ClearPath Chatbot - Backend API

A RAG-powered customer support chatbot built with FastAPI for the ClearPath PM SaaS platform. The backend has three layers:

• BM25 Retrieval Pipeline -- Extracts and chunks 30 PDF documents, builds a BM25Okapi index, and retrieves the top 5 relevant chunks per query.
• Deterministic Router -- A rule-based classifier that routes simple queries to a smaller model and complex queries to a larger model based on keyword matching, word count, and sentence structure.
• Output Evaluator -- Inspects every LLM response and flags issues like missing context, refusals, or internal data leaks.

---

Local Setup

1. Clone the repository

git clone $repo_link
cd backend

2. Create and activate a virtual environment

python -m venv .venv

Windows:

.venv\Scripts\activate

macOS / Linux:

source .venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Set up environment variables

Create a .env file in the backend directory:

Groq_Llama_33_70b_versatile=your_api_key_here
Groq_Llama_31_8b_instant=your_api_key_here

5. Build the BM25 index

python -m rag.ingest

This reads all PDFs from rag/ClearPath/clearpath_docs/, chunks them, and saves the BM25 index to rag/index_store/.

6. Start the server

uvicorn main:app --reload --port 8000

---

Models Used

Two Groq-hosted LLMs are used, selected at runtime by the deterministic router:

• llama-3.1-8b-instant -- For simple queries (short, direct questions).
• llama-3.3-70b-versatile -- For complex queries (comparisons, multi-part questions, analytical requests).

---

Bonus Challenges Attempted

Conversation Memory

Conversation history is maintained in-memory using a conversation_id. Each new session gets a unique ID, and follow-up messages with the same ID continue the conversation. The backend passes the last 3 exchanges (6 messages) to the LLM for context.

Live Deploy

The frontend is deployed on Vercel and is live at: https://clear-path-frontend.vercel.app/

The backend is configured for Vercel deployment using vercel.json.

---

Known Issues and Limitations

• Aggressive Routing -- The deterministic router is strict to prioritize answer quality. Keywords like "how do", "why", or "explain" immediately trigger the complex classification, so even basic questions like "how do I reset my password" get routed to the 70B model.

• Ephemeral Memory -- The conversation store is a Python dictionary (_conversations) held in memory. All conversation history is lost when the server restarts.