ADR 003: PostgreSQL with pgvector for Vector Storage
Status
Accepted
Date
2024-02-15
Context
Mutx needs to store and query vector embeddings for agent memory and semantic search capabilities.
Decision
Use PostgreSQL with the pgvector extension for all data storage including:
- Structured metadata (agents, deployments, users)
- Vector embeddings for semantic search
- Time-series data for agent logs
Consequences
Positive
- Single database: Simplified operational overhead
- Vector support: pgvector provides efficient similarity search
- Mature: PostgreSQL is well-understood and stable
- ACID compliant: Strong consistency guarantees
Negative
- Vector performance: Not as optimized as dedicated vector DBs (Pinecone, Weaviate)
- Scaling: Vertical scaling limits vs. distributed vector databases
Alternatives Considered
- Pinecone: Rejected - added cost and vendor lock-in
- Weaviate: Rejected - operational complexity
- Elasticsearch: Rejected - overkill for our use case