A Google Drive folder that answers questions, with a citation on every claim

Everyone wants to “chat with our documents.” The hard part is not getting an answer, it is trusting one. Before writing any code we read 51 premium Upwork build jobs for exactly this system, and the four things clients complained about were always the same four. Each one became a feature.

• Hallucination

  A confident answer with no basis in the documents. Every claim now carries a citation, and the application checks it in code before the answer is displayed.

• No provenance

  “Your holiday allowance is 25 days,” with no source. Each citation now resolves to a chunk, a document, a heading, a page and a Drive link you can open.

• Stale versions

  Quoting the superseded 2025 policy. The classifier labels documents current or obsolete from their dates and filenames, and the agent has to say which one it used.

• Silent failure

  A scanned PDF with no text layer gets skipped, or worse, indexed as empty. It now goes to a human review queue with the reason attached.

Two of those requirements were named more often than citations were: a human approval gate, and a way to measure accuracy. So the review queue and the eval set are core to this build, not extras bolted on at the end.

Share one folder with a service account, read-only. Within an hour you can browse a map of what is inside it and ask it questions from three different places. The design rule was that every layer has one job and never does another layer’s.

• Citations checked in code: the model is asked to cite; the application decides whether the citation is real. Every marker must resolve to a retrieved chunk, and every number in the answer must appear in a cited passage or be the sum of two that do.
• One answer function: the web chat, the Slack command and the MCP server all call the same entry point, so retrieval, validation and logging can never drift between surfaces.
• Hybrid search in one SQL function: keyword ranking and vector ranking over the same candidates, fused by rank position rather than by blending incompatible scores. One index, one round trip, filters applied before ranking.
• A human gate that suspends the pipeline: an unreadable file or a low-confidence classification pauses the workflow in the database. A person approves or declines, and that decision resumes the exact run.
• Thirty golden questions in the repo: recall, faithfulness and refusal accuracy measured on every change and stored per commit, so a regression shows up as a number rather than as a complaint.
• Automation kept replaceable: the scheduler triggers and delivers. It never parses a document or writes a prompt, so all the judgment stays in version-controlled code that can be tested.