Confluence

Half the answers to “why is it like this” questions are already written down — in Confluence. Requirements, architecture decisions, procedures, runbooks: things a person will not retell in a question and the model does not know on its own. The integration lets Buff read the spaces in the connection’s scope when an answer needs them, and takes questions right in the page discussion. One integration works with both self-hosted (Data Center, Server) and Cloud.

• A support agent answers by the procedure, not from memory

  A customer asks for a refund twenty days after paying. The refund procedure is written in the support procedures in Confluence, but the agent remembers it only roughly, and asking the manager means waiting until the evening.

  Result. The AI agent finds the procedure page, reads the section on refunds and answers by it — with a link to the page, so the answer can be checked rather than taken on trust.

• A newcomer asks about leave where they read the procedures

  How leave works, the payment schedule and who grants access are described in the “For employees” section, but a new person asks in the chat, and HR answers the same thing for the tenth time.

  Result. The answer comes with a link to the right section of the procedures; the department’s rule “answer with a link to the procedure, not a retelling” is kept in the knowledge base and applied to every answer.

• A question right on the page

  Someone reads a procedure and does not understand how it applies to their case. Going to another tool for a question is an extra step people usually skip.

  Result. They mention the bot in a comment on the page. The answer arrives there, in the same discussion, and stays next to the text the question was about.

• A runbook instead of questioning the on-call engineer

  An incident at night, a new on-call engineer, and the steps are described in the wiki — in a five-page section.

  Result. The bot assembles the answer from the right pages of the section and answers to the point, with links to each. The on-call engineer reads three paragraphs instead of five pages.

1. @buff how do we number releases?

   a comment on any page of a selected space

   tool: search_pages + get_page

   By the YY.MM.N scheme — the year, the month and the release’s sequence number within the month; the scheme’s internal name is “Greenwich”. Releases go out on Thursday; the on-call engineer decides on postponement. Source: the “Release procedure” page in the DOC space.

2. @buff what constraints did we record for the external API integration?

   a question in the Buff app or a comment on the page

   tool: search_pages

   I found three pages: “External API integration” (a limit of 100 requests per minute and mandatory idempotency), “Retry decision” (exponential backoff, at most 5 attempts) and a discussion where it was agreed not to keep responses longer than a day. Links to all three are below.

3. /research why we dropped the shared cache

   a comment on the page with the decision

   tool: get_page

   The word after the mention picks the work: without it — a short answer, “research” — deep research with a document, “task” — a draft development task. Research reads the page, the pages related to it and the code.

4. @buff what does this section say about backups?

   a comment on the section’s parent page

   tool: get_page_children

   The section has four child pages; two are about backups: “Backup” (the schedule, a 90-day retention period) and “Restore” (the procedure and who approves it). Both are linked.

1. Step 1

   Installation takes one click

   Open the “Marketplace” section on the connector’s local page. Before installing you can see what the integration will get: the addresses it will be able to reach, which credentials it will ask for and how many tools the model will see. The connector downloads a signed bundle from our registry and checks the signature before writing anything to disk.

   Connector

   

   ← scroll the frame →

2. Step 2

   Setup: an address, access and a list of spaces

   For self-hosted the address looks like `https://wiki.corp.local`; for Cloud — `https://your-company.atlassian.net/wiki`. Next come the bot’s credentials (for Cloud, an email and an API token) and the space keys: you can see them in a page address, `/spaces/DOC/pages/…`; empty means all spaces open to the account. The secret is stored on your machine; the integration does not see it.

   Connector

   

   ← scroll the frame →

3. Step 3

   Create a separate account for the bot

   Give it read access only to the spaces where it should work, and permission to comment. The bot should not be a Confluence administrator: all it needs is to read pages and leave comments.

4. Step 4

   The same from the command line

   Installing without a GUI is an equal option, not a fallback: `connector install confluence`, then `connector instance add confluence --config base_url=https://wiki.corp.local --config spaces=DOC,OPS`.

5. Step 5

   Events arrive on their own

   Polling always works and needs no ports: the integration reads the comments of the selected spaces itself (with no list, of all spaces open to the account) — every 20 seconds by default; the interval can be changed in the connection settings. If you have Data Center and want it faster, turn on event receiving in the connector and add a webhook: it is not parsed for content, it simply wakes the poll, so the answer arrives at once rather than on the next pass.

   Connector

   

   ← scroll the frame →

The integration runs as a separate process under its own operating system user: it does not see the connector’s key, does not read the files of other connections and cannot reach the network past the connector. More in the security section.

• Confluence Data Center, Server or Cloud — one and the same integration.
• A bot account with permission to read the spaces needed and to comment: a token for self-hosted, an email and an API token for Cloud.
• Network access from the connector’s machine to Confluence — nothing needs to be opened from your network to the outside.
• A connector run with the right to create a separate OS user for the integration (root or CAP_SETUID).
• Tasks and code are connected separately: Confluence is a source of knowledge, not a tracker or a repository.

Will you copy our knowledge base to your side?

No. The integration reads pages at the model’s request when an answer requires it, and transfers or indexes nothing in advance. What is read enters the context of a specific answer — not a store.

We have Cloud, not Data Center. Do we need a different integration?

No, this same one. We checked it on live servers of both editions: their APIs match in everything the integration needs. Sign-in differs — Cloud uses an email and an API token, and the connection form has a separate field for that. For Cloud, give the address together with `/wiki`: `https://your-company.atlassian.net/wiki`.

Will the bot be able to change anything on the pages?

No. It reads pages and writes comments; it cannot create or edit pages — the integration has no such tool at all.

How do I limit what it sees?

In two ways at once: by the permissions of the bot’s account in Confluence and by the list of spaces in the connection settings. The list works both ways. If it is empty, the bot works in all spaces its account can see, so create a separate account for the bot and give it access only to the spaces needed — or list them in the connection.

Does a big page mean a big bill?

A page is read whole and is not cut: a truncated procedure is a silently lost requirement. But it is read only when the model goes for it, and the answer shows which pages were read.