grafana / vectorapi Goto Github PK

Similar to /query, except taking in a text input instead of a vector input

Options:

• A separate endpoint (e.g. /v1/collections/<collection_name>/search) that only takes text input
• Modify /v1/collections/<collection_name>/query to have flexible input (either text input or vector input)

We should maybe also support a generic search without going to a specific collection first, e.g.

• /v1/search that takes a collection_name in the input
• Stretch goal: multiple collection names to search across multiple collections?

Currently we expect a model to be specific at query time, but the model use should just be the same model used to initialize the collection.

Might be better to just store the model name in the collection details.

Baseline done in #10

At the moment we only run sync at the start up which means that if a collection is created by another pod then this collection won't come up by other vectorapi pods.

Add sync on pgvector client get_collection() to solve this issue.

Getting the following error

vectorapi-api-1       | asyncpg.exceptions.TooManyConnectionsError: sorry, too many clients already

When running upserts during collection setup.

reproduce steps:

1. start vectorapi: make up
2. run custom initialization script (e.g. make store in llm-experiment-lab)

For testing and docker build

• build docker image (done in #9)
• run unit tests in CI
• run integration tests in CI (with db)

• pgvector