A production-style conversational RAG agent — grounded answers, verifiable citations, and multi-turn memory, built on LangChain LCEL and served through Streamlit.

Live demo: https://support-rag-agent.streamlit.app/

Most "RAG demo" repos hallucinate sources and forget what the user just said. This one doesn't:

• Every claim is traceable. Answers cite the exact FAQ section and line number they came from — not a vague "according to our docs."
• It actually remembers. Ask "Do you ship to India?" then "How much does that cost?" — the second question resolves correctly using conversation history, not a fresh, context-free query.
• It fails safely. No API key configured, missing data, or an unanswerable question — the agent tells you plainly instead of guessing.

User question
     │
     ▼
History-aware retriever   →  rewrites "how much does that cost?" into a standalone
                              query using prior turns, then retrieves matching FAQ
                              sections from FAISS
     │
     ▼
Cited answer chain         →  LLM answers strictly from retrieved context, required
                              to cite (Source, Section, Line) for every claim
     │
     ▼
Per-session memory          →  both the rewrite step and the answer step see the
                               full chat history, keyed by session_id

The FAQ is split by ## Section headers. Each chunk retains source, section, and line_start metadata, which is rendered back into the LLM's context and enforced in the system prompt — the model is instructed to cite only sections that actually appear in front of it, never to invent one:

Shipping to India costs $24.99 (Source: gigacorp_faq.txt, Section: "Shipping Policy", line ~1).

Prerequisites: Python 3.12, a free Groq API key (or an OpenAI/Anthropic key if you prefer).

1. Clone the repository

git clone https://github.com/kailashv2/support-rag-agent.git
cd support-rag-agent

2. Create and activate a virtual environment

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\Activate.ps1

3. Install dependencies

All dependencies are listed in requirements.txt, pinned to exact versions for reproducibility:

pip install -r requirements.txt

4. Configure your API key

Copy the example env file and add your key:

cp .env.example .env

Open .env and set:

GROQ_API_KEY=your-key-here

Alternatively, skip this and paste the key directly into the app's sidebar at runtime.

5. Run the app

streamlit run app.py

Opens at http://localhost:8501.

6. Run the tests (optional but recommended)

pytest tests/ -v

Covers section parsing, metadata integrity, and failure handling on malformed input — no API key or network call required.

1. Push this repo to GitHub.
2. share.streamlit.io → New app → select this repo, branch main, file app.py.
3. Under Advanced settings, set the Python version to 3.12 and add a secret:

   GROQ_API_KEY = "your-key-here"

4. Deploy — first load downloads the embedding model (~90MB), expect a 30–60s cold start.

1. New Space → SDK: Streamlit.
2. Push this repo's contents to the Space.
3. Settings → Repository secrets → add GROQ_API_KEY.

• PDF instead of .txt: swap parse_sections() in src/knowledge_base.py for PyPDFLoader + RecursiveCharacterTextSplitter, keep page number in metadata.
• Scale beyond one instance: replace the in-memory _sessions dict in SupportAgent with a Redis-backed ChatMessageHistory.
• New LLM provider: add a branch to _load_llm() in src/chat_agent.py and an entry to PROVIDER_KEY_NAMES in app.py.

MIT — use freely, attribution appreciated.