Semantic Search

An open-source Python library for semantic search, featuring:

• FAISS for rapid vector similarity.
• SentenceTransformers for high-quality embeddings.
• Pluggable database backends (MongoDB, SQLite, Redis, PostgreSQL, MySQL).

With this library, you can efficiently store, index, and retrieve documents based on semantic similarity.

Table of Contents

1. Features
2. Databases & FAISS Table
3. Installation
4. Usage
5. Testing
6. FAQ

Features

• Multi-DB Support – Choose MongoDB, SQLite, Redis, PostgreSQL, or MySQL.
• FAISS Index – Build a fast, in-memory index for vector searches.
• Easy API – add_document(), build_faiss_index(), retrieve().
• Scalable – Handle millions of embeddings with FAISS.
• Open Source – Contributions are welcome!

Databases & FAISS Table

FAISS:

• IndexFlatL2 used for demonstration (simple & effective for smaller data).
• For large data: consider IVF, HNSW, or GPU-based indexing.

Installation

1. Clone or Download this repo:

   git clone https://github.com/username/semantic-search.git
   cd semantic-search
   

2. Install Dependencies:

    pip install -r requirements.txt
    
    Make sure to adjust dependencies (faiss-cpu vs. faiss-gpu) depending on your environment.
    If you plan to use PyTorch on GPU, install torch with CUDA support.

Usage

Below is a basic example using MongoDB as the database backend. You can switch to other backends by changing db_type.

from semantic_search import SemanticSearch, DatabaseFactory

# 1. Create a database connection (MongoDB example)
db = DatabaseFactory.create_database(
    db_type="mongodb",
    mongo_uri="mongodb://localhost:27017/",
    db_name="semantic_db",
    collection_name="documents"
)

# 2. Initialize SemanticSearch with the DB
search_engine = SemanticSearch(database=db)

# 3. If documents already exist, build FAISS index
try:
    search_engine.build_faiss_index()
except ValueError:
    print("No existing documents found. The FAISS index was not built.")

# 4. Add a new document
search_engine.add_document("Deep learning for NLP is a powerful tool.")

# 5. Retrieve similar documents
query = "Best techniques for NLP deep learning?"
results = search_engine.retrieve(query, top_k=3)
print("Results:", results)

Testing

This library includes pytest tests in the tests/ directory. To run them locally:

1. Install dev dependencies (pytest, pylint, etc.) from requirements.txt.
   
2. Run

python -m pytest tests/

FAQ

1. Why do I see a type error about faiss_index.add(x)?

FAISS’s Python bindings are generated via SWIG, so static type checkers (like Pyright) think the signature is add(n, x). We add # type: ignore to bypass this false positive. Runtime usage works fine with add(x).

2. Is .cpu() needed for embeddings?

• By default, SentenceTransformers returns NumPy arrays if you pass convert_to_tensor=False, so .cpu() is not needed.
• If you use convert_to_tensor=True, you get a PyTorch tensor. Convert it with:

  embedding = embedding.cpu().numpy().astype(np.float32)
  

3. Which database is best?

• MongoDB or PostgreSQL for large data.
• Redis for fast in-memory lookups.
• SQLite for small local apps.

4. Can I store embeddings in the DB & FAISS on disk?

Currently, IndexFlatL2 is in-memory only. To store FAISS on disk, use other FAISS indexes with I/O support or HDF5-based approach.

5. How do I contribute?

Fork this repo, create a new branch, and submit a PR with changes.
Add tests in tests/.