An AI assistant that answers from your docs and cites sources

Teams keep their answers in shared drives, and searching them is slow. I built this n8n workflow template to retrieve relevant passages and attach source citations. The sections below describe its configuration; this case study does not demonstrate a live deployment or an end-to-end test.

An n8n workflow lists a Google Drive folder every night and checks each file against an index table. Only new or changed files are downloaded and split into passages. Each passage keeps its file name, link and last-modified date. Before a file counts as indexed, the workflow reads a passage back out of the vector store to prove the write landed.

A question first retrieves the closest passages. If the best match falls below a threshold, the reply says that nothing matched well enough and names what was searched. Above the threshold, the model answers from the retrieved text alone, and citations are attached before the reply is sent. If a cited file is older than the staleness window, the reply says so.

Retrieval makes answers far more reliable, but a model can still misread a passage. The citation lets anyone check the source in one click. Every question is logged with whether it was answered and which files were cited, which also shows which documents are missing.

The full workflow is at fractionalhq.uk/use-cases/document-qa-citations.

Booboo, my own open-source project, applies the same rule across a whole operation: every answer traces back to where it came from. It also emits that knowledge as plain markdown with wikilinks, which Obsidian resolves by filename, so a person can read the brain with no tooling at all. Below is the real output of booboo vault run on the fictional Pemberton Grand demo brain: the folder tree it wrote, and one page exactly as it came out.