Project 02 / Hackathon
LogBase
Massive machine logs are hard to question and easy to misread. LogBase turns the BGL and OpenStack corpora into evidence-grounded answers: an AI agent calls read-only MCP tools that search and aggregate inside the database and return only bounded evidence or exact counts, never the raw corpus.
What we built
- Eight read-only MCP tools served over stdio, usable from any MCP client
- Full-text search over an inverted index with phrase, AND, and exclude syntax
- Exact counting and trends — per-group totals, time buckets, and operational hotspots
- Distinctive-term analysis computed in-database with native term-dictionary aggregates
- Role-aware answers for engineers, team leads, and managers over the same corpus
- Automatic charting of numerical answers and opt-in cited investigation recaps
Built with
- Python 3.12 and the Model Context Protocol (MCP)
- SereneDB for full-text search and native aggregate analytics
- Docker Compose for a reproducible local database
- The Loghub BGL and OpenStack log datasets (~4.75M lines)
- A team of three, built during a hackathon
A hackathon project built with two other students. Every result carries the executed SQL, elapsed time, and a truncated flag, so a limited answer is never mistaken for the whole picture, and observed evidence stays separate from recommendations. All searching and counting runs inside the database — the agent never pulls the raw corpus into the model.