pgvector Semantic Search

Use this skill to design semantic search on PostgreSQL with pgvector in a way that balances relevance, filtering, and operational simplicity.

Clarify First

• What is being embedded and searched.
• Whether semantic search stands alone or augments keyword search.
• Metadata filters, freshness needs, and expected corpus size.
• Latency and recall expectations.
• How results will be evaluated.

Design Priorities

• Retrieval quality starts with chunking and embedding choices, not just index tuning.
• Metadata filtering often matters as much as vector similarity.
• Hybrid retrieval may outperform pure semantic search.
• Evaluate precision and failure modes with real queries.
• Keep update and re-embedding workflows in view.

Output

• Recommended retrieval strategy.
• Schema shape for vectors and metadata.
• Indexing and filtering considerations.
• Evaluation and quality-check suggestions.
• Tradeoffs around scale, cost, and operational complexity.

Common Mistakes

• Assuming better embeddings fix weak chunking.
• Ignoring metadata filters or freshness.
• Measuring only latency and not result quality.
• Treating vector search as a drop-in substitute for every search problem.