tzurot-db-vector

Invoke with /tzurot-db-vector for migration and database operations.

Rules for queries and caching are in .claude/rules/03-database.md - they apply automatically.

Migration Procedure

1. Create Migration

# Interactive (prompts for name)
pnpm ops db:safe-migrate

# Non-interactive (AI assistants, CI)
pnpm ops db:safe-migrate --name <migration_name>

This command:

1. Tries prisma migrate dev --create-only (interactive path)
2. Falls back to prisma migrate diff if stdin is not a TTY (non-interactive)
3. Removes known drift patterns from prisma/drift-ignore.json
4. Reports what was sanitized
5. Shows the clean migration for review

All pnpm ops db:* commands work in non-interactive environments.

2. Review SQL (CRITICAL)

The safe-migrate command auto-sanitizes these, but always verify:

• DROP INDEX "idx_memories_embedding" — removed (IVFFlat vector index)
• DROP INDEX "memories_chunk_group_id_idx" — removed (partial index)
• CREATE INDEX "memories_chunk_group_id_idx" without WHERE — removed (non-partial)

3. Apply Migration

# Apply locally + regenerate Prisma client
pnpm ops db:migrate

# Regenerate PGLite test schema
pnpm ops test:generate-schema

# Apply to Railway
pnpm ops db:migrate --env dev
pnpm ops db:migrate --env prod --force  # Prod requires --force

db:migrate uses prisma migrate deploy (not migrate dev) for all environments. This avoids interactive prompts and drift-detection loops from sanitized indexes.

Drift Detection & Fix

# Check all migrations for drift
pnpm ops db:check-drift

# Fix specific migration (non-destructive)
pnpm ops db:fix-drift 20251213200000_add_tombstones

When safe to fix: Formatting/whitespace changes only When NOT to fix: Actual SQL logic was changed → create new migration

Database Inspection

# Show tables, indexes, migrations
pnpm ops db:inspect

# Inspect specific table
pnpm ops db:inspect --table memories

# Show indexes only
pnpm ops db:inspect --indexes

Vector Operations

Store Embedding

const embeddingStr = `[${embedding.join(',')}]`;
await prisma.$executeRaw`
  INSERT INTO memories (id, "personalityId", content, embedding, "createdAt")
  VALUES (gen_random_uuid(), ${id}::uuid, ${content}, ${embeddingStr}::vector, NOW())
`;

Similarity Search

// Cosine distance: 0 = identical, 2 = opposite
const results = await prisma.$queryRaw<SimilarMemory[]>`
  SELECT id, content, 1 - (embedding <-> ${embeddingStr}::vector) as similarity
  FROM memories
  WHERE "personalityId" = ${personalityId}::uuid
  ORDER BY embedding <-> ${embeddingStr}::vector
  LIMIT ${limit}
`;

Protected Indexes (CRITICAL)

| Index | Type | Why Protected | | ----------------------------- | -------------- | ----------------------------------------------- | | idx_memories_embedding | IVFFlat vector | Prisma doesn't support Unsupported type indexes | | memories_chunk_group_id_idx | Partial B-tree | Prisma can't represent WHERE clauses |

Recover If Dropped

-- IVFFlat vector index
CREATE INDEX IF NOT EXISTS idx_memories_embedding
ON memories USING ivfflat (embedding vector_cosine_ops)
WITH (lists = 50);

-- Partial index for chunk groups
DROP INDEX IF EXISTS "memories_chunk_group_id_idx";
CREATE INDEX "memories_chunk_group_id_idx" ON "memories"("chunk_group_id")
WHERE "chunk_group_id" IS NOT NULL;

PGLite Schema Regeneration

After any migration:

pnpm ops test:generate-schema
# Output: packages/test-utils/schema/pglite-schema.sql

References

• Prisma docs: https://www.prisma.io/docs
• pgvector docs: https://github.com/pgvector/pgvector
• Schema: prisma/schema.prisma
• Rules: .claude/rules/03-database.md