Implement vector search

[surister]

CrateDB supports vector search, there are two components for this feature, the column datatype FLOAT_VECTOR and the knn_match function to perform vector search.

Django does not support vector fields natively, we'd have to create it ourselves.

cc @amotl

[amotl]

Excellent, thanks. Maybe we can find a Django dialect from another vector database vendor for inspirational purposes, when applicable. In some adapters, I am deriving from / leaning on existing implementations for PostgreSQL's pgvector, in order to adhere to their interfaces, just swapping the implementation.

Those resources could possibly be relevant in this context.

vector / semantic search

• https://unfoldai.com/vector-database-django-pgvector/
• https://medium.com/@simeon.emanuilov/integrating-a-vector-database-into-django-using-pgvector-72322b9debbe
• https://dev.to/buddhiraz/using-pgvector-with-your-django-app-2o3j
• https://kacperlukawski.github.io/django-semantic-search/
• https://github.com/pkavumba/django-vectordb
• https://2025.djangocon.us/talks/beyond-filters-modern-search-and-more-with-vectors-in-django/

fts / full-text search

• https://www.julienphalip.com/blog/adding-search-to-a-django-site-in-a-snap/
• https://stackoverflow.com/questions/49867384/django-searchvector-using-icontains

[amotl]

Those resources could possibly be relevant in this context.

Ah, right. pgvector already includes a relevant implementation for Django. Maybe you can derive from that 1:1, when possible?

• https://github.com/pgvector/pgvector-python/tree/master/pgvector/django

FWIW, the LangChain adapter for CrateDB is based on the corresponding PostgreSQL adapter, and also reuses / wraps around / derives from its PGVector extension, also based on the same fundamental pgvector package.

• https://github.com/crate/langchain-cratedb/blob/v0.2.0/langchain_cratedb/vectorstores/main.py

FWIW, on the other hand, the CrateDB SQLAlchemy vendorizes two utility functions from pgvector, and adds own sugar to connect it with the framework.

• https://github.com/crate/sqlalchemy-cratedb/blob/0.41.0/src/sqlalchemy_cratedb/type/vector.py

What I want to say is: If you think it makes sense, don't be shy to add pgvector as a dependency to save keystrokes.

[amotl]

Regarding the django-semantic-search package by @kacperlukawski, it could be nice to have a presence as a backend, next to Qdrant. 🙏