README.md

HR Policy Assistant (RAG)

A retrieval-augmented generation (RAG) chatbot that answers employee questions about a company's HR policy document. Built with LangChain, Groq-hosted LLMs, Jina embeddings, and a Qdrant Cloud vector store, with input/output safety guardrails and LangSmith tracing.

How it works

1. Ingest — hr_assistant/document_loader.py loads data/hr_policy.txt, and hr_assistant/splitter.py splits it into chunks (CHUNK_SIZE=500, CHUNK_OVERLAP=60).
2. Embed & store — hr_assistant/embeddings.py creates embeddings with Jina (jina-embeddings-v2-base-en), and hr_assistant/vector_store.py uploads/loads them from a Qdrant Cloud collection (reused on subsequent runs instead of re-embedding).
3. Retrieve — hr_assistant/tools.py wraps the vector store retriever (top-k = 3) as a search_hr_policy tool.
4. Agent — hr_assistant/agent.py builds a LangChain agent (openai/gpt-oss-20b via Groq) that calls the search tool to ground its answers.
5. Guardrails — hr_assistant/guardrails.py runs a separate Groq safety model (openai/gpt-oss-safeguard-20b) to screen both the incoming question (prompt injection, requests for other employees' data) and the outgoing answer (PII leaks, unauthorized promises, suspicious links) before it reaches the user.
6. Everything is wired together in hr_assistant/pipeline.py (build_hr_assistant() / ask()), used by both entry points below.

Entry points

• python main.py — CLI demo that asks a few sample HR questions.
• streamlit run app.py — interactive chat UI.
• rag.ipynb — notebook version for experimentation.

Project layout

hr_assistant/
  config.py          settings, env vars, system prompt
  document_loader.py load the HR policy text file
  splitter.py         chunk the document
  embeddings.py       Jina embeddings model
  vector_store.py     Qdrant Cloud build/load/retriever
  tools.py             search tool for the agent
  llm.py                Groq LLM setup
  agent.py             LangChain agent construction
  guardrails.py        input/output safety checks
  pipeline.py           wires everything together (build_hr_assistant, ask)
  logger.py             file logging (logs/)
  tracing.py             LangSmith tracing check
data/hr_policy.txt      source HR policy document
docs/                    notes on logging, LangSmith, Qdrant Cloud migration, guardrail attack testing
NOTES/                   reference PDFs

Setup

1. Install uv:

   pip install uv
   

2. Create and activate a virtual environment:

   uv venv ragenv
   ragenv\Scripts\activate
   

3. Install dependencies:

   uv pip install -r requirements.txt
   

4. Create a .env file with:

   GROQ_API_KEY=...
   JINA_API_KEY=...
   QDRANT_URL=...
   QDRANT_API_KEY=...
   QDRANT_COLLECTION_NAME=hr_policy
   LANGSMITH_TRACING=false
   LANGSMITH_ENDPOINT=...
   LANGSMITH_API_KEY=...
   LANGSMITH_PROJECT=...
   

Git basics

git add .
git commit -m "Some message"
git push