Retrieval · LTIMindtree
Root-cause analysis over machine logs with hybrid-retrieval RAG
A routed, multi-agent RAG system for enterprise logs: 89% relevance on technical queries over 500GB+ a day. The prototype placed 1st runner-up at LTIMindtree Hack The Future 2.0.
The problem
When something breaks in a large enterprise system, the answer is usually in the logs — somewhere in hundreds of gigabytes of them. Engineers spent hours grepping, correlating timestamps and guessing. We wanted to ask the logs a question and get a grounded answer.
Approach
- Hybrid retrieval because logs are full of exact tokens — error codes, hostnames, stack frames — that pure embeddings blur.
- A custom router across 7 specialised prompts, so a latency question and a crash question get different treatment.
- Separate agents for retrieval, reasoning and output generation, which made each step testable on its own.
Stack: LangChain, LlamaIndex, Pinecone, Weaviate, FastAPI.
Result
89% relevance on technical queries, at 500GB+ of logs per day and 30M+ embeddings, cutting complex query resolution from hours to minutes. The first version was built for LTIMindtree's Hack The Future 2.0 with team Innov8ors, where it placed 1st runner-up (₹75,000).