From 13cbb35977b7999c9c08670e44dabecb2e48aeaf Mon Sep 17 00:00:00 2001
From: Bazhen Rzheutskii <barzhomi@gmail.com>
Date: Fri, 29 May 2026 00:30:43 +0400
Subject: [PATCH] fix(indexing): prevent LanceDB metadata type coercion causing
 full re-index on restart
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

LanceDB infers the value column type from the first row in _createMetadataData().
Since vector_size (1024) was stored as a number, all subsequent string values
(embedding_provider, embedding_model_id) and booleans (indexing_complete) were
silently coerced to NaN/0/1, corrupting metadata.

On restart, _getStoredEmbeddingProfile() received NaN values, causing
needsRecreation=true every time — the database was dropped and recreated,
the file-hash cache cleared, and a full re-index triggered on every VS Code reopen.

Store all metadata values as strings to ensure consistent column type.
---
 .../vector-store/lancedb-vector-store.ts      | 20 ++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)

diff --git a/packages/kilo-indexing/src/indexing/vector-store/lancedb-vector-store.ts b/packages/kilo-indexing/src/indexing/vector-store/lancedb-vector-store.ts
index 39bc14760d..d4433d2e30 100644
--- a/packages/kilo-indexing/src/indexing/vector-store/lancedb-vector-store.ts
+++ b/packages/kilo-indexing/src/indexing/vector-store/lancedb-vector-store.ts
@@ -136,7 +136,7 @@ export class LanceDBVectorStore implements IVectorStore {
     return [
       {
         key: KEY.size,
-        value: this.vectorSize,
+        value: String(this.vectorSize),
       },
       {
         key: KEY.provider,
@@ -148,11 +148,11 @@ export class LanceDBVectorStore implements IVectorStore {
       },
       {
         key: KEY.dimension,
-        value: this.profile.dimension,
+        value: String(this.profile.dimension),
       },
       {
         key: KEY.complete,
-        value: false,
+        value: "false",
       },
     ]
   }
@@ -572,7 +572,7 @@ export class LanceDBVectorStore implements IVectorStore {
       }
       const metadataTable = await db.openTable(this.metadataTableName)
       const metadataResults = await metadataTable.query().where(`key = '${KEY.complete}'`).toArray()
-      const indexed = metadataResults.length > 0 ? metadataResults[0].value : false
+      const indexed = metadataResults.length > 0 ? String(metadataResults[0].value) === "true" : false
       log.info("LanceDB indexing metadata evaluated", {
         workspacePath: this.workspacePath,
         pointCount,
@@ -590,14 +590,16 @@ export class LanceDBVectorStore implements IVectorStore {
       throw new Error(`Invalid metadata key: ${key}`)
     }
     await metadataTable.delete(`key = '${key}'`)
-    await metadataTable.add([{ key, value }])
+    // All values must be strings to prevent LanceDB from inferring the value column
+    // type as number from the first row, which corrupts subsequent string/boolean values.
+    await metadataTable.add([{ key, value: String(value) }])
   }
 
   private async _persistEmbeddingProfile(metadataTable: Table): Promise<void> {
     await this._upsertMetadata(metadataTable, KEY.provider, this.profile.provider)
     await this._upsertMetadata(metadataTable, KEY.model, this.profile.modelId)
-    await this._upsertMetadata(metadataTable, KEY.dimension, this.profile.dimension)
-    await this._upsertMetadata(metadataTable, KEY.size, this.vectorSize)
+    await this._upsertMetadata(metadataTable, KEY.dimension, String(this.profile.dimension))
+    await this._upsertMetadata(metadataTable, KEY.size, String(this.vectorSize))
   }
 
   /**
@@ -609,7 +611,7 @@ export class LanceDBVectorStore implements IVectorStore {
       const db = await this.getDb()
       const metadataTable = await db.openTable(this.metadataTableName)
       await this._persistEmbeddingProfile(metadataTable)
-      await this._upsertMetadata(metadataTable, KEY.complete, true)
+      await this._upsertMetadata(metadataTable, KEY.complete, "true")
       log.info("Marked indexing as complete")
     } catch (error) {
       log.error("Failed to mark indexing as complete", { error })
@@ -626,7 +628,7 @@ export class LanceDBVectorStore implements IVectorStore {
       const db = await this.getDb()
       const metadataTable = await db.openTable(this.metadataTableName)
       await this._persistEmbeddingProfile(metadataTable)
-      await this._upsertMetadata(metadataTable, KEY.complete, false)
+      await this._upsertMetadata(metadataTable, KEY.complete, "false")
       log.info("Marked indexing as incomplete (in progress)")
     } catch (error) {
       log.error("Failed to mark indexing as incomplete", { error })