JSON Documents Vector Search API

This is a simple API for vector similarity search built with FastAPI. It uses a JSONVectorSearch database utility for storing and querying vector embeddings.

Endpoints

• POST /create - Creates a new empty collection by calling db.clean()
• POST /add - Adds documents to the collection. Takes a list of ids, documents, and optional metadatas. Calls db.add() to index them.
• POST /query - Searches the collection. Takes a query string and number of n_results to return. Calls db.query_str() and returns the results.
• DELETE /clean - Deletes all documents in the collection by calling db.clean().

Usage

The documents passed to /add should be dictionaries containing a "vector" key with the vector embedding to index.

The query string passed to /query will be converted to a vector using the same process as indexing, and a similarity search performed to find the closest matching vectors.

The API expects vectors of fixed size for a given collection.

Requirements

• FastAPI
• Pydantic

Database

The JSONVectorSearch database utility handles storing the indexed vectors and metadata in JSON files. It provides methods for adding, cleaning, and querying the collection.

The vector search is approximate, optimized for speed over accuracy. The database is kept in-memory for low latency queries, but not persisted across restarts.

Next Steps

• Add persistence for the database
• Improve vector search accuracy
• Add more query options like filtering and faceting
• Containerize with Docker for easy deployment