Skip to content

Latest commit

 

History

History
139 lines (104 loc) · 4.29 KB

README.md

File metadata and controls

139 lines (104 loc) · 4.29 KB
Raw

Use FAISS for Self-Development?

🚀 FAISS to the rescue! If you're tired of high costs from services like Pinecone, especially for large-scale vector storage, using FAISS on your own AWS instance can save your budget. For example, hosting your own FAISS on a t3.medium EC2 instance can drastically cut down costs compared to the recurring fees for retrieval and upserts on Pinecone. With FAISS, you get scalable, high-performance vector searching at a fraction of the price—no subscription costs or hidden fees! Why pay more when you can build smarter? 😎

FAISS Vector Store API

This project provides a Flask-based API for managing FAISS vector stores with OpenAI embeddings. You can create vector stores, add vectors to them, and perform similarity searches.

Features

  • Create FAISS vector stores: Easily initialize new FAISS vector stores with a unique token.
  • Add vectors: Use OpenAI embeddings to generate vectors and add them to the store.
  • Similarity search: Perform top-K searches to find the most similar vectors based on input text.
  • JSON serializable responses: Ensure clean and valid JSON responses for all API endpoints.
  • Memory management: Vector stores are loaded and released from memory efficiently.

Run

python3 main.py

API Endpoints

1. Create a Vector Store

  • Endpoint: /create/<name>
  • Method: POST
  • Description: Creates a new FAISS vector store and returns a unique token.

Example request:

curl -X POST http://localhost:5000/create/my_store

Response:

{
  "token": "unique_token"
}

2. Add Vectors to a Store

  • Endpoint: /vector-token/index
  • Method: POST
  • Description: Adds vectors to the store using an auth token, with automatic metadata handling.

Example request:

curl -X POST http://localhost:5000/vector-token/index \
     -H "Content-Type: application/json" \
     -d '{"auth_token":"your_token", "texts": "your input text", "subject": "Test Subject"}'

Response:

{
  "message": "Vector added successfully"
}

3. Search for Vectors

  • Endpoint: /vector_token/search/<k>
  • Method: POST
  • Description: Performs a top-K similarity search based on input query context.

Example request:

curl -X POST http://localhost:5000/vector_token/search/5 \
     -H "Content-Type: application/json" \
     -d '{"auth_token":"your_token", "input_query_context": "search text"}'

Response:

{
  "results": [
    {
      "vector_id": 1,
      "distance": 290884,
      "metadata": {
        "input_text": "I wana sleep 4",
        "added_on": "2024-09-28-22-17",
        "vector_id": 1
      }
    },
    {
      "vector_id": 0,
      "distance": 999999,
      "metadata": {
        "input_text": "dummy_vector",
        "added_on": "2024-09-28-22-17",
        "vector_id": 0
      }
    }
  ]
}

Key Implementation Details

  • OpenAI Embeddings: The API uses OpenAI's embedding model (text-embedding-ada-002) to generate embeddings from text.
  • Vector Store Management: Each vector store is associated with a unique token. Stores are saved to disk and can be loaded or released from memory when needed.
  • Indexing and Searching: FAISS is used for efficient vector storage and similarity searching.

Improvements

  • Optimize Memory Handling: Automatically release vector stores from memory after every operation.
  • Distance Scaling: Convert distances to integers by multiplying by 1,000,000 to avoid floating-point precision issues in JSON responses.
  • Error Handling: Added error handling for invalid tokens and indices.

How to Contribute

  • Clone the repository:

    git clone https://github.com/wolfmib/alinex-faiss.git

  • Set up the environment:

    • Install Python dependencies: pip install -r requirements.txt
    • Create a .env file and add your OpenAI API key.

  • Run Tests:

    python -m unittest discover -s tests

License

This project is licensed under the MIT License. Please remember to attribute contributions where applicable.

How to Use It:

  • Replace the placeholders like your_token with actual values.
  • The Markdown is structured to be readable and organized with clear sections for features, API endpoints, and usage examples.

Let me know if you need any further adjustments!