Resume

Adaptive RAG Chatbot

An intelligent retrieval system that dynamically adapts query processing to reduce hallucination and improve response quality in LLMs.

šŸš€ Live DemoGitHub ā†—
Adaptive RAG Chatbot

Problem

Large Language Models often generate confident but incorrect responses when relevant context is missing. Traditional RAG systems rely on static retrieval pipelines, which treat all queries equally ā€” leading to unnecessary latency for simple queries and insufficient context for complex ones.

Solution

Designed a query routing mechanism using LangGraph to classify queries and dynamically select retrieval strategies
Implemented hybrid retrieval combining vector search (Qdrant) with contextual filtering for improved relevance
Added a re-ranking layer to prioritize high-signal documents and remove noisy results
Integrated fallback mechanisms (FAISS) and memory (MongoDB) for fault-tolerant conversational continuity

Tech Stack

LangGraphQdrantFAISSMongoDBFastAPIGPT-4o

Architecture

Query ā†’ Intent Classification ā†’ Adaptive Routing ā†’ Hybrid Retrieval (Qdrant/FAISS + Filters) ā†’ Re-ranking ā†’ Context Construction ā†’ LLM ā†’ Response

Challenges

Balancing latency and accuracy was critical ā€” deeper retrieval improves answer quality but increases response time. This was solved using adaptive routing to trigger complex pipelines only when needed. Ensuring factual grounding required strict context injection and prompt design to minimize hallucination.

What Iā€™d Improve Next

  • ā€¢ Introduce learning-to-rank models for advanced re-ranking
  • ā€¢ Add feedback-driven reinforcement loop for retrieval optimization
  • ā€¢ Optimize ANN indexing (HNSW tuning) for large-scale datasets