Question: Can I use pgvector/Postgres on the backend with the pixeltable?

[qdrddr]

Can I use the full instance of pgvector/Postgres on the backend with the pixeltable?

[pierrebrunelle]

Short answer: No, you can't bring your own pgvector/Postgres instance. Pixeltable isn't just using PostgreSQL—it's built on top of PostgreSQL with extensive custom implementations that require complete control of the database schema and lifecycle.

Why it's not swappable:

Custom Metadata Schema: Pixeltable creates its own system tables (systeminfo, dirs, tables, etc.) with complex JSONB metadata that tracks table versioning, column dependencies, incremental computation state, and view/snapshot lineage.

Incremental Computation Engine: The core value prop—automatic recomputation when data/code changes—requires tight integration with PostgreSQL's transaction system. Pixeltable tracks exactly which computed columns need updates based on dependency graphs stored in its metadata tables.

Multimodal Type System: Pixeltable's types (pxt.Image, pxt.Video) use hybrid storage: structured metadata in PostgreSQL, actual media files in ~/.pixeltable/media/, with automatic URL mapping between them.

Unified Query Planning: Vector similarity, computed column evaluation, and SQL filtering all get coordinated through Pixeltable's custom query planner that generates execution plans spanning both PostgreSQL queries and Python UDF execution.

Schema Migration Management: Pixeltable manages its own schema versioning with automatic upgrades that would conflict with external database management.

Bottom line: Pixeltable is the database, orchestration, and retrieval layers, not just a wrapper that uses PostgreSQL. You can decide to use one of them or all of them for the needs of your application alongside other services if you have requirements such as latency, availability and want to use Pixeltable mostly for orchestrating multimodal AI workloads.