Retrieve before writing
Retrieval-Augmented Generation means first retrieving relevant information, then supplying the model with the question and those excerpts. It is not retraining the model on all company data.
Embeddings represent content as numerical vectors to help find semantically related text. Search can combine these with traditional keywords. A good match still needs context: document version, date and whether it applies to the task.
Sources make answers checkable
AI Link brings company knowledge and connected sources directly into chat. Citations lead back to the basis of a statement. Synchronisation and indexing prepare information for retrieval and may store document content or derived representations in the workspace.
Retrieval must return only authorised content. Deleted documents and revoked access need to be reflected in the data lifecycle. Useful tests therefore include revoked permissions and outdated sources.
Three questions for a pilot
Familiar business questions provide a reliable basis for evaluation. Retrieved sources and the generated text are assessed separately, showing where retrieval and responses already work well and where improvement is needed.
- Was the right document version retrieved?
- Does the source support every important claim?
- Does the system say when an answer is missing?
Sources & further reading
Original research, provider documentation and further reading on the concepts and figures discussed here.