How rohithreddy39/mediq-rag works
MediQ is a Medicare-policy RAG assistant: administrators ask questions and receive answers grounded in retrieved policy passages, with source citations and safety checks. The documented workflow runs from PDF ingestion through chunking, tagging and embedding, into dense-plus-sparse retrieval, prompt construction, generation and response checks. The product also exposes a Streamlit chat UI and REST API, and includes an evaluation pipeline. Some UI and evaluation internals are unsampled, so their connections are limited to relationships shown in excerpts, the source index or README.
Main entry points: pipeline.py Ingestion pipeline · eval_runner.py Evaluation runner · main.py REST API · app.py Streamlit chat UI
Components 21
Document ingestion
- Configuration
mediq/config.py - PDF parser
mediq/ingestion/pdf_parser.py - Text chunker
mediq/ingestion/chunker.py - Metadata tagger
mediq/ingestion/metadata_tagger.py - Embedder
mediq/ingestion/embedder.py - Ingestion pipeline
mediq/ingestion/pipeline.py - Processed chunks
data/processed/chunks.json
Hybrid retrieval
- FAISS vector store
mediq/retrieval/vector_store.py - BM25 index
mediq/retrieval/bm25_index.py - Hybrid retriever
mediq/retrieval/hybrid_retriever.py
Answer generation
- Guardrails
mediq/generation/guardrails.py - Prompt builder
mediq/generation/prompt_builder.py - LLM client
mediq/generation/llm_client.py
User interfaces
- REST API
mediq/api/main.py - Streamlit chat UI
mediq/ui/app.py
Evaluation
- Evaluation runner
mediq/evaluation/eval_runner.py - Evaluation scorers
mediq/evaluation/scorers.py - Golden dataset
mediq/evaluation/golden_dataset.json
Other components
- Administrator
- Medicare PDFs
- OpenAI API
Connections 29
- Administrator → Streamlit chat UI: asks questions
README.md - Administrator → REST API: submits queries
README.md - Medicare PDFs → Ingestion pipeline: provides documents
README.md - Ingestion pipeline → PDF parser: parses PDFs
mediq/ingestion/pipeline.py - Ingestion pipeline → Text chunker: chunks pages
mediq/ingestion/pipeline.py - Ingestion pipeline → Metadata tagger: tags chunks
mediq/ingestion/pipeline.py - Ingestion pipeline → Embedder: embeds chunks
mediq/ingestion/pipeline.py - Ingestion pipeline → Processed chunks: writes processed data
README.md - Configuration → Ingestion pipeline: configures paths
mediq/config.py - Processed chunks → REST API: loads chunks
mediq/api/main.py - Processed chunks → Streamlit chat UI: loads chunks
mediq/ui/app.py - REST API → Hybrid retriever: retrieves evidence
mediq/api/main.py - REST API → Prompt builder: builds prompt
mediq/api/main.py - REST API → LLM client: generates answer
mediq/api/main.py - REST API → Guardrails: checks safety
mediq/api/main.py - Streamlit chat UI → Hybrid retriever: retrieves evidence
mediq/ui/app.py - Streamlit chat UI → Prompt builder: builds prompt
mediq/ui/app.py - Streamlit chat UI → LLM client: generates answer
mediq/ui/app.py - Streamlit chat UI → Guardrails: checks safety
mediq/ui/app.py - Hybrid retriever → FAISS vector store: searches dense
mediq/retrieval/hybrid_retriever.py - Hybrid retriever → BM25 index: searches sparse
mediq/retrieval/hybrid_retriever.py - Hybrid retriever → Prompt builder: returns ranked chunks
- LLM client → OpenAI API: calls model
mediq/generation/llm_client.py - Evaluation runner → Hybrid retriever: evaluates retrieval
mediq/evaluation/eval_runner.py - Evaluation runner → Prompt builder: builds test prompts
mediq/evaluation/eval_runner.py - Evaluation runner → LLM client: generates test answers
mediq/evaluation/eval_runner.py - Evaluation runner → Guardrails: checks scope
mediq/evaluation/eval_runner.py - Evaluation runner → Evaluation scorers: scores results
mediq/evaluation/eval_runner.py - Evaluation runner → Golden dataset: reads test cases