AI Investigation & Knowledge Platform
RAG platform that turns a power plant's manuals and engineering drawings into cited investigation reports and instant answers.
Overview
A production AI platform for the control room and engineering teams of a pumped-storage hydroelectric plant. Operators describe a trip, alarm or site observation and the platform drafts a structured investigation report grounded in the plant's own documentation. A second module, a conversational technical assistant, answers follow-up questions with sources.
The Challenge
Investigations meant searching thousands of pages of manuals, procedures and engineering drawings by hand. The answer often depended on which engineer remembered where something was written, and diagrams were effectively unsearchable.
The Solution
Investigation report generator
Turns an event description into a structured report with findings, likely causes and references back to the source documents.
Drawing intelligence
Engineering drawings are read with vision AI so the content of diagrams becomes searchable, not just the text around them.
Conversational technical assistant
Multi-turn Q&A with follow-up memory, answers grounded strictly in the plant's knowledge base, multi-document synthesis with citations.
Document pipeline
Upload, parsing, chunking, embedding and indexing for PDFs and drawings.
How It Works
Docs & drawings uploaded
Parsing + vision analysis
Voyage AI embeddings
Qdrant vector index
Hybrid retrieval + ranking
Claude report / answer with sources
Architecture
Ingestion API
FastAPI endpoints for upload, parsing and chunking, with safe handling of concurrent uploads.
Vision layer
Claude vision describes engineering drawings so they can be embedded and retrieved.
Vector index
Voyage AI embeddings stored in Qdrant.
Retrieval engine
Vector search plus keyword boosts, phrase rescue and breadth-penalty damping.
Report generator
Claude produces structured reports with citations to the exact source documents.
Web app
React, Vite and Tailwind interface for reports, chat and document management.
Engineering Highlights
- Retrieval tuned against real QA batches submitted by plant engineers, so accuracy is measured on real questions, not demos.
- Keyword-boost disambiguation and breadth-penalty damping stop large equipment catalogues from drowning out the right document.
- Embeddings migrated from local sentence-transformers to Voyage AI for a clear accuracy gain.
- Fixed a production race condition where concurrent uploads could corrupt files.
Outcome
- Live in production for control room operators and engineers.
- Delivered across multiple paid phases, each with a 5-star review.
- Ongoing engagement focused on retrieval accuracy and new features.