A production-grade Retrieval-Augmented Generation system for dense business and research document corpora. Designed for research analysts and strategy consultants who need evidence-grounded, cited answers across multi-document collections — annual reports, DRHPs, industry reports, expert transcripts.

Live demo: deep-research-rag-production.up.railway.app

Analysts spend hours manually reading documents to locate specific data points, cross-reference claims across sources, and build evidence-grounded views. Keyword search (Ctrl+F) is shallow. A basic chatbot is unreliable — it hallucinates and cannot cite sources.

This system solves that with a 7-stage pipeline that retrieves the most relevant context before generating an answer, and cites every claim to a specific source document and page number.

Component Library / Service API framework FastAPI + Uvicorn Vector store ChromaDB Dense embeddings sentence-transformers (all-MiniLM-L6-v2) Sparse retrieval rank-bm25 Reranking cross-encoder/ms-marco-MiniLM-L-6-v2 LLM generation Groq API (compound-beta-mini) Containerisation Docker + Docker Compose Deployment Railway (Dockerised, Southeast Asia region) Evaluation Local RAGAS-equivalent metrics (no API calls)

Evaluated on 13 ground-truth question-answer pairs across the ingested corpus.

Metric Score What it means Context Precision 0.991 Retrieved chunks are almost entirely relevant Context Recall 1.000 Every reference sentence is covered by the retrieved context Faithfulness 0.385 See note below Answer Relevancy 0.462 Correlates with faithfulness — same root cause

Note on Faithfulness: The 0.385 is a measurement artefact, not hallucination. The cross-encoder faithfulness metric penalises terse one-liner answers (produced for numerical fact queries) because short answer sentences score in a different logit range than full sentences when paired with long context paragraphs. Conceptual questions — where the model produces multi-sentence answers — all scored 1.000. Fix: modify the generation prompt to require full-sentence answers for all query types.

Then open http://localhost:7860 for the dark-mode browser interface with example questions, inline answers, and page-level source citations.

Interactive docs: http://localhost:8000/docs

The ChromaDB vector store is baked into the Docker image. HuggingFace model weights are cached in a named Docker volume (huggingface_cache) so they download once and persist across container restarts.

Returns pipeline readiness status.

Runs the full 7-stage pipeline.

422 — empty or missing question field

500 — pipeline error (check detail field)