Vector Databases

Compozy decouples workflow metadata (PostgreSQL or SQLite) from vector storage. Your choice of metadata database dictates which vector providers are available.

Driver Compatibility

Metadata Driver	Vector Provider	Notes
PostgreSQL	`pgvector`	Native integration when the `vector` extension is installed.
PostgreSQL	Qdrant, Redis, Filesystem	Supported via external services if you prefer to separate workloads.
SQLite	Qdrant (recommended)	External vector database required for knowledge bases.
SQLite	Redis with RediSearch	Works for low-latency deployments with managed Redis.
SQLite	Filesystem	Development-only option for lightweight prototypes.

SQLite deployments cannot use pgvector. Attempting to configure provider: pgvector results in: pgvector provider is incompatible with SQLite driver. Select an external vector database instead.

pgvector (PostgreSQL Only)

Enable pgvector

CREATE EXTENSION IF NOT EXISTS vector;

Configuration

knowledge:
  vector_dbs:
    - id: main
      provider: pgvector
      dimension: 1536

Works alongside metadata in the same PostgreSQL cluster.
Ensure dimension matches the embedding model output.
Managed databases (Neon, Supabase, RDS) often provide pgvector as a toggle or automatic extension.

Qdrant (Recommended with SQLite)

Qdrant configuration

knowledge:
  vector_dbs:
    - id: main
      provider: qdrant
      url: http://localhost:6333
      dimension: 1536
      collection: compozy_embeddings

Production-ready open-source vector database.
Supports filtering, payloads, and distributed deployments.
Works with both PostgreSQL and SQLite metadata drivers.

Redis with RediSearch

Redis vector store

knowledge:
  vector_dbs:
    - id: cache
      provider: redis
      url: redis://localhost:6379
      dialect: search
      dimension: 1536
      index: compozy:kb

Utilize managed Redis offerings with RediSearch enabled.
Ideal when you already depend on Redis for caching or rate limiting.
Memory-bound; monitor dataset size closely.

Filesystem Provider (Development Only)

Filesystem provider

knowledge:
  vector_dbs:
    - id: local
      provider: filesystem
      path: ./data/embeddings
      dimension: 1536

Stores embeddings as structured files on disk.
Great for unit tests or quick demos.
Not recommended for production due to limited concurrency and durability.

Decision Matrix

Requirement	Recommended Provider
Production deployment with PostgreSQL	`pgvector`
Production deployment with SQLite	`qdrant`
Edge deployment with minimal services	`filesystem` (development) or `qdrant` (managed)
Shared cache infrastructure already in place	`redis`

Operational Tips

Dimension Consistency
Ensure every provider uses the same embedding dimension across collections.
Health Checks
Add readiness probes for external vector services to avoid startup race conditions.
Access Control
Secure API keys or tokens for Qdrant/Redis; never store them in plain text inside repositories.
Monitoring
Track latency and memory usage, especially when running Redis or Qdrant alongside Compozy.

Database Overview

Understand how metadata driver selection impacts vector storage.

PostgreSQL Driver

Configure pgvector with PostgreSQL.

SQLite Driver

Set up external vector databases alongside SQLite.

Vector Databases

On this page