Introduction

In the current landscape of Large Language Models (LLMs), retrieving accurate information from domain-specific documents is a major challenge. Standard Retrieval-Augmented Generation (RAG) systems often fail when dealing with complex terminology and structured files across highly diverse sectors like Healthcare, Law, and Agriculture.

To solve this, I designed TRISEVA (a Multi-Agent Retrieval-Augmented System) as part of my M.Tech dissertation at BITS Pilani.

Architecture

TRISEVA employs a modular multi-agent architecture where specialized agents collaborate to analyze, retrieve, and formulate explainable answers:

  1. Domain Classifier Agent: Routes user queries to the specific domain context (Healthcare, Legal/Government, or Agriculture).
  2. Retrieval Specialist Agent: Queries local vector stores optimized for specific terminologies and document hierarchies.
  3. Explainer Agent: Interprets the retrieved context and structures the output to show source citations and confidence scores, providing transparency.

Key Technologies Used

  • LangChain & LangGraph: For defining stateful, collaborative agent graphs.
  • Python: Core programming language.
  • TensorFlow / PyTorch: For embedding tuning and text processing.
  • ChromaDB / PGVector: Used to cache and query semantic embeddings.

Conclusion

By breaking down the retrieval and reasoning tasks into specialized agents, TRISEVA delivers explainable, highly accurate domain answers, reducing hallucinations significantly and making QA reliable for critical fields.