> ## Documentation Index
> Fetch the complete documentation index at: https://docs.orkestral.pro/llms.txt
> Use this file to discover all available pages before exploring further.

# Knowledge base

> A wiki-style brain per workspace, generated from your code.

The knowledge base is a wiki-style brain for each workspace. It holds interlinked pages auto-generated from your repositories, so your AI team always answers with real context about your code, not guesses.

Everything is built and searched **on your machine**. No cloud, no telemetry.

<CardGroup cols={2}>
  <Card title="Wiki pages" icon="brain">
    Structured pages describing modules, flows, and decisions in your workspace.
  </Card>

  <Card title="Wikilinks + graph" icon="code-branch">
    Pages link to each other, forming a navigable graph of your project.
  </Card>

  <Card title="Local RAG" icon="magnifying-glass">
    BM25 lexical search plus on-device embeddings retrieve the right page.
  </Card>

  <Card title="Agents that learn" icon="robot">
    Agents read the KB before acting and write learnings back to it.
  </Card>
</CardGroup>

## What the knowledge base is

Each workspace gets its own knowledge base, a private wiki that lives next to your repositories in `~/.orkestral`. It captures what your codebase is and how it works, in plain language, so that both you and the agents can reason about the project quickly.

Think of it as institutional memory: instead of re-reading every file on every request, the team consults a curated set of pages that summarize architecture, conventions, and past decisions.

## Pages, wikilinks, and the graph

A knowledge base is made of **pages**. Each page covers one idea, a module, a service, a data flow, a convention, or a decision.

<Steps>
  <Step title="Pages">
    Every page has a title and a body. Pages describe a concrete part of your workspace and stay focused on a single topic.
  </Step>

  <Step title="Wikilinks">
    Pages reference each other with wikilinks. When one concept depends on another, a link connects them so you can follow the thread.
  </Step>

  <Step title="Graph view">
    The **graph view** renders every page as a node and every wikilink as an edge. Use it to see clusters, find orphaned topics, and understand how your project hangs together.
  </Step>
</Steps>

<Tip>
  Follow wikilinks the way you would in any wiki. Start from a high-level page and drill down into the modules it mentions.
</Tip>

## How it is built from your repos

The knowledge base is generated from the repositories in your workspace, then made searchable with a hybrid retrieval engine, all running locally.

<AccordionGroup>
  <Accordion title="BM25 lexical search" icon="magnifying-glass">
    A classic keyword ranking algorithm matches exact terms, function names, error strings, file paths. It excels when you know the precise words you are looking for.
  </Accordion>

  <Accordion title="Local embeddings (RAG)" icon="brain">
    On-device embeddings capture meaning, not just keywords. They retrieve pages that are conceptually related even when the wording differs. Combined with BM25, this is retrieval-augmented generation (RAG) running entirely on your hardware.
  </Accordion>

  <Accordion title="Fully local" icon="folder">
    Indexing and search happen on your machine. Your code never leaves your device to populate or query the knowledge base.
  </Accordion>
</AccordionGroup>

<Note>
  The hybrid approach matters: BM25 nails exact identifiers, while embeddings catch the "what is the thing that does X?" questions. Together they surface the most relevant page for each query.
</Note>

## How the agents use it

The knowledge base is wired into the team's workflow, not a side feature.

<CardGroup cols={2}>
  <Card title="Consult before answering" icon="bullseye">
    Before an agent plans or writes code, it retrieves the most relevant pages so its answer is grounded in your actual project.
  </Card>

  <Card title="Write learnings back" icon="bolt">
    As the team works through issues and reviews, new understanding gets written back to the KB, so the brain keeps improving.
  </Card>
</CardGroup>

This creates a loop: the more your team works in a workspace, the richer its knowledge base becomes, and the better grounded every future answer is.

<Warning>
  Knowledge bases are per workspace. Switching workspaces switches the brain, pages and graph from one workspace do not bleed into another.
</Warning>

## Next steps

<CardGroup cols={2}>
  <Card title="Agent team" icon="users" href="/en/agents">
    See how the CEO and specialists delegate work and consult the KB.
  </Card>

  <Card title="Workspaces" icon="folder-tree" href="/en/workspaces">
    Learn how repositories map to a workspace and its knowledge base.
  </Card>
</CardGroup>