Purpose: The goal of FIndex is to learn more about vector databases by building a practical system for face recognition and similarity search. This project demonstrates how to generate, store, and query facial embeddings using a vector database and RESTful APIs.

FIndex is a face recognition and search system powered by embeddings, Qdrant vector database, and a lightweight REST API written in .NET 8. It allows you to embed faces from images, store their vectors, and search visually similar faces efficiently.

• A FastAPI-based microservice that takes an image and returns a 512-dimensional facial embedding.
• Uses InsightFace under the hood.

• High-performance vector search engine for similarity search.
• Stores all embedded facial vectors.

• Accepts image uploads via REST.
• Sends them to the embedding service, gets vector back.
• Searches similar vectors in Qdrant and returns matches with confidence scores.

• Console app to upload faces and create database.

• Asp.net web api service that takes an image and returns a 512-dimensional facial embedding.
• Powered by ONNX Runtime and OpenCvSharp, using a pre-trained ResNet50 (WebFace600K) model. Download Link: Google Drive.

• Standalone console sample that demonstrates embedding a small face dataset and performing searches backed by PostgreSQL with the pgvector extension.
• Expects the same webface_r50.onnx model as the embedder; place it under FIndex.VectorDataSample/Data/ after downloading from the link above.

1. Ensure Docker Desktop (or another Docker host) is running locally.
2. (Optional) Download the InsightFace models you want to serve and place them under ~/.insightface/models; they are mounted into the embedding container so they persist across rebuilds.
3. Build and start the stack:

docker compose up --build

The compose file spins up the Python embedding service, the C# embedder, Qdrant, and the public API. All services are wired together on an internal Docker network, so no additional setup is required.

To stop the services press Ctrl+C, or run docker compose down in another terminal to remove the containers.

• The API is wired to the Python embedding service by default (embedding-service in docker-compose.yml, exposed on port 8001).
• The C# embedder (embedder-service, exposed on port 5443) is available for experimentation. You can point your client applications to it directly, or adjust the API to call it instead of the Python endpoint if you prefer an all-.NET stack.
• If you only need one implementation, comment out or remove the other service from docker-compose.yml before running docker compose up.

This project consists of three main services: the Python-based embedding service, the Qdrant vector database, and the .NET API. Below are the details for accessing each service:

This project's code is licensed under the MIT License. You are free to use, modify, and distribute the code with proper attribution.

Some images used in this project are for personal use and local experimentation only.

By using this repository, you agree not to use any of the stored or uploaded images (I'm not that pretty).