vlite is a lightweight, high-performance vector database designed for AI applications like agents and ChatGPT Plugins, addressing the need for a simple, fast, and portable solution for managing embeddings. It offers a novel CTX file format, akin to browser cookies for embeddings, enabling efficient storage, retrieval, and context management.

How It Works

vlite leverages a novel approach using NumPy for efficient vector operations, claiming superior performance and smaller disk footprints compared to alternatives like Chroma. Its design prioritizes simplicity, eliminating the need for server setup or complex configurations. The CTX file format is central to its portability and composability, allowing embeddings to be easily moved and integrated across different applications.

Quick Start & Requirements

• Install: pip install vlite
• Install with OCR: pip install vlite[ocr]
• Prerequisites: Python, NumPy. OCR support requires surya.
• Usage examples and LangChain integration are provided in the README.

Highlighted Details

• Claims <1.1s search time for 500k documents with binary embeddings.
• Includes built-in embedding generation via mixedbread-embed-large.
• Supports ingestion of Text, PDF, CSV, PPTX, and webpages, with PDF OCR.
• Benchmarked as >77.95% faster on indexing, >422% faster on retrieval, and >3.6x smaller than Chroma.

Maintenance & Community

• Contributions acknowledged from Claude and Ray.
• Open to contributions via issues or pull requests.

Licensing & Compatibility

• License: AGPL-3.0.
• AGPL-3.0 is a strong copyleft license, requiring derivative works to also be open-sourced under AGPL-3.0. This may restrict commercial use in closed-source applications without careful consideration or licensing agreements.

Limitations & Caveats

The AGPL-3.0 license imposes significant obligations for any modifications or integrations, potentially limiting its use in proprietary software. The project appears to be actively developed, and specific performance claims should be validated against user-specific workloads.