Case Study

CV RAG Chatbot

A live Retrieval-Augmented Generation chatbot — ask questions about my professional background, powered by LangChain + FAISS.

LangChain LCEL FAISS FastEmbed (ONNX) Groq Llama 3.3 Streamlit RAGAS 2024
🤖 Try the Live Chatbot →

The Problem

My GitHub profile doesn't tell the full story. Most of my best work is under NDA, and a static CV can't answer follow-up questions. I needed a way for recruiters and potential clients to interactively explore my background — ask specific questions and get context-aware answers.

This is also a live demonstration of RAG architecture — showing that I can build, evaluate, and deploy LLM applications, not just talk about them.

The Architecture

graph TB A[User Question] --> B[Streamlit UI] B --> C[LangChain LCEL Orchestrator] C --> D[FAISS Vector Store] C --> E[Groq Llama 3.3] D --> F[CV + Project Docs] E --> G[Generated Answer] G --> B H[RAGAS Evaluation] --> I[Faithfulness] H --> J[Answer Relevancy] H --> K[Context Precision] style C fill:#6c5ce7,stroke:#7c6df0,color:#fff style D fill:#00d2ff,stroke:#00b8e6,color:#0a0a0f style H fill:#f7c948,stroke:#e0b830,color:#0a0a0f

Why It's Hard

  • Retrieval quality is everything. If the retriever pulls irrelevant chunks, the LLM hallucinates. I used RAGAS to measure faithfulness, answer relevancy, and context precision — iterating on chunking strategy and embedding models until metrics were solid.
  • Chunking strategy matters. CV data is structured (sections, bullet points). Naive chunking breaks semantic coherence. I implemented section-aware chunking that respects document structure.
  • Evaluation-driven development. Instead of guessing whether the chatbot was good, I built a RAGAS evaluation pipeline with synthetic test questions. Every change to chunking, prompting, or retrieval was measured.
  • Deployment on Streamlit Cloud. Free tier, no server management. But cold starts and memory limits required optimizing the FAISS index size and model loading.

Technical Stack

  • LangChain LCEL — declarative chain composition: document loading, splitting, retrieval, and prompt management
  • FAISS — vector store for semantic search over CV and project documents
  • FastEmbed (ONNX-based) — local embeddings, no API dependency, fast cold starts
  • Groq + Llama 3.3 — LLM for answer generation via Groq's high-throughput inference API
  • Streamlit — UI framework, deployed on Streamlit Community Cloud
  • RAGAS — evaluation framework measuring faithfulness, answer relevancy, and context precision

What I'd Do Differently

  • Add hybrid search (BM25 + semantic). Pure semantic search sometimes misses exact keyword matches. A hybrid approach would improve recall for specific skill/technology queries.
  • Implement query rewriting. User questions are often vague. Rewriting them into more specific queries before retrieval would improve context precision.
  • Add source citations in the UI. Showing which CV section each answer came from builds trust. Users should be able to click through to the source.
  • Experiment with reranking. A cross-encoder reranker after initial retrieval could further improve context precision, especially for ambiguous queries.

Key Takeaways

Building a RAG chatbot is easy. Building one that's actually useful is hard. The difference is evaluation — measuring retrieval quality, iterating on chunking, and being honest about when the system doesn't know something. RAGAS made this measurable instead of subjective.
← Back to Projects