AI Empowers · Energy
Private ChatGPT on company documents
Employees ask questions in plain language and get answers from internal documents, with sources.
- Energy
- Private chatbot
- Document intelligence
The problem
Staff needed answers from a large set of internal documents, without leaking data and without a model making things up.
The solution
Retrieval-augmented generation instead of fine-tuning: documents chunked with overlap and access metadata, hybrid search, and answers that cite their sources or say there is no answer.
The outcome
A private assistant that respects document permissions. We evaluated Redis, FAISS, Databricks, Weaviate and Pinecone along the way.
The full story
Why not fine-tune
Fine-tuning a model on private data makes answers hard to trace, ignores who may see which document, needs retraining for every update and still has a knowledge cutoff. So we kept the knowledge and the language model apart.
How it works
Documents are split into overlapping chunks with section context and access metadata. A hybrid search, keywords plus vectors, finds the relevant passages, and only those go to the model, with one instruction: answer from them and cite them, or say there is no answer.
At scale
The first version indexed about 10,000 technical documents and was tested on real questions in French. It is now used in about nine countries and eighteen power plants. It has grown into an agent built on the Microsoft Agentic Framework that reads real-time plant parameters from the data warehouse, compares them with historical cases and sits inside the application people already use on the plant floor. It supports operators, who stay responsible for complex or critical decisions.
Tech
Ask Pai
Ask Pai about this case
Pai is an AI. It only knows what is written on this page.
Pai is an AI. Answers come from this case page. Don't share sensitive data.
Chats are kept for 30 days unless you switch that off in Privacy and choices.