Answers grounded inyour own documents.

Fine-tuning changes the model's weights; RAG changes what the model sees at the moment it answers, by retrieving relevant passages first. RAG is generally faster to update — a new document is available as soon as it's indexed — and keeps answers traceable to a source, which fine-tuning does not.

Most text-bearing sources: PDFs, wikis, tickets, contracts, structured records and exports from systems you already run. Each source needs its own ingestion and chunking approach to retrieve well, which is most of the actual engineering.

By enforcing the same access controls at retrieval time that the source systems already use, rather than relying on the model to withhold information it was shown. A document a user can't open in the source system isn't retrievable through the assistant either.

Separately from the model's writing quality. Retrieval evaluation checks whether the right passage was surfaced at all — precision and recall against a labelled set of real queries — before ever judging the answer generated from it.

Not necessarily. The right choice depends on data volume, existing infrastructure and update frequency — sometimes a managed vector store is right, sometimes hybrid search over infrastructure you already run is a better fit.