The problem
Documents were classified and filed automatically, but finding information inside them still meant opening files one by one — metadata and tags help you find a document, not an answer buried on page 12 of it.
The approach
Kept the existing AI Builder + Power Automate pipeline for classification and routing, and added a RAG (Retrieval-Augmented Generation) layer on top so the system can answer questions grounded in the actual document content — not classify-and-file, but classify-and-understand.
What it does
- Classifies and routes incoming documents automatically via AI Builder, as before
- Splits document text into chunks and embeds them into a vector index, so content becomes searchable by meaning, not just keyword
- On a user question, retrieves the most relevant chunks and passes them to an LLM as context — the answer is grounded in retrieved text, not the model's memory
- Surfaces the source document alongside every answer, so users can verify the claim rather than trust it blindly
- Copilot integration lets this run as a conversational interface inside the tools people already use, instead of a separate search portal
Why RAG here
Fine-tuning a model on the document library would go stale the moment new documents arrive, and re-training isn't a realistic response to a new file being uploaded. RAG keeps generation and retrieval separate — new documents just get embedded and added to the index, so the system's knowledge updates the moment a file lands, with no retraining step.
Outcome
Turned a filing system into a queryable knowledge base. Instead of hunting through folders, users ask a question and get an answer with the source document attached — while the underlying classification and compliance routing kept working exactly as it did before.