🤖 Zyro Dynamics HR Assistant

A retrieval-augmented HR policy chatbot that answers employee questions using only the company's actual HR documents — with strict grounding to prevent hallucinated policies, automatic out-of-scope detection, and full request tracing.

🔗 Live demo: hrragmasterclass.streamlit.app

---

Why this isn't a generic "PDF chatbot"

Most RAG tutorials stop at "embed documents, retrieve, answer." This one adds two things that actually matter in a real HR deployment:

1. A scope gate runs before retrieval. A lightweight classifier call checks whether a question is even HR-related before the RAG pipeline runs — so "what's our revenue last year?" or "rank up in Valorant" gets a clean refusal instead of the model improvising an answer from leaked context or general knowledge.
2. The prompt is built to refuse, not guess. If the retrieved context doesn't contain the answer, the model is instructed to say so explicitly rather than filling the gap — the single most common failure mode in HR/legal/policy bots.

How it works

graph LR
    A[User Question] --> B{Out-of-Scope<br/>Classifier}
    B -->|NO| C[Refusal Message]
    B -->|YES| D[MMR Retriever<br/>FAISS]
    D --> E[Grounded Prompt<br/>+ Context]
    E --> F[Groq LLM<br/>llama-3.1-8b-instant]
    F --> G[Answer + Sources]
    F -.trace.-> H[LangSmith]

Loading

1. Ingestion: HR policy PDFs are loaded from zyro-dynamics-hr-corpus/, split into 1000-character chunks (200-character overlap), and embedded with sentence-transformers/all-MiniLM-L6-v2.
2. Indexing: Chunks are stored in a FAISS vector index, cached once per session via @st.cache_resource.
3. Scope check: Before any retrieval happens, the question is passed to a binary YES/NO classifier prompt that decides if it's actually an HR policy question. Off-topic questions are refused immediately — no LLM call wastes a retrieval pass on them.
4. Retrieval: In-scope questions go through MMR (Maximal Marginal Relevance) search — not plain top-k similarity — pulling from a wider candidate pool (fetch_k=30) and selecting 8 chunks that balance relevance with diversity (lambda_mult=0.7), so the answer isn't built from 8 near-duplicate chunks.
5. Generation: Retrieved chunks are inserted into a strict grounding prompt that explicitly forbids outside knowledge, inference, and invented details, and mandates a fixed refusal string when the answer isn't in context.
6. Observability: Every chain run is traced to LangSmith (zyro_rag_challenge project) for debugging and inspection.

Tech stack

Layer	Tool
Orchestration	LangChain
LLM	Groq (llama-3.1-8b-instant)
Embeddings	HuggingFace all-MiniLM-L6-v2
Vector store	FAISS
Retrieval strategy	MMR (diversity-aware)
Tracing	LangSmith
UI	Streamlit

Running locally

git clone https://github.com/Shrikar-Dev/HR_RAG_MASTERCLASS.git
cd HR_RAG_MASTERCLASS
pip install -r requirements.txt

Create a .env file in the project root:

GROQ_API_KEY=your_groq_api_key_here
LANGSMITH_API_KEY=your_langsmith_api_key_here

Then run:

streamlit run app.py

The app will build the FAISS index from the PDFs in zyro-dynamics-hr-corpus/ on first load (cached afterward via Streamlit's resource cache).

Example questions

• What is the work from home policy?
• How does employee onboarding work?
• What expenses can be reimbursed?
• What is the leave approval process?
• What are the remote work security requirements?

Try an off-topic question (e.g. "what's the best Valorant strategy?") to see the scope-gate refusal in action.

Known limitations

• The out-of-scope classifier and the answer-generation step are two separate LLM calls — this adds latency. A future version could combine them or use a smaller/faster model purely for classification.
• No conversation memory across turns yet — each question is answered independently of chat history.
• Source attribution currently shows file names only, not page numbers.

Project structure

HR_RAG_MASTERCLASS/
├── zyro-dynamics-hr-corpus/   # HR policy PDFs (source documents)
├── app.py                     # Streamlit app + RAG pipeline
├── Starter_Notebook.ipynb     # Original development/exploration notebook
├── requirements.txt
└── .gitignore

License

MIT