREFIX, graphId, nodeKind, fieldPath);
+    return chunkD7ZEOS56_cjs.vectorPhysicalName(TABLE_PREFIX, graphId, nodeKind, fieldPath);
   },
   ownedTables(slot) {
     const table = this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath);
-    const name = chunk2JXB3NYJ_cjs.quoteIdentifier(table);
+    const name = chunkD7ZEOS56_cjs.quoteIdentifier(table);
     const createDdl = [
       `CREATE TABLE IF NOT EXISTS ${name} (
   "graph_id" TEXT NOT NULL,
@@ -84,7 +94,7 @@ var libsqlVectorStrategy = {
     }
     return [
       {
-        logicalName: `${chunk2JXB3NYJ_cjs.VECTOR_CONTRIBUTION_PREFIX}:${slot.nodeKind}.${slot.fieldPath}`,
+        logicalName: `${chunkD7ZEOS56_cjs.VECTOR_CONTRIBUTION_PREFIX}:${slot.nodeKind}.${slot.fieldPath}`,
         owner: "libsql-native",
         tableName: table,
         createDdl,
@@ -93,7 +103,7 @@ var libsqlVectorStrategy = {
     ];
   },
   buildUpsert(slot, params, timestamp) {
-    const table = chunkYAZE7QZR_cjs.quotedTableName(
+    const table = chunk7YSP4V4R_cjs.quotedTableName(
       this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
     );
     const value = vector32Literal(params.embedding, "embedding");
@@ -106,8 +116,27 @@ var libsqlVectorStrategy = {
       `
     ];
   },
+  buildUpsertBatch(slot, params, timestamp) {
+    const table = chunk7YSP4V4R_cjs.quotedTableName(
+      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
+    );
+    const valueRows = drizzleOrm.sql.join(
+      params.rows.map(
+        (row) => drizzleOrm.sql`(${params.graphId}, ${row.nodeId}, ${vector32Literal(row.embedding, "embedding")}, ${timestamp}, ${timestamp})`
+      ),
+      drizzleOrm.sql`, `
+    );
+    return [
+      drizzleOrm.sql`
+        INSERT INTO ${table} ("graph_id", "node_id", "embedding", "created_at", "updated_at")
+        VALUES ${valueRows}
+        ON CONFLICT ("graph_id", "node_id")
+        DO UPDATE SET "embedding" = excluded."embedding", "updated_at" = excluded."updated_at"
+      `
+    ];
+  },
   buildDelete(slot, params) {
-    const table = chunkYAZE7QZR_cjs.quotedTableName(
+    const table = chunk7YSP4V4R_cjs.quotedTableName(
       this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
     );
     return [
@@ -117,46 +146,56 @@ var libsqlVectorStrategy = {
       `
     ];
   },
-  buildSearch(slot, params) {
+  buildSearch(slot, params, candidates) {
     const table = this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath);
-    const quoted = chunkYAZE7QZR_cjs.quotedTableName(table);
+    const quoted = chunk7YSP4V4R_cjs.quotedTableName(table);
     const embeddingColumn = drizzleOrm.sql`${quoted}."embedding"`;
     const distance = distanceExpression(
       embeddingColumn,
       params.queryEmbedding,
       params.metric
     );
-    const score = chunk2JXB3NYJ_cjs.vectorScoreExpression(distance, params.metric);
+    const score = chunkD7ZEOS56_cjs.vectorScoreExpression(distance, params.metric);
     if (usesAnnIndex(slot)) {
       const indexName = libsqlIndexName(slot);
       const query = vector32Literal(params.queryEmbedding, "queryEmbedding");
       const conditions2 = [drizzleOrm.sql`${quoted}."graph_id" = ${params.graphId}`];
       if (params.minScore !== void 0) {
         conditions2.push(
-          chunk2JXB3NYJ_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
+          chunkD7ZEOS56_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
         );
       }
+      if (candidates !== void 0) {
+        conditions2.push(drizzleOrm.sql`${quoted}."node_id" IN (${candidates})`);
+      }
+      const pageK = params.limit + (params.offset ?? 0);
+      const annK = candidates === void 0 ? pageK : pageK * CANDIDATE_FILTER_OVERFETCH;
+      const pageClause2 = chunkCUHM5CWJ_cjs.vectorPageClause(params.limit, params.offset);
       return drizzleOrm.sql`
         SELECT ${quoted}."node_id" AS node_id, ${score} AS score
-        FROM vector_top_k(${indexName}, ${query}, ${params.limit})
+        FROM vector_top_k(${indexName}, ${query}, ${annK})
         JOIN ${quoted} ON ${quoted}.rowid = id
         WHERE ${drizzleOrm.sql.join(conditions2, drizzleOrm.sql` AND `)}
         ORDER BY ${distance} ASC
-        LIMIT ${params.limit}
+        ${pageClause2}
       `;
     }
     const conditions = [drizzleOrm.sql`${quoted}."graph_id" = ${params.graphId}`];
+    if (candidates !== void 0) {
+      conditions.push(drizzleOrm.sql`${quoted}."node_id" IN (${candidates})`);
+    }
     if (params.minScore !== void 0) {
       conditions.push(
-        chunk2JXB3NYJ_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
+        chunkD7ZEOS56_cjs.vectorMinScoreCondition(distance, params.metric, params.minScore)
       );
     }
+    const pageClause = chunkCUHM5CWJ_cjs.vectorPageClause(params.limit, params.offset);
     return drizzleOrm.sql`
       SELECT ${quoted}."node_id" AS node_id, ${score} AS score
       FROM ${quoted}
       WHERE ${drizzleOrm.sql.join(conditions, drizzleOrm.sql` AND `)}
       ORDER BY ${distance} ASC
-      LIMIT ${params.limit}
+      ${pageClause}
     `;
   },
   distanceExpression(embeddingColumn, queryEmbedding, metric) {
@@ -174,24 +213,25 @@ var libsqlVectorStrategy = {
   buildDropIndex(slot) {
     if (!usesAnnIndex(slot)) return void 0;
     const indexName = libsqlIndexName(slot);
-    return drizzleOrm.sql.raw(`DROP INDEX IF EXISTS ${chunk2JXB3NYJ_cjs.quoteIdentifier(indexName)}`);
+    return drizzleOrm.sql.raw(`DROP INDEX IF EXISTS ${chunkD7ZEOS56_cjs.quoteIdentifier(indexName)}`);
   },
   buildDropStorage(slot) {
-    const table = chunk2JXB3NYJ_cjs.quoteIdentifier(
+    const table = chunkD7ZEOS56_cjs.quoteIdentifier(
       this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath)
     );
     const statements = [];
     if (usesAnnIndex(slot)) {
       statements.push(
-        `DROP INDEX IF EXISTS ${chunk2JXB3NYJ_cjs.quoteIdentifier(libsqlIndexName(slot))}`
+        `DROP INDEX IF EXISTS ${chunkD7ZEOS56_cjs.quoteIdentifier(libsqlIndexName(slot))}`
       );
     }
     statements.push(`DROP TABLE IF EXISTS ${table}`);
     return statements;
   }
 };
+var CANDIDATE_FILTER_OVERFETCH = 4;
 function libsqlIndexName(slot) {
-  return chunk2JXB3NYJ_cjs.vectorPhysicalName(
+  return chunkD7ZEOS56_cjs.vectorPhysicalName(
     INDEX_PREFIX,
     slot.graphId,
     slot.nodeKind,
@@ -199,11 +239,11 @@ function libsqlIndexName(slot) {
   );
 }
 function libsqlVectorIndexDdl(table, slot) {
-  const indexName = chunk2JXB3NYJ_cjs.quoteIdentifier(libsqlIndexName(slot));
+  const indexName = chunkD7ZEOS56_cjs.quoteIdentifier(libsqlIndexName(slot));
   const metric = libsqlMetricToken(slot.metric);
-  return `CREATE INDEX IF NOT EXISTS ${indexName} ON ${chunk2JXB3NYJ_cjs.quoteIdentifier(table)}(libsql_vector_idx("embedding", 'metric=${metric}'));`;
+  return `CREATE INDEX IF NOT EXISTS ${indexName} ON ${chunkD7ZEOS56_cjs.quoteIdentifier(table)}(libsql_vector_idx("embedding", 'metric=${metric}'));`;
 }
 
 exports.libsqlVectorStrategy = libsqlVectorStrategy;
-//# sourceMappingURL=chunk-3Z6TW3CT.cjs.map
-//# sourceMappingURL=chunk-3Z6TW3CT.cjs.map
+//# sourceMappingURL=chunk-5AQIKK56.cjs.map
+//# sourceMappingURL=chunk-5AQIKK56.cjs.map

package/dist/chunk-5AQIKK56.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/query/dialect/vector/libsql-strategy.ts"],"names":["assertFiniteEmbedding","sql","vectorPhysicalName","quoteIdentifier","VECTOR_CONTRIBUTION_PREFIX","quotedTableName","vectorScoreExpression","conditions","vectorMinScoreCondition","pageClause","vectorPageClause"],"mappings":";;;;;;;AA6CA,IAAM,YAAA,GAAe,QAAA;AACrB,IAAM,YAAA,GAAe,WAAA;AAKrB,IAAM,qBAAA,GAAwB,KAAA;AAE9B,IAAM,mBAAA,GAA0C;AAAA,EAC9C,SAAA,EAAW,IAAA;AAAA,EACX,OAAA,EAAS,CAAC,QAAA,EAAU,IAAI,CAAA;AAAA;AAAA,EAExB,UAAA,EAAY,CAAC,MAAA,EAAQ,MAAM,CAAA;AAAA,EAC3B,aAAA,EAAe,qBAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMf,yBAAA,EAA2B;AAAA,IACzB,IAAA,EAAM,aAAA;AAAA,IACN,kBAAA,EAAoB;AAAA;AAExB,CAAA;AAGA,SAAS,aAAa,IAAA,EAA2B;AAC/C,EAAA,OAAO,KAAK,SAAA,KAAc,MAAA;AAC5B;AAGA,SAAS,kBAAkB,MAAA,EAA8B;AACvD,EAAA,QAAQ,MAAA;AAAQ,IACd,KAAK,QAAA,EAAU;AACb,MAAA,OAAO,QAAA;AAAA,IACT;AAAA,IACA,KAAK,IAAA,EAAM;AACT,MAAA,OAAO,IAAA;AAAA,IACT;AAAA,IACA,KAAK,eAAA,EAAiB;AACpB,MAAA,MAAM,IAAI,KAAA;AAAA,QACR;AAAA,OACF;AAAA,IACF;AAAA,IACA,SAAS;AACP,MAAA,MAAM,WAAA,GAAqB,MAAA;AAC3B,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,2BAAA,EAA8B,MAAA,CAAO,WAAW,CAAC,CAAA,CAAE,CAAA;AAAA,IACrE;AAAA;AAEJ;AAOA,SAAS,eAAA,CAAgB,WAA8B,IAAA,EAAmB;AACxE,EAAAA,uCAAA,CAAsB,WAAW,IAAI,CAAA;AACrC,EAAA,OAAOC,0BAAe,CAAA,CAAA,EAAI,SAAA,CAAU,IAAA,CAAK,GAAG,CAAC,CAAA,CAAA,CAAG,CAAA,CAAA,CAAA;AAClD;AAEA,SAAS,kBAAA,CACP,eAAA,EACA,cAAA,EACA,MAAA,EACK;AACL,EAAA,MAAM,KAAA,GAAQ,eAAA,CAAgB,cAAA,EAAgB,gBAAgB,CAAA;AAC9D,EAAA,QAAQ,MAAA;AAAQ,IACd,KAAK,QAAA,EAAU;AACb,MAAA,OAAOA,cAAA,CAAA,oBAAA,EAA0B,eAAe,CAAA,EAAA,EAAK,KAAK,CAAA,CAAA,CAAA;AAAA,IAC5D;AAAA,IACA,KAAK,IAAA,EAAM;AACT,MAAA,OAAOA,cAAA,CAAA,mBAAA,EAAyB,eAAe,CAAA,EAAA,EAAK,KAAK,CAAA,CAAA,CAAA;AAAA,IAC3D;AAAA,IACA,KAAK,eAAA,EAAiB;AACpB,MAAA,MAAM,IAAI,KAAA;AAAA,QACR;AAAA,OACF;AAAA,IACF;AAAA,IACA,SAAS;AACP,MAAA,MAAM,WAAA,GAAqB,MAAA;AAC3B,MAAA,MAAM,IAAI,KAAA,CAAM,CAAA,2BAAA,EAA8B,MAAA,CAAO,WAAW,CAAC,CAAA,CAAE,CAAA;AAAA,IACrE;AAAA;AAEJ;AAEO,IAAM,oBAAA,GAAuC;AAAA,EAClD,IAAA,EAAM,eAAA;AAAA,EACN,YAAA,EAAc,mBAAA;AAAA,EAEd,SAAA,CAAU,OAAA,EAAS,QAAA,EAAU,SAAA,EAAW;AACtC,IAAA,OAAOC,oCAAA,CAAmB,YAAA,EAAc,OAAA,EAAS,QAAA,EAAU,SAAS,CAAA;AAAA,EACtE,CAAA;AAAA,EAEA,YAAY,IAAA,EAA4C;AACtD,IAAA,MAAM,KAAA,GAAQ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS,CAAA;AACxE,IAAA,MAAM,IAAA,GAAOC,kCAAgB,KAAK,CAAA;AAKlC,IAAA,MAAM,SAAA,GAAY;AAAA,MAChB,8BAA8B,IAAI,CAAA;AAAA;AAAA;AAAA,uBAAA,EAGf,KAAK,UAAU,CAAA;AAAA;AAAA;AAAA;AAAA,EAAA;AAAA,KAKpC;AAEA,IAAA,IAAI,YAAA,CAAa,IAAI,CAAA,EAAG;AAGtB,MAAA,SAAA,CAAU,IAAA,CAAK,oBAAA,CAAqB,KAAA,EAAO,IAAI,CAAC,CAAA;AAAA,IAClD;AAEA,IAAA,OAAO;AAAA,MACL;AAAA,QACE,WAAA,EAAa,GAAGC,4CAA0B,CAAA,CAAA,EAAI,KAAK,QAAQ,CAAA,CAAA,EAAI,KAAK,SAAS,CAAA,CAAA;AAAA,QAC7E,KAAA,EAAO,eAAA;AAAA,QACP,SAAA,EAAW,KAAA;AAAA,QACX,SAAA;AAAA,QACA,aAAA,EAAe;AAAA;AACjB,KACF;AAAA,EACF,CAAA;AAAA,EAEA,WAAA,CAAY,IAAA,EAAM,MAAA,EAA+B,SAAA,EAA2B;AAC1E,IAAA,MAAM,KAAA,GAAQC,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AACA,IAAA,MAAM,KAAA,GAAQ,eAAA,CAAgB,MAAA,CAAO,SAAA,EAAW,WAAW,CAAA;AAC3D,IAAA,OAAO;AAAA,MACLJ,cAAA;AAAA,oBAAA,EACgB,KAAK,CAAA;AAAA,gBAAA,EACT,MAAA,CAAO,OAAO,CAAA,EAAA,EAAK,MAAA,CAAO,MAAM,KAAK,KAAK,CAAA,EAAA,EAAK,SAAS,CAAA,EAAA,EAAK,SAAS,CAAA;AAAA;AAAA,oCAAA,EAElD,KAAK,oBAAoB,SAAS;AAAA,MAAA;AAAA,KAEpE;AAAA,EACF,CAAA;AAAA,EAEA,gBAAA,CACE,IAAA,EACA,MAAA,EACA,SAAA,EACgB;AAChB,IAAA,MAAM,KAAA,GAAQI,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AACA,IAAA,MAAM,YAAYJ,cAAA,CAAI,IAAA;AAAA,MACpB,OAAO,IAAA,CAAK,GAAA;AAAA,QACV,CAAC,GAAA,KACCA,cAAA,CAAA,CAAA,EAAO,MAAA,CAAO,OAAO,KAAK,GAAA,CAAI,MAAM,CAAA,EAAA,EAAK,eAAA,CAAgB,IAAI,SAAA,EAAW,WAAW,CAAC,CAAA,EAAA,EAAK,SAAS,KAAK,SAAS,CAAA,CAAA;AAAA,OACpH;AAAA,MACAA,cAAA,CAAA,EAAA;AAAA,KACF;AAIA,IAAA,OAAO;AAAA,MACLA,cAAA;AAAA,oBAAA,EACgB,KAAK,CAAA;AAAA,eAAA,EACV,SAAS;AAAA;AAAA;AAAA,MAAA;AAAA,KAItB;AAAA,EACF,CAAA;AAAA,EAEA,WAAA,CAAY,MAAM,MAAA,EAA+C;AAC/D,IAAA,MAAM,KAAA,GAAQI,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AACA,IAAA,OAAO;AAAA,MACLJ,cAAA;AAAA,oBAAA,EACgB,KAAK;AAAA,2BAAA,EACE,MAAA,CAAO,OAAO,CAAA,iBAAA,EAAoB,MAAA,CAAO,MAAM;AAAA,MAAA;AAAA,KAExE;AAAA,EACF,CAAA;AAAA,EAEA,WAAA,CAAY,IAAA,EAAM,MAAA,EAA4B,UAAA,EAAuB;AACnE,IAAA,MAAM,KAAA,GAAQ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS,CAAA;AACxE,IAAA,MAAM,MAAA,GAASI,kCAAgB,KAAK,CAAA;AACpC,IAAA,MAAM,eAAA,GAAkBJ,iBAAM,MAAM,CAAA,YAAA,CAAA;AACpC,IAAA,MAAM,QAAA,GAAW,kBAAA;AAAA,MACf,eAAA;AAAA,MACA,MAAA,CAAO,cAAA;AAAA,MACP,MAAA,CAAO;AAAA,KACT;AACA,IAAA,MAAM,KAAA,GAAQK,uCAAA,CAAsB,QAAA,EAAU,MAAA,CAAO,MAAM,CAAA;AAE3D,IAAA,IAAI,YAAA,CAAa,IAAI,CAAA,EAAG;AAKtB,MAAA,MAAM,SAAA,GAAY,gBAAgB,IAAI,CAAA;AACtC,MAAA,MAAM,KAAA,GAAQ,eAAA,CAAgB,MAAA,CAAO,cAAA,EAAgB,gBAAgB,CAAA;AACrE,MAAA,MAAMC,cAAoB,CAACN,cAAA,CAAA,EAAM,MAAM,CAAA,cAAA,EAAiB,MAAA,CAAO,OAAO,CAAA,CAAE,CAAA;AACxE,MAAA,IAAI,MAAA,CAAO,aAAa,MAAA,EAAW;AACjC,QAAAM,WAAAA,CAAW,IAAA;AAAA,UACTC,yCAAA,CAAwB,QAAA,EAAU,MAAA,CAAO,MAAA,EAAQ,OAAO,QAAQ;AAAA,SAClE;AAAA,MACF;AAOA,MAAA,IAAI,eAAe,MAAA,EAAW;AAC5B,QAAAD,YAAW,IAAA,CAAKN,cAAA,CAAA,EAAM,MAAM,CAAA,eAAA,EAAkB,UAAU,CAAA,CAAA,CAAG,CAAA;AAAA,MAC7D;AAGA,MAAA,MAAM,KAAA,GAAQ,MAAA,CAAO,KAAA,IAAS,MAAA,CAAO,MAAA,IAAU,CAAA,CAAA;AAC/C,MAAA,MAAM,IAAA,GACJ,UAAA,KAAe,MAAA,GAAY,KAAA,GAAQ,KAAA,GAAQ,0BAAA;AAC7C,MAAA,MAAMQ,WAAAA,GAAaC,kCAAA,CAAiB,MAAA,CAAO,KAAA,EAAO,OAAO,MAAM,CAAA;AAC/D,MAAA,OAAOT,cAAA;AAAA,eAAA,EACI,MAAM,0BAA0B,KAAK,CAAA;AAAA,0BAAA,EAC1B,SAAS,CAAA,EAAA,EAAK,KAAK,CAAA,EAAA,EAAK,IAAI,CAAA;AAAA,aAAA,EACzC,MAAM,OAAO,MAAM,CAAA;AAAA,cAAA,EAClBA,cAAA,CAAI,IAAA,CAAKM,WAAAA,EAAYN,cAAA,CAAA,KAAA,CAAU,CAAC;AAAA,iBAAA,EAC7B,QAAQ,CAAA;AAAA,QAAA,EACjBQ,WAAU;AAAA,MAAA,CAAA;AAAA,IAEhB;AAGA,IAAA,MAAM,aAAoB,CAACR,cAAA,CAAA,EAAM,MAAM,CAAA,cAAA,EAAiB,MAAA,CAAO,OAAO,CAAA,CAAE,CAAA;AACxE,IAAA,IAAI,eAAe,MAAA,EAAW;AAC5B,MAAA,UAAA,CAAW,IAAA,CAAKA,cAAA,CAAA,EAAM,MAAM,CAAA,eAAA,EAAkB,UAAU,CAAA,CAAA,CAAG,CAAA;AAAA,IAC7D;AACA,IAAA,IAAI,MAAA,CAAO,aAAa,MAAA,EAAW;AACjC,MAAA,UAAA,CAAW,IAAA;AAAA,QACTO,yCAAA,CAAwB,QAAA,EAAU,MAAA,CAAO,MAAA,EAAQ,OAAO,QAAQ;AAAA,OAClE;AAAA,IACF;AACA,IAAA,MAAM,UAAA,GAAaE,kCAAA,CAAiB,MAAA,CAAO,KAAA,EAAO,OAAO,MAAM,CAAA;AAC/D,IAAA,OAAOT,cAAA;AAAA,aAAA,EACI,MAAM,0BAA0B,KAAK,CAAA;AAAA,WAAA,EACvC,MAAM;AAAA,YAAA,EACLA,cAAA,CAAI,IAAA,CAAK,UAAA,EAAYA,cAAA,CAAA,KAAA,CAAU,CAAC;AAAA,eAAA,EAC7B,QAAQ,CAAA;AAAA,MAAA,EACjB,UAAU;AAAA,IAAA,CAAA;AAAA,EAEhB,CAAA;AAAA,EAEA,kBAAA,CAAmB,eAAA,EAAiB,cAAA,EAAgB,MAAA,EAAQ;AAC1D,IAAA,OAAO,kBAAA,CAAmB,eAAA,EAAiB,cAAA,EAAgB,MAAM,CAAA;AAAA,EACnE,CAAA;AAAA,EAEA,iBAAiB,IAAA,EAAuB;AACtC,IAAA,IAAI,CAAC,YAAA,CAAa,IAAI,CAAA,EAAG,OAAO,MAAA;AAChC,IAAA,OAAOA,cAAA,CAAI,GAAA;AAAA,MACT,oBAAA;AAAA,QACE,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS,CAAA;AAAA,QAC1D;AAAA;AACF,KACF;AAAA,EACF,CAAA;AAAA,EAEA,eAAe,IAAA,EAAuB;AACpC,IAAA,IAAI,CAAC,YAAA,CAAa,IAAI,CAAA,EAAG,OAAO,MAAA;AAChC,IAAA,MAAM,SAAA,GAAY,gBAAgB,IAAI,CAAA;AACtC,IAAA,OAAOA,eAAI,GAAA,CAAI,CAAA,qBAAA,EAAwBE,iCAAA,CAAgB,SAAS,CAAC,CAAA,CAAE,CAAA;AAAA,EACrE,CAAA;AAAA,EAEA,iBAAiB,IAAA,EAAyB;AACxC,IAAA,MAAM,KAAA,GAAQA,iCAAA;AAAA,MACZ,KAAK,SAAA,CAAU,IAAA,CAAK,SAAS,IAAA,CAAK,QAAA,EAAU,KAAK,SAAS;AAAA,KAC5D;AACA,IAAA,MAAM,aAAuB,EAAC;AAG9B,IAAA,IAAI,YAAA,CAAa,IAAI,CAAA,EAAG;AACtB,MAAA,UAAA,CAAW,IAAA;AAAA,QACT,CAAA,qBAAA,EAAwBA,iCAAA,CAAgB,eAAA,CAAgB,IAAI,CAAC,CAAC,CAAA;AAAA,OAChE;AAAA,IACF;AACA,IAAA,UAAA,CAAW,IAAA,CAAK,CAAA,qBAAA,EAAwB,KAAK,CAAA,CAAE,CAAA;AAC/C,IAAA,OAAO,UAAA;AAAA,EACT;AACF;AAUA,IAAM,0BAAA,GAA6B,CAAA;AAEnC,SAAS,gBAAgB,IAAA,EAA0B;AACjD,EAAA,OAAOD,oCAAA;AAAA,IACL,YAAA;AAAA,IACA,IAAA,CAAK,OAAA;AAAA,IACL,IAAA,CAAK,QAAA;AAAA,IACL,IAAA,CAAK;AAAA,GACP;AACF;AAEA,SAAS,oBAAA,CAAqB,OAAe,IAAA,EAA0B;AACrE,EAAA,MAAM,SAAA,GAAYC,iCAAA,CAAgB,eAAA,CAAgB,IAAI,CAAC,CAAA;AACvD,EAAA,MAAM,MAAA,GAAS,iBAAA,CAAkB,IAAA,CAAK,MAAM,CAAA;AAC5C,EAAA,OAAO,8BAA8B,SAAS,CAAA,IAAA,EAAOA,kCAAgB,KAAK,CAAC,2CAA2C,MAAM,CAAA,IAAA,CAAA;AAC9H","file":"chunk-5AQIKK56.cjs","sourcesContent":["/**\n * libSQL native vector strategy.\n *\n * libSQL ships vector search in core — no extension to load — so this\n * strategy is wired unconditionally by `createLibsqlBackend`. It backs each\n * `(nodeKind, fieldPath)` with its own `F32_BLOB(N)` table (the spike proved\n * the dimension MUST live in the column type for `libsql_vector_idx` to\n * build), brute-forces with `vector_distance_cos` / `vector_distance_l2`, and\n * accelerates with DiskANN (`libsql_vector_idx` + `vector_top_k`) when the\n * slot declares an approximate index.\n *\n * Metric support is `cosine` + `l2` (no `inner_product`), matching sqlite-vec\n * and advertised as data on `capabilities`. The generic `\"hnsw\"` index intent\n * (the portable \"give me an ANN index\" signal that `embedding()` defaults to)\n * is realized here as libSQL's DiskANN index; `\"ivfflat\"` is not supported and\n * a slot declaring it is materialized brute-force-only.\n *\n * The `pgvectorStrategy` and `sqliteVecStrategy` siblings follow the same\n * `VectorStrategy` contract for their engines.\n */\nimport { type SQL, sql } from \"drizzle-orm\";\n\nimport { quotedTableName } from \"../../../backend/drizzle/operations/shared\";\nimport { type StrategyTableContribution } from \"../../../backend/table-contribution\";\nimport {\n  type DeleteEmbeddingParams,\n  type UpsertEmbeddingBatchParams,\n  type UpsertEmbeddingParams,\n  type VectorCapabilities,\n  type VectorMetric,\n  type VectorSearchParams,\n} from \"../../../backend/types\";\nimport {\n  assertFiniteEmbedding,\n  quoteIdentifier,\n  VECTOR_CONTRIBUTION_PREFIX,\n  vectorMinScoreCondition,\n  vectorPhysicalName,\n  vectorScoreExpression,\n  type VectorSlot,\n  type VectorStrategy,\n} from \"../vector-strategy\";\nimport { vectorPageClause } from \"./pagination\";\n\n/** Physical-name prefixes for the per-field table and its DiskANN index. */\nconst TABLE_PREFIX = \"tg_vec\";\nconst INDEX_PREFIX = \"tg_vecidx\";\n\n/**\n * libSQL `F32_BLOB` caps at 65,536 dimensions (per Turso docs / spike §2).\n */\nconst LIBSQL_MAX_DIMENSIONS = 65_536;\n\nconst LIBSQL_CAPABILITIES: VectorCapabilities = {\n  supported: true,\n  metrics: [\"cosine\", \"l2\"],\n  // \"hnsw\" = portable ANN intent, realized as DiskANN. \"none\" = brute-force.\n  indexTypes: [\"hnsw\", \"none\"],\n  maxDimensions: LIBSQL_MAX_DIMENSIONS,\n  // `vector_top_k` is a table function: no filter pushdown, no way to re-enter\n  // the index. A filtered DiskANN search over-fetches by\n  // CANDIDATE_FILTER_OVERFETCH and filters afterwards, so it under-fills once\n  // more than the headroom is filtered out. Unlike sqlite-vec (pushdown) and\n  // pgvector >= 0.8 (iterative scan), libSQL has no recovery at all.\n  filteredApproximateSearch: {\n    mode: \"post-filter\",\n    guaranteesFullPage: false,\n  },\n};\n\n/** Whether a slot's declared index type maps to a real libSQL ANN index. */\nfunction usesAnnIndex(slot: VectorSlot): boolean {\n  return slot.indexType === \"hnsw\";\n}\n\n/** libSQL metric token accepted by `libsql_vector_idx(col, 'metric=…')`. */\nfunction libsqlMetricToken(metric: VectorMetric): string {\n  switch (metric) {\n    case \"cosine\": {\n      return \"cosine\";\n    }\n    case \"l2\": {\n      return \"l2\";\n    }\n    case \"inner_product\": {\n      throw new Error(\n        \"libSQL vector search does not support inner_product. Use 'cosine' or 'l2'.\",\n      );\n    }\n    default: {\n      const _exhaustive: never = metric;\n      throw new Error(`Unsupported vector metric: ${String(_exhaustive)}`);\n    }\n  }\n}\n\n/**\n * Formats an embedding as the text argument to `vector32(...)`. The value is\n * bound as a parameter (not interpolated) — `vector32` parses it into the\n * F32 blob. Finiteness is validated first so a NaN names its index.\n */\nfunction vector32Literal(embedding: readonly number[], name: string): SQL {\n  assertFiniteEmbedding(embedding, name);\n  return sql`vector32(${`[${embedding.join(\",\")}]`})`;\n}\n\nfunction distanceExpression(\n  embeddingColumn: SQL,\n  queryEmbedding: readonly number[],\n  metric: VectorMetric,\n): SQL {\n  const query = vector32Literal(queryEmbedding, \"queryEmbedding\");\n  switch (metric) {\n    case \"cosine\": {\n      return sql`vector_distance_cos(${embeddingColumn}, ${query})`;\n    }\n    case \"l2\": {\n      return sql`vector_distance_l2(${embeddingColumn}, ${query})`;\n    }\n    case \"inner_product\": {\n      throw new Error(\n        \"libSQL vector search does not support inner_product. Use 'cosine' or 'l2'.\",\n      );\n    }\n    default: {\n      const _exhaustive: never = metric;\n      throw new Error(`Unsupported vector metric: ${String(_exhaustive)}`);\n    }\n  }\n}\n\nexport const libsqlVectorStrategy: VectorStrategy = {\n  name: \"libsql-native\",\n  capabilities: LIBSQL_CAPABILITIES,\n\n  tableName(graphId, nodeKind, fieldPath) {\n    return vectorPhysicalName(TABLE_PREFIX, graphId, nodeKind, fieldPath);\n  },\n\n  ownedTables(slot): readonly StrategyTableContribution[] {\n    const table = this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath);\n    const name = quoteIdentifier(table);\n\n    // No standalone graph_id index: the PRIMARY KEY (graph_id, node_id) already\n    // covers `WHERE graph_id = ?` via its leading column, so a separate index\n    // would be pure write amplification.\n    const createDdl = [\n      `CREATE TABLE IF NOT EXISTS ${name} (\n  \"graph_id\" TEXT NOT NULL,\n  \"node_id\" TEXT NOT NULL,\n  \"embedding\" F32_BLOB(${slot.dimensions}) NOT NULL,\n  \"created_at\" TEXT NOT NULL,\n  \"updated_at\" TEXT NOT NULL,\n  PRIMARY KEY (\"graph_id\", \"node_id\")\n);`,\n    ];\n\n    if (usesAnnIndex(slot)) {\n      // The DiskANN index DDL is deterministic raw SQL; fold it into the\n      // contribution so materialization creates table + index together.\n      createDdl.push(libsqlVectorIndexDdl(table, slot));\n    }\n\n    return [\n      {\n        logicalName: `${VECTOR_CONTRIBUTION_PREFIX}:${slot.nodeKind}.${slot.fieldPath}`,\n        owner: \"libsql-native\",\n        tableName: table,\n        createDdl,\n        runtimeEnsure: true,\n      },\n    ];\n  },\n\n  buildUpsert(slot, params: UpsertEmbeddingParams, timestamp): readonly SQL[] {\n    const table = quotedTableName(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    const value = vector32Literal(params.embedding, \"embedding\");\n    return [\n      sql`\n        INSERT INTO ${table} (\"graph_id\", \"node_id\", \"embedding\", \"created_at\", \"updated_at\")\n        VALUES (${params.graphId}, ${params.nodeId}, ${value}, ${timestamp}, ${timestamp})\n        ON CONFLICT (\"graph_id\", \"node_id\")\n        DO UPDATE SET \"embedding\" = ${value}, \"updated_at\" = ${timestamp}\n      `,\n    ];\n  },\n\n  buildUpsertBatch(\n    slot,\n    params: UpsertEmbeddingBatchParams,\n    timestamp,\n  ): readonly SQL[] {\n    const table = quotedTableName(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    const valueRows = sql.join(\n      params.rows.map(\n        (row) =>\n          sql`(${params.graphId}, ${row.nodeId}, ${vector32Literal(row.embedding, \"embedding\")}, ${timestamp}, ${timestamp})`,\n      ),\n      sql`, `,\n    );\n    // `excluded.\"embedding\"` reuses the row's already-converted F32_BLOB\n    // value, so the multi-row form needs no per-row conversion in the\n    // update arm (unlike buildUpsert's single bound value).\n    return [\n      sql`\n        INSERT INTO ${table} (\"graph_id\", \"node_id\", \"embedding\", \"created_at\", \"updated_at\")\n        VALUES ${valueRows}\n        ON CONFLICT (\"graph_id\", \"node_id\")\n        DO UPDATE SET \"embedding\" = excluded.\"embedding\", \"updated_at\" = excluded.\"updated_at\"\n      `,\n    ];\n  },\n\n  buildDelete(slot, params: DeleteEmbeddingParams): readonly SQL[] {\n    const table = quotedTableName(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    return [\n      sql`\n        DELETE FROM ${table}\n        WHERE \"graph_id\" = ${params.graphId} AND \"node_id\" = ${params.nodeId}\n      `,\n    ];\n  },\n\n  buildSearch(slot, params: VectorSearchParams, candidates?: SQL): SQL {\n    const table = this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath);\n    const quoted = quotedTableName(table);\n    const embeddingColumn = sql`${quoted}.\"embedding\"`;\n    const distance = distanceExpression(\n      embeddingColumn,\n      params.queryEmbedding,\n      params.metric,\n    );\n    const score = vectorScoreExpression(distance, params.metric);\n\n    if (usesAnnIndex(slot)) {\n      // DiskANN: vector_top_k returns rowids from the per-field table; join\n      // back, then scope by graph and compute the real distance/score.\n      // `vector_top_k` is table-global, so single-graph deployments are exact;\n      // multi-graph recall is bounded by the over-fetched k (documented).\n      const indexName = libsqlIndexName(slot);\n      const query = vector32Literal(params.queryEmbedding, \"queryEmbedding\");\n      const conditions: SQL[] = [sql`${quoted}.\"graph_id\" = ${params.graphId}`];\n      if (params.minScore !== undefined) {\n        conditions.push(\n          vectorMinScoreCondition(distance, params.metric, params.minScore),\n        );\n      }\n      // `vector_top_k` is a table function with no filter pushdown, so a\n      // candidate filter can only be applied AFTER the ANN retrieval.\n      // Over-fetch the DiskANN k to leave headroom for filtered-out rows;\n      // recall is bounded by the over-fetch, and the page under-fills once\n      // more than the headroom is filtered out. Declared, not merely noted:\n      // `capabilities.filteredApproximateSearch.mode === \"post-filter\"`.\n      if (candidates !== undefined) {\n        conditions.push(sql`${quoted}.\"node_id\" IN (${candidates})`);\n      }\n      // `pageK` covers the requested page (`limit + offset`); the\n      // candidate-filter over-fetch multiplies on top of it.\n      const pageK = params.limit + (params.offset ?? 0);\n      const annK =\n        candidates === undefined ? pageK : pageK * CANDIDATE_FILTER_OVERFETCH;\n      const pageClause = vectorPageClause(params.limit, params.offset);\n      return sql`\n        SELECT ${quoted}.\"node_id\" AS node_id, ${score} AS score\n        FROM vector_top_k(${indexName}, ${query}, ${annK})\n        JOIN ${quoted} ON ${quoted}.rowid = id\n        WHERE ${sql.join(conditions, sql` AND `)}\n        ORDER BY ${distance} ASC\n        ${pageClause}\n      `;\n    }\n\n    // Brute-force scan.\n    const conditions: SQL[] = [sql`${quoted}.\"graph_id\" = ${params.graphId}`];\n    if (candidates !== undefined) {\n      conditions.push(sql`${quoted}.\"node_id\" IN (${candidates})`);\n    }\n    if (params.minScore !== undefined) {\n      conditions.push(\n        vectorMinScoreCondition(distance, params.metric, params.minScore),\n      );\n    }\n    const pageClause = vectorPageClause(params.limit, params.offset);\n    return sql`\n      SELECT ${quoted}.\"node_id\" AS node_id, ${score} AS score\n      FROM ${quoted}\n      WHERE ${sql.join(conditions, sql` AND `)}\n      ORDER BY ${distance} ASC\n      ${pageClause}\n    `;\n  },\n\n  distanceExpression(embeddingColumn, queryEmbedding, metric) {\n    return distanceExpression(embeddingColumn, queryEmbedding, metric);\n  },\n\n  buildCreateIndex(slot): SQL | undefined {\n    if (!usesAnnIndex(slot)) return undefined;\n    return sql.raw(\n      libsqlVectorIndexDdl(\n        this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n        slot,\n      ),\n    );\n  },\n\n  buildDropIndex(slot): SQL | undefined {\n    if (!usesAnnIndex(slot)) return undefined;\n    const indexName = libsqlIndexName(slot);\n    return sql.raw(`DROP INDEX IF EXISTS ${quoteIdentifier(indexName)}`);\n  },\n\n  buildDropStorage(slot): readonly string[] {\n    const table = quoteIdentifier(\n      this.tableName(slot.graphId, slot.nodeKind, slot.fieldPath),\n    );\n    const statements: string[] = [];\n    // Drop the DiskANN index first so libSQL reclaims its shadow tables,\n    // then the table itself.\n    if (usesAnnIndex(slot)) {\n      statements.push(\n        `DROP INDEX IF EXISTS ${quoteIdentifier(libsqlIndexName(slot))}`,\n      );\n    }\n    statements.push(`DROP TABLE IF EXISTS ${table}`);\n    return statements;\n  },\n};\n\n/**\n * DiskANN over-fetch multiplier applied when a candidate filter is present:\n * `vector_top_k` cannot pre-filter, so fetch `4k` neighbors and filter after.\n * Mirrors the hybrid facade's 4x over-fetch. Recall for the filtered search\n * is bounded by this headroom — if more than `3k` of the top `4k` neighbors\n * are filtered out, fewer than `k` rows return. That is the observable\n * meaning of `capabilities.filteredApproximateSearch.guaranteesFullPage`.\n */\nconst CANDIDATE_FILTER_OVERFETCH = 4;\n\nfunction libsqlIndexName(slot: VectorSlot): string {\n  return vectorPhysicalName(\n    INDEX_PREFIX,\n    slot.graphId,\n    slot.nodeKind,\n    slot.fieldPath,\n  );\n}\n\nfunction libsqlVectorIndexDdl(table: string, slot: VectorSlot): string {\n  const indexName = quoteIdentifier(libsqlIndexName(slot));\n  const metric = libsqlMetricToken(slot.metric);\n  return `CREATE INDEX IF NOT EXISTS ${indexName} ON ${quoteIdentifier(table)}(libsql_vector_idx(\"embedding\", 'metric=${metric}'));`;\n}\n"]}