FastEmbed Experiments

This repository contains experiments and services using FastEmbed, a lightweight, fast, Python library for embedding generation.

Projects

1. Embedding API

A containerized FastAPI service for generating dense and sparse embeddings.

Directory: services/embedding-api

Quick Start (Docker)

# Build the image (use DOCKER_BUILDKIT=0 if you encounter 403 errors with base image)
DOCKER_BUILDKIT=0 docker build -t fastembed-api:latest services/embedding-api

# Run the container
docker run -d -p 8000:8000 --name fastembed-api fastembed-api:latest

# Check status
curl http://localhost:8000/health

Local Development

You can run the service locally using the provided helper scripts:

# Start the service (background)
./services/embedding-api/start.sh

# Stop the service
./services/embedding-api/stop.sh

Logs are written to services/embedding-api/service.log.

Endpoints

• POST /embed/text: Generate dense embeddings.
  • Input: {"documents": ["text1", "text2"]}
  • Output: {"embeddings": [[0.1, ...], [0.2, ...]]}
  • Default Model: BAAI/bge-small-en-v1.5
• POST /embed/sparse: Generate sparse (SPLADE) embeddings.
  • Input: {"documents": ["text1", "text2"]}
  • Output: {"embeddings": [{"indices": [...], "values": [...]}, ...]}
  • Default Model: prithivida/Splade_PP_en_v1
• POST /embed/image: Generate dense image embeddings.
  • Input: {"images": ["base64_string_1", "base64_string_2"]}
  • Output: {"embeddings": [[0.1, ...], [0.2, ...]]}
  • Default Model: Qdrant/clip-ViT-B-32-vision

Configuration is managed via .env (see services/embedding-api/src/config.py).

2. Notebooks

Exploratory notebooks for various embedding tasks:

• try_fastembed_text_embeding.ipynb: Basic dense text embedding.
• try_fastembed_sparse_embedding.ipynb: Sparse embedding (SPLADE) exploration.
• try_fastembed_image_embedding.ipynb: Image embedding experiments.