2025Deployed; multi-layered retrieval solved garbled-PDF relevance
ExamAI: RAG Study Assistant
Upload lecture notes (even the garbled ones) and get answers, follow-up questions and MCQ quizzes, plus the exact retrieved context behind every answer.
- LangChain
- Gemini
- Groq Llama3
- Pinecone
- BM25
- Streamlit
Built during the NSK.AI Bootcamp after my course notes kept defeating naive retrieval. Academic PDFs extract badly; simple vector search returned junk. The fix was a multi-layered retrieval strategy:
- HyDE query enrichment using Groq’s Llama3 before embedding
- Hybrid search combining dense vectors with BM25 keyword matching
- Cross-encoder reranking of the final candidate set
Stack: Google Gemini 2.5 Flash, sentence-transformers/all-MiniLM-L6-v2, Pinecone, LangChain, Streamlit.
The breakthrough took days of debugging, and it turned exam prep from panic into process.
Notable outcome
Deployed; multi-layered retrieval solved garbled-PDF relevance