This page traces the path from a connected Source to a published answer or Document, and how that knowledge then reaches both people and coding agents.
The knowledge lifecycle#
Dosu manages knowledge in four stages, each mapped to a section of these docs.
- Connect. You add Sources to a Library. A Source is a tool Dosu learns from, such as a GitHub repository, a Slack workspace, or the web.
- Generate. Dosu drafts answers, Documents, and changelogs from what it finds in those Sources.
- Maintain. When code changes, Dosu flags the Documents that may be affected and proposes updates, so knowledge keeps up with the work.
- Organize. You group related Documents with Topics and shape generated Documents with Templates, all inside a Library.
The sections below follow a single request through this lifecycle.
Step 1: Connect and index Sources#
When you add a Source to a Library, Dosu indexes its content so it can retrieve relevant information later. For a GitHub repository, that includes code, issues, pull requests, discussions, and the wiki. For Slack, it includes channel messages and threads. Dosu watches connected code repositories and re-indexes automatically, so the knowledge base reflects the current state of your work.
Web is the exception. Dosu does not crawl, index, or store website content. It reaches the web at query time using search, so a Web Source returns live results rather than a stored copy.
Step 2: Retrieve the right context#
When a question comes in, Dosu searches the indexed Sources in the relevant Library to find context that bears on it. That can include code, documentation, past issues, and conversations. Dosu can also answer questions about Dosu itself by searching the official product documentation.
Retrieval is scoped to the Library behind the Agent that received the question, which is how a repository Agent and a Slack channel Agent each draw on the right Sources instead of the whole organization at once.
Step 3: Draft an answer or a Document#
From the retrieved context, Dosu generates a response that synthesizes what it found, with citations linking back to the source material. The same drafting step produces two kinds of output:
- An answer posted where the question was asked, for example a reply on a GitHub issue or in a Slack thread.
- A Document, a saved knowledge artifact that lives in a Library. Save This Answer turns a good answer into a Document, and Generate Docs drafts new Documents straight from your Sources.
Every draft carries a confidence score that reflects how much relevant information Dosu found. High means strong supporting context, medium suggests a closer look, and low is held back rather than published.
Step 4: Review before anything goes out#
You stay in control of what Dosu posts. New Documents and auto-generated responses land as drafts on the Review page, where you can:
- Accept a draft and publish it as written.
- Edit it first, then publish.
- Decline it if no response is needed.
The same Accept, Edit, and Decline choices appear in the in-PR and in-MR Knowledge review comment, so reviewers can act without leaving the pull request. Until you are confident in response quality, you can keep an Agent in preview mode and review privately before Dosu ever replies in public.
Maintaining knowledge over time#
Publishing a Document is what lets Dosu keep it current. When a pull request or merge request is opened, Dosu checks the changes against your published Documents and comments with links to the ones that may be affected. When that PR or MR merges, Dosu reviews the changes again and proposes updates, with a link back to what changed. This is the knowledge-review and stale-docs behavior, and it runs through the same Accept, Edit, and Decline flow. A Monitor watches a Source and flags Documents that may be going stale.
How knowledge is served#
Once knowledge lives in a Library, Dosu serves it back in two directions.
To people:
- Answers in GitHub and Slack when someone mentions Dosu or an Agent replies on its own.
- The web app, where you can ask Dosu directly in Chat. A GitHub or Slack reply answers once per mention, while the web app supports a back-and-forth conversation with follow-ups, plus capabilities those surfaces do not have, like generating and maintaining Documents.
To coding agents:
- The MCP Server lets assistants like Claude Code, Cursor, and VS Code Copilot search your Library and pull project-specific context on demand.
Serving the same reviewed knowledge to humans and agents keeps answers consistent and cuts the cost of running coding agents across many repositories and teams.