fix: repair pgvector migration to actually convert embedding column from text to vector(1536)
All checks were successful
Deploy to Production / Build and Deploy (push) Successful in 4s

The original migration used a fragile add-copy-drop-rename pattern with
_jsonb casts that silently failed, leaving the embedding column as text.
Replace with a direct ALTER COLUMN TYPE ... USING embedding::vector(1536)
that is fully idempotent and handles all partial states from previous
failed attempts.
This commit is contained in:
Antigravity
2026-05-12 08:00:37 +00:00
parent bbe54cf656
commit feaeb075ce
2 changed files with 62 additions and 30 deletions

37
DEPLOY-ISSUES-3.md Normal file
View File

@@ -0,0 +1,37 @@
# Search Broken — embedding column not converted to vector type
## Date: 2026-05-12
## Problem
The search fails with this error:
```
operator does not exist: text <=> vector
HINT: No operator matches the given name and argument types. You might need to add explicit type casts.
```
## Root cause
The `NoteEmbedding.embedding` column is still type `text` (old JSON string format), NOT `vector(1536)`.
The Prisma migration marked itself as applied but the actual column type conversion was never executed.
The SQL query tries to use the `<=>` cosine distance operator on a text column, which fails.
## Current state
- pgvector extension IS installed (CREATE EXTENSION worked)
- But the embedding column was NOT converted from text to vector(1536)
- There are 102 rows in NoteEmbedding with JSON string embeddings
- The migration SQL needs to: ALTER COLUMN embedding TYPE vector(1536) using proper casting
## What needs to happen
1. Check the actual column type: SELECT column_name, data_type, udt_name FROM information_schema.columns WHERE table_name = 'NoteEmbedding' AND column_name = 'embedding';
2. The migration SQL must convert the column. The embedding values are stored as JSON strings like "[0.1, 0.2, ...]" — need to strip brackets, then cast to vector.
3. The conversion SQL should be something like:
ALTER TABLE "NoteEmbedding" ALTER COLUMN embedding TYPE vector(1536) USING embedding::vector(1536);
OR if stored as JSON string:
ALTER TABLE "NoteEmbedding" ALTER COLUMN embedding TYPE vector(1536) USING (replace(replace(embedding, '[', ''), ']', ''))::vector(1536);
4. Also check if the tsvector column and trigger on Note table were created properly.
5. The semantic-search.service.ts code uses $queryRawUnsafe with <=> operator — make sure the SQL is correct for pgvector.
## Files to check/fix
- prisma/migrations/20260512120000_pgvector_and_fts_search/migration.sql — the actual migration SQL
- lib/ai/services/semantic-search.service.ts — the search service using vector queries
- lib/ai/services/embedding.service.ts — embedding service
- schema.prisma — NoteEmbedding model

View File

@@ -1,48 +1,43 @@
-- Phase 1: Enable pgvector extension -- Phase 1: Enable pgvector extension
CREATE EXTENSION IF NOT EXISTS vector; CREATE EXTENSION IF NOT EXISTS vector;
-- Phase 2: Convert embedding column from text/JSON to vector(1536) if needed -- Phase 2: Convert embedding column from text to vector(1536)
-- Idempotent: detects current column type and only converts when needed.
-- Handles all partial states from previous failed migration attempts:
-- A) embedding is text → direct ALTER COLUMN TYPE conversion
-- B) embedding already vector → skip
-- C) embedding missing, _vec_tmp exists → rename
DO $$ DO $$
DECLARE DECLARE
_udt text; _emb_type text;
_vec_tmp_exists boolean; _tmp_type text;
BEGIN BEGIN
SELECT udt_name INTO _udt SELECT udt_name INTO _emb_type
FROM information_schema.columns FROM information_schema.columns
WHERE table_schema = 'public' WHERE table_schema = 'public'
AND table_name = 'NoteEmbedding' AND table_name = 'NoteEmbedding'
AND column_name = 'embedding'; AND column_name = 'embedding';
IF _udt IS NOT NULL AND _udt != 'vector' THEN IF _emb_type = 'vector' THEN
SELECT EXISTS ( RETURN;
SELECT 1 FROM information_schema.columns END IF;
WHERE table_schema = 'public'
AND table_name = 'NoteEmbedding'
AND column_name = '_vec_tmp'
) INTO _vec_tmp_exists;
IF NOT _vec_tmp_exists THEN IF _emb_type IS NOT NULL THEN
ALTER TABLE "NoteEmbedding" ADD COLUMN "_vec_tmp" vector(1536); ALTER TABLE "NoteEmbedding" DROP COLUMN IF EXISTS "_vec_tmp";
END IF; ALTER TABLE "NoteEmbedding"
ALTER COLUMN "embedding" TYPE vector(1536)
USING "embedding"::vector(1536);
RETURN;
END IF;
UPDATE "NoteEmbedding" SELECT udt_name INTO _tmp_type
SET "_vec_tmp" = ("embedding"::jsonb)::text::vector(1536) FROM information_schema.columns
WHERE "embedding" IS NOT NULL WHERE table_schema = 'public'
AND "_vec_tmp" IS NULL; AND table_name = 'NoteEmbedding'
AND column_name = '_vec_tmp';
ALTER TABLE "NoteEmbedding" DROP COLUMN "embedding"; IF _tmp_type IS NOT NULL THEN
ALTER TABLE "NoteEmbedding" RENAME COLUMN "_vec_tmp" TO "embedding"; ALTER TABLE "NoteEmbedding" RENAME COLUMN "_vec_tmp" TO "embedding";
ELSIF _udt IS NULL THEN
SELECT EXISTS (
SELECT 1 FROM information_schema.columns
WHERE table_schema = 'public'
AND table_name = 'NoteEmbedding'
AND column_name = '_vec_tmp'
) INTO _vec_tmp_exists;
IF _vec_tmp_exists THEN
ALTER TABLE "NoteEmbedding" RENAME COLUMN "_vec_tmp" TO "embedding";
END IF;
END IF; END IF;
END $$; END $$;