what is 8hr-marathon fr?

This repository, titled Enterprise Agentic RAG, is a system that answers questions by first retrieving relevant information from a set of documents and then generating an answer with a language model, an approach commonly called RAG. It is aimed at enterprise style document question answering and is built to tell apart genuinely relevant technical content from irrelevant or noisy documents before answering. The system checks incoming questions with a safety layer called NeMo Guardrails before doing anything else, blocking off topic requests, jailbreak attempts, and prompt injection. A planner then decides whether a question is simple conversation or something that needs document retrieval. Technical questions go through a retriever that searches a Qdrant vector database, then a local reranker called FlashRank narrows the results down before a final responder generates the answer, with LangGraph keeping track of conversation memory across the whole flow. Language model calls are routed through a gateway called Portkey, which can automatically fall back to a backup key if the primary one fails, and the underlying model used is Llama 3.3 70B served through Groq. Document embeddings for search come from Google's Gemini embedding model. Document parsing for PDF, HTML, TXT, DOCX, and PPTX files happens entirely on the user's own machine, with no external service needed. The project also includes tracing through Pydantic Logfire and LangSmith, and an evaluation suite built on RAGAS with six metrics plus a Streamlit app to view results. To run it, you create a Python virtual environment, install the pinned dependencies, and fill in a .env file with API keys for Groq, Portkey, Qdrant, Logfire, LangSmith, and Gemini. After running an ingestion command that parses and indexes the sample documents, you start a FastAPI backend and a separate Streamlit chat interface in two terminals. An optional evaluation app can be run against the live backend afterward. The README includes eleven separate documentation guides covering the system's architecture, environment variables, and known gotchas in detail.