mirzaasadijaz/agentic-rag-langgraph repository overview

This image provides a self-correcting, agentic RAG (Retrieval-Augmented Generation) application where the underlying agent handles query routing, retrieval, grading, generation, and self-checking. It processes document queries using a combination of local and API-driven technologies, ensuring every retry loop is capped by a defined limit for guaranteed termination.

Key features included in this image:

• Agentic Control Flow: Utilizes a LangGraph StateGraph to route queries, grade retrieved chunk relevance, rewrite queries for better retrieval attempts, and fall back to Tavily web search if local documents cannot provide an answer.
• Advanced Retrieval: Implements hybrid retrieval fusing Chroma dense vectors with BM25 keyword search using weighted Reciprocal Rank Fusion, followed by local FlashRank cross-encoder reranking to ensure high accuracy.
• Dual LLM Architecture: Splits workloads between a grading model (such as Groq) for fast routing and decisions, and a generation model (such as Google Gemini) for the final response.
• Interactive UI: Includes a Streamlit chat interface featuring live agent timelines, per-answer source tracking, LaTeX and code rendering, and live document indexing via sidebar uploads.
• Secure & Optimized Container: Configured to run as an unprivileged, non-root user (uid 1000) with API secrets isolated from the build environment, utilizing CPU-only PyTorch by default to reduce image size.