Company Knowledge Assistant That Cites Its Sources

A confidently wrong answer about your own refund policy is more expensive than no answer at all.

The same questions get asked every week, and the answers live in SOPs, a shared drive, old tickets and one person’s head. A generic chatbot bolted onto that makes it worse.

Every node with its icon, as on the n8n canvasstarts a runa model or a parser, wired into the node above itHover a node for its name and type

• 01Change Detection & Re-Index

  13 nodes · 13 connections · 4 notes

  every six hours, only what changed

• 02Gap-Log Digest

  9 nodes · 8 connections · 3 notes

  the weekly list of what it could not answer

• 03Ingest & Index

  25 nodes · 27 connections · 5 notes

  Drive, PDFs and tickets: chunk, embed, store

• 04Slack Answer Path

  22 nodes · 22 connections · 5 notes

  a question in, a cited answer or a refusal out

• 05Web Widget Answer Path

  26 nodes · 28 connections · 5 notes

  the same, for a site widget

The gates, 10 If and Switch nodes, decide whether anything happens next: a countable part of the system, not a claim about it. The 22 sticky notes carry the reasoning on the canvas, at the point it applies.

HoverTap a card to read what it prevents.

• No verbatim quote, no answer

  No verbatim quote, no answer

  Case, whitespace and curly quotes are normalised, then the quote must appear in what was retrieved. A failed check routes to a refusal, never to a softened answer.

• Default-deny, inside the search

  Default-deny, inside the search

  Permissions are applied inside the similarity search, not checked on the answer afterwards, so a restricted document cannot leak through a prompt.

• Unchanged documents are never re-embedded

  Unchanged documents are never re-embedded

  A content hash is compared first, so re-reading a folder that has not changed costs nothing.

• Old text goes before new text arrives

  Old text goes before new text arrives

  A changed document’s old chunks are deleted before its new ones are written, or it would answer with both versions and nobody could tell which.

• Deleted at the source, deleted here

  Deleted at the source, deleted here

  A file trashed in Drive loses its record and chunks, so nothing cites a file that no longer exists.

• The assistant ignores itself

  The assistant ignores itself

  Messages from bots are dropped first — the classic loop of a Slack bot replying to its own answer, forever.

• Built, not yet deployed. The workflows pass the importability checks; they have not run end to end against live Supabase, OpenAI and Slack accounts.
• WhatsApp is not built. Slack and a site-widget webhook are; WhatsApp would be a new trigger into the same retrieval and quote check.
• Tickets are read from a table a separate export has to fill; there is no Zendesk, Freshdesk or Intercom connector here.
• The model’s confidence is logged, not gated: a low-confidence answer that passes the quote check still posts.