Clearpath Chatbot Support System

This project is a functional, production-ready customer support chatbot for "Clearpath", a fictional project management SaaS. It implements an advanced RAG (Retrieval-Augmented Generation) pipeline, a deterministic model router, and an output evaluator to ensure safe and accurate responses from the LLM.

🌐 Live Deployment

The project is currently deployed and accessible at: http://13.49.75.239

---

How to Run the Project Locally

Prerequisites:

• Python 3.9+
• A Groq API Key

1. Setup the Environment Clone this repository and navigate to the project root:

# Create and activate a Virtual Environment
python3 -m venv venv

# On Windows:
.\venv\Scripts\Activate.ps1
# On Mac/Linux:
source venv/bin/activate

# Install the dependencies
pip install fastapi uvicorn groq chromadb sentence-transformers pydantic pypdf2 python-dotenv python-multipart

2. Set up the Environment Variables Create a .env file in the root directory or export the variable directly:

# On Windows (PowerShell):
$env:GROQ_API_KEY="your-api-key"

# On Mac/Linux:
export GROQ_API_KEY="your-api-key"

3. Start the Backend Server The backend main.py is configured to handle the /query endpoint and serve the frontend statically. Run the startup command from the root directory:

python -m uvicorn backend.main:app --reload

The server will now be running at http://localhost:8000.

4. Access the Chat Interface Open your web browser and go to http://localhost:8000. The frontend interface is served directly from the backend. Type a message and hit "Send"!

---

Groq Models & Environment Config

This project utilizes two models via the Groq API. Queries are routed dynamically by the rule-based logic in main.py to balance cost, speed, and reasoning capabilities:

1. Simple Queries: llama-3.1-8b-instant
   • Used for greetings, short questions, and single-fact lookups. Also used for the Query Condensation step.
2. Complex Queries: llama-3.3-70b-versatile
   • Used for queries containing interrogative words (why/how/explain), multi-step questions, or queries exceeding 10 words.

Environment Config: The application simply requires the GROQ_API_KEY to be set in the environment.

---

Bonus Challenges Attempted

According to the optional Bonus Challenges rubric, the following three challenges were successfully implemented:

1. Conversation memory: The chatbot maintains conversation memory across turns. This is implemented via an active sliding-window memory (last 4 messages) managed by a backend session dictionary (CONVERSATIONS). A Query Condensation Layer using the 8B model was also implemented to rewrite pronoun-heavy follow-up questions before hitting the vector database.
2. Eval harness: A dedicated eval_harness.py test suite was written. It contains a dozen test queries covering factual retrieval, out-of-bounds questions, and competitor mentions. Running python backend/eval_harness.py automatically evaluates the system against these expected answers and reports Pass/Fail for each.
3. Live deploy: The full backend and frontend stack was deployed to a business-grade cloud provider (AWS EC2) and is accessible via a public URL interface.

---

Additional Self-Improved Features

Beyond the assignment rubric, several advanced RAG concepts were engineered to ensure production-grade reliability:

1. Advanced Two-Stage Retrieval (Reranking): Upgraded from standard dense retrieval to a two-stage pipeline. ChromaDB first fetches the top 15 chunks using all-MiniLM-L6-v2, and then a powerful Neural Cross-Encoder model (ms-marco-MiniLM-L-6-v2) reranks them mathematically, returning only the top 3 most precisely relevant chunks. I also tuned the negative logits threshold to allow for typo tolerance.
2. Lexical Grounding Flagging: Built a custom algorithmic hallucination checker. Before returning the final answer, the system extracts all long nouns from the LLM's response and verifies that at least 50% (for simple models) to 60% (for complex models) of them actually exist in the retrieved PDF chunks. If an LLM hallucinates new features not found in the manuals, it perfectly flags the output with low_grounding.
3. Query Condensation Layer: Built an LLM-powered query rewrite step using the fast 8B model. When a user asks a follow-up question containing pronouns or implicit context (e.g., "how much does it cost?"), the condensation layer rewrites the query using the entire conversation history into a standalone, search-optimized string (e.g., "what is the price of the enterprise plan?"). This dramatically improves vector semantic similarity matching for multi-turn conversations.

---

Known Issues & Limitations

1. In-Memory State Management: The conversation history is currently stored in a Python dictionary (CONVERSATIONS) inside the server memory. While fast, this means that if the server crashes or restarts, all active user conversational histories are lost. In a real production system, this would be moved to Redis or a database.
2. No Streaming: Output streaming was purposely disabled in favor of strict JSON responses (/query) to ensure the frontend could confidently parse the structured evaluator warnings (metadata.flags) without complex stream chunk parsing. This increases perceived latency for the user.
3. Cross-Encoder Latency "Cold Start": Because the Cross-Encoder model runs locally on CPU (especially noticeable on the EC2 instance without a GPU), the very first query takes slightly longer as PyTorch loads the model into active memory.