h2oai/h2ogpt

Leverages persistent vector databases (Chroma, Weaviate, FAISS) with semantic chunking and HYDE retrieval to enable efficient document Q&A without few-shot prompting, while supporting multimodal inputs including vision models (LLaVa, Claude-3, GPT-4-Vision), voice I/O via Whisper and TTS, and image generation. Provides OpenAI-compatible API endpoints and integrates with multiple inference backends (vLLM, HF TGI, Together.ai, Groq) plus agent frameworks like AutoGen for code execution and tool calling with automatic model selection.

12,004 stars.