docs-keeper uses large language models to draft documentation. We believe in being explicit about where AI is involved, what it can and can't do, and where a human stays in the loop. This page is a plain-language summary; the binding terms are in our Terms of Service.
1. What is AI-generated
Every documentation change docs-keeper proposes — the prose, the code examples, the changelog entries — is drafted by AI. We surface this transparently: generated pull requests are clearly labelled as machine-drafted and attributed to the docs-keeper bot, never to a human author.
2. Which models we use
Drafts are produced by third-party large language models routed through OpenRouter (Anthropic Claude models by default). Code is parsed with Tree-sitter, embedded with Voyage AI, and reranked with Cohere as part of retrieval. Model selection can change as better models ship; the human-review safeguard below does not.
3. Safeguards that run on every draft
- Retrieval grounding.Drafts are grounded in your actual code via context-aware retrieval, not generated from the model's memory alone.
- Quality gate. A separate evaluation pass scores each draft and can send it back for revision before it ever reaches you.
- Hallucination check. Drafts are re-sampled and checked for internal contradiction; low-confidence spans are flagged.
- Code-example validation. Where enabled, code snippets are compiled in an isolated sandbox before a PR is opened.
4. The human stays in control
docs-keeper opens a pull request. It never pushes to your default branch and never auto-merges. You — or a human reviewer on your team — read, edit, and approve every documentation change before it lands. You are responsible for the accuracy of what you merge.
5. Your data is not used for training
Your repositories are never used to train models. We process the diff and surrounding context of merged PRs to draft documentation, and we do not persist your full source code. See the Privacy Policy and Data Processing Addendum for specifics.
6. Limitations
AI output can be wrong, incomplete, or out of date despite our safeguards. Treat every draft as a starting point that requires review, not as authoritative output. If you spot a systemic issue, tell us at [email protected].