Causal Analysis RAG System
Inter IIT Tech Meet 14.0
A retrieval system that knows when it doesn't know enough yet — and keeps digging until it does.
Overview
For Inter IIT Tech Meet 14.0, our team built a multi-agent Retrieval-Augmented Generation system designed for causal, analytical questions — the kind where a single retrieval pass usually isn't enough. The system shipped as 5 independent FastAPI microservices, containerized and orchestrated with Docker Compose for horizontal scaling. It won a Silver Medal.
The Problem
Standard RAG retrieves once and answers immediately. For causal or multi-hop analytical questions — "why did X happen, and what else does it affect?" — a single pass often surfaces only part of the evidence chain, producing shallow or confidently wrong answers. We needed retrieval that could recognize its own gaps and keep searching.
Approach
3-hop iterative retrieval with sufficiency checks
Architected a 3-hop LangGraph retriever where an LLM judges whether the retrieved evidence is sufficient after each hop. If not, it refines the query and retrieves again — up to 3 hops — closing gaps that single-pass RAG would miss entirely.
Weighted ensemble reranking
Combined ColBERT, Cohere, and Voyage rerankers into a single weighted ensemble, surfacing the top-10 most relevant evidence chunks from a much larger candidate pool — balancing each model's strengths instead of betting on one.
Long-term memory with GraphRAG
Integrated Zep Cloud's GraphRAG memory layer so follow-up questions in the same session are enriched with prior context automatically, skipping redundant retrieval calls and keeping multi-turn conversations coherent.
Privacy-safe by default
Built an automated Presidio-based PII redaction pipeline covering 11 entity types, with exponential-backoff retries and a TF-IDF fallback reranker for when primary rerankers are unavailable — keeping the system usable and safe under load.
Independently scalable microservices
Split the system into 5 FastAPI microservices (ingestion, retrieval, reranking, memory, generation) deployed via Docker Compose, so each stage can scale horizontally on its own.
Results
- Won a Silver Medal at Inter IIT Tech Meet 14.0, competing against teams from every IIT
- 3-hop retrieval loop materially improved evidence completeness on causal, multi-hop queries
- 11-entity-type PII redaction pipeline with TF-IDF fallback kept the system privacy-safe even under reranker failure
- 5-microservice architecture validated horizontal scaling under Docker Compose
Stack
More projects