# Knowledge base The knowledge base is a set of markdown files. The daemon indexes them at startup and makes them available to agents and to people through search, MCP tools, and [MCP resources](https://modelcontextprotocol.io/specification/2025-11-25/server/resources). ## Source directories Three directories, each with a distinct role: | Directory | Purpose | |-----------|---------| | `wiki/` | How the org works, architecture, decisions | | `blueprints/` | What correct looks like | | `recipes/` | Step-by-step how to do things | Add a Markdown file to one of these directories. Then restart the daemon. ## Frontmatter All fields are optional. The daemon also indexes files without frontmatter. ```yaml --- title: Git Workflow # display title; defaults to the first H1, then the filename slug editable: true # allow propose_edit to target this doc (default false) category: conventions # tree position under the doc's URI scheme — single level tags: [git, process] # cross-cutting topic labels, lowercase-kebab description: How the team manages Git branches and changes # retrieval description --- ``` `category` groups docs into a branch in `bento kb`. Uncategorized docs sit at the scheme root. `tags` render as a suffix in the listing, filter via `bento kb --tag `, and improve retrieval context. `description` provides a short retrieval summary when the document uses QMD or Chroma. ## Method and delivery The [`knowledge:` block](/pipelines/config#knowledge) configures how knowledge reaches an agent run, one pipeline at a time. It has two settings: `method`, which sets how the daemon retrieves, and `delivery`, which sets how results reach the agent. Both compose with a `search` escape hatch. See [Method and delivery](/knowledge-base/modes) for the full reference. ## Querying the knowledge base Bento offers three ways to query the knowledge base. 1. Through MCP on your harness, such as Claude Code or Codex: ```bash (claude) "how do we handle git branches" => Returns LLM response ``` 2. Through the command line, with `bento ask`: ```bash bento ask "how do we handle git branches" bento ask "how do we handle git branches" => Returns LLM response ``` 3. Directly, as MCP resources: ```bash "Do we have any git resources in the wiki?" => Returns LLM list of wiki resources, e.g.: - wiki://git-workflow - wiki://git-commit-message-standards ``` :::warning A connected client injects MCP resources into the context without showing you, under the `wiki://`, `blueprint://`, and `recipe://` URI schemes. Account for that context cost when you query. The same behavior is what lets you structure and link knowledge across documents. ::: ## Why the separation The three methods give you a fast feedback loop, and they let you confirm that both a person and an agent retrieve the same knowledge consistently. Use them as a test. If no method here retrieves a document you know exists, an agent fails to surface it too when it counts, such as mid-run in a pipeline. ## Notes * **Docs are indexed at startup, not on save.** Adding or editing a file requires `bento daemon restart` to take effect. * **Prefer prose over tables for searchable content.** Tables and bullet lists embed poorly as standalone chunks. A prose sentence covering the same convention surfaces more reliably for natural-language queries. * See [MCP](/knowledge-base/mcp) for detailed information.