Two-stage retrieval

TurboQwenRAG v1.1.0

update adding reranker, persistent index, anti-hallucination guard, multi-turn chat, and a full evaluation suite on top of the initial release.

Stack

• Embedder : Qwen3-Embedding-0.6B - instruction-aware, 8K context (up-to 32K context), 1024-dim
• Vector DB : turbovec IdMapIndex - TurboQuant 4-bit, Rust+SIMD, .tvim persistence
• Reranker : Qwen3-Reranker-0.6B - cross-encoder two-stage retrieval
• LLM : Qwen3-8B Q4_K_M - llama.cpp + full CUDA offload (~5GB VRAM)
• UI : Gradio 5.50.0 - dark theme, multi-tab, shareable Colab link

Features

• Auto load/build index (turbovec .tvim - no rebuild on restart)
• Confidence threshold (anti-hallucination guard, default 0.30)
• Multi-turn chat history (last 3 turns injected as context)
• RAG evaluation suite (hit rate, context coverage, confidence score)
• Supports PDF, TXT, and Markdown ingestion via PyMuPDF + LangChain

Initial Release

TurboQwenRAG v1.0.0

First public release of a fully local, zero-cost RAG system built on the Qwen3 ecosystem and turbovec TurboQuant.

Stack

• Embedder : Qwen3-Embedding-0.6B - instruction-aware, 8K context (up-to 32K context), 1024-dim
• Vector DB : turbovec IdMapIndex - TurboQuant 4-bit, Rust+SIMD, .tvim persistence
• LLM : Qwen3-8B Q4_K_M - llama.cpp + full CUDA offload (~5GB VRAM)
• UI : Gradio 5.50.0 - dark theme, multi-tab, shareable Colab link

Features

• End-to-end RAG pipeline
• turbovec IdMapIndex with TurboQuant 4-bit compression and .tvim persistence
• Instruction-aware embedding (separate prompts for query vs document)
• Gradio dark theme UI with file upload and chat interface
• Supports PDF, TXT, and Markdown via PyMuPDF + LangChain

Notes

Reranker, confidence threshold, multi-turn chat, and evaluation suite are not yet included in this version — see v1.1.0.