pgvector-setup

# pgvector-setup

## Instructions

This skill provides complete pgvector setup for Supabase databases, enabling vector search capabilities for AI applications, RAG systems, and semantic search.

### Phase 1: Enable pgvector Extension

1. Run the setup script to enable pgvector:
   ```bash
   bash scripts/setup-pgvector.sh [SUPABASE_DB_URL]
   ```
   This creates the pgvector extension and sets up basic embedding tables.

2. Choose your embedding dimensions based on your model:
   - OpenAI text-embedding-3-small: 1536 dimensions
   - OpenAI text-embedding-3-large: 3072 dimensions
   - Cohere embed-english-v3.0: 1024 dimensions
   - Custom models: Check model documentation

### Phase 2: Create Embedding Tables

1. Use the embedding table template:
   ```bash
   # Copy template and customize for your use case
   cat templates/embedding-table-schema.sql
   ```

2. Customize the schema:
   - Adjust vector dimensions to match your model
   - Add metadata columns (tags, timestamps, user_id, etc.)
   - Configure RLS policies for security

3. Apply the schema:
   ```bash
   psql $SUPABASE_DB_URL < templates/embedding-table-schema.sql
   ```

### Phase 3: Create Vector Indexes

**Choose index type based on your data size:**

**HNSW (Recommended for most cases):**
- Best for: < 1M vectors, high recall requirements
- Pros: Fast queries, good recall, works well with small-medium datasets
- Cons: Slower inserts, higher memory usage
- Run: `bash scripts/create-indexes.sh hnsw [TABLE_NAME] [DIMENSION]`

**IVFFlat:**
- Best for: > 1M vectors, write-heavy workloads
- Pros: Faster inserts, lower memory
- Cons: Requires training, lower recall
- Run: `bash scripts/create-indexes.sh ivfflat [TABLE_NAME] [DIMENSION]`

**Performance Tuning:**
- HNSW m parameter (default 16): Higher = better recall, more memory
- HNSW ef_construction (default 64): Higher = better quality, slower builds
- IVFFlat lists (default sqrt(rows)): More lists = faster queries, lower recall

### Phase 4: Implement Se