TechyNilesh/DeepImageSearch

What it solves

DeepImageSearch is a Python library designed to simplify the creation of AI-powered image search systems. It removes the complexity of managing multimodal embeddings, vector indexing, and metadata storage, allowing users to find images using natural language (text-to-image), visually similar images (image-to-image), or a combination of both (hybrid search).

How it works

The library uses multimodal embedding models (such as CLIP, SigLIP, and EVA-CLIP) to convert images and text into a shared vector space. It then indexes these vectors using a vector store (FAISS, ChromaDB, or Qdrant) and tracks image details via a metadata store (JSON or PostgreSQL).

Key components include:

• Embedding Manager: Handles the generation of vectors from images and text.
• Indexer: Processes images and optionally generates captions using LLMs via OpenAI-compatible APIs.
• Hyrbid Search: Combines text and image queries using weighted fusion to refine results.
• Agentic Integration: Provides an MCP server and LangChain tools to allow AI agents (like Claude) to perform image searches.

Who it’s for

Developers building image-centric applications, AI agents, and RAG pipelines that require efficient retrieval of visual assets based on semantic meaning rather than just filenames or tags.

Highlights

• Multimodal Search: Supports text-to-image, image-to-image, and hybrid search modes.
• Flexible Vector Stores: Out-of-the-box support for FAISS, ChromaDB, and Qdrant.
• LLM Captioning: Automatically generates image captions using any OpenAI SDK-compatible provider.
• Agent-Ready: Includes an MCP server for Claude and a LangChain tool for agentic workflows.
• Production-Ready Metadata: Supports PostgreSQL for scalable image record management.
• Hardware Acceleration: Auto-detects and supports CUDA, MPS (Apple Silicon), and CPU.