IRS RAG Chatbot

A Retrieval-Augmented Generation app that scrapes the entire IRS Internal Revenue Manual (~1,287 pages), embeds it into a vector database, and lets you chat with the corpus through a Streamlit interface — with inline citations and per-source summaries.

---

What it does

• Bulk-scrapes the IRS Internal Revenue Manual end-to-end — walks the table of contents, follows every partN/irm_* link, and ingests them in one click.
• Drop-in document chat — PDF uploads and arbitrary URLs work the same way as IRS pages.
• Cites its sources, then explains them — answers come with bracketed citations ([1], [2]…) and an optional second pass that summarizes each cited chunk in plain English.
• Refuses to hallucinate — if the retrieved context doesn't contain the answer, the prompt forces a literal "I don't have that information" with no citations.

Why I built it

I wanted a project that exercised the full RAG stack without leaning on a hosted "chat with PDF" service. The IRS Manual was a deliberate pick: it's massive, hierarchical, full of cross-references, and the kind of document where a wrong answer with a confident citation is actively harmful — so the citation discipline and the "refuse if absent" guardrail were the parts I cared most about getting right.

Bulk IRS ingestion

One click walks the IRS Internal Revenue Manual table of contents, follows every partN/irm_* link, and streams each page through the ingestion path below.

How it works

Retrieval path: user query → all-mpnet-base-v2 embedding → ChromaDB top-k similarity search → cited context injected into a Gemini 2.0 Flash prompt → response parsed for [n] citations → cited chunks re-summarized in a second LLM call (toggleable).

Ingestion path: PDF (PyMuPDF, page-level with markdown tables) or URL (WebBaseLoader) → regex cleanup → RecursiveCharacterTextSplitter (1000 / 200 overlap) → embedded and persisted in ChromaDB.

Tech stack

Layer	Choice	Why
LLM	Gemini 2.0 Flash	Fast, cheap, generous free tier — good fit for a portfolio demo
Embeddings	sentence-transformers/all-mpnet-base-v2	Strong general-purpose model, runs locally, no API cost
Vector store	ChromaDB (persistent)	Zero-config local persistence; survives restarts
Orchestration	LangChain	Loaders, splitters, Chroma wrapper
PDF parsing	PyMuPDF	Preserves page boundaries + extracts tables as markdown
Web scraping	WebBaseLoader + requests + regex	Bulk-walks the IRS sitemap
UI	Streamlit	Minimal-effort UI for a single-user RAG demo
Tracing	LangSmith	Optional, for inspecting retrieval / prompt behavior

Run it locally

git clone https://github.com/T-Rzeznik/irs-rag-chatbot.git
cd irs-rag-chatbot
pip install -r requirements.txt
streamlit run app.py

Then either:

• Paste your API keys into the API Keys expander in the UI, or
• Copy .env.example to .env, fill it in, and they'll be picked up.

Keys you'll need: HuggingFace, Google AI Studio, and (optional) LangSmith.

Project layout

.
├── app.py                  # Streamlit UI + chat + retrieval orchestration
├── document_processor.py   # PDF / URL ingestion + IRS bulk scraper
├── requirements.txt
├── .env.example
└── docs/                   # Architecture diagram + screenshots

License

MIT — feel free to fork it, learn from it, or rip it apart.