Brains: Your AI Knowledge Base

What is a brain?

A brain is the collection of information your AI assistant can search when answering questions. It can include documentation sources (websites that There There crawls), manually written articles, and MCP connections to external services.

Brains are linked to channels. When an agent uses the AI chat on a ticket, the assistant searches the brains linked to that ticket's channel. This means different channels can draw from different brains, which is useful if your workspace handles support for multiple products. A single brain can be shared across multiple channels, and a single channel can search multiple brains.

To link brains to a channel, go to the channel's settings and tick one or more brains in the brains list. You can also set this during the initial channel setup.

Documentation sources

Documentation sources are websites that There There crawls and indexes. By connecting your product documentation, help center, or API reference, the assistant can provide accurate, sourced answers to customer questions.

Adding a documentation source

1. Go to Settings > Brains and select the brain you want to configure (or create a new one).
2. Open the Sources tab and click to add a new source.
3. Provide a name for the source (e.g., "Product Documentation").
4. Enter the base URL of your documentation site (e.g., https://docs.example.com).
5. Optionally specify include patterns to limit crawling to specific sections, or exclude patterns to skip irrelevant pages. Patterns accept wildcards (e.g. /docs/*).
6. Save the source to begin crawling.

How indexing works

After you add a source, There There automatically starts crawling the website. The crawl status progresses through several stages: Pending, Crawling, Completed, or Failed. You can monitor the progress in real time, including the number of pages crawled and chunks created. Click Pages next to a source to open the source detail screen, which shows the page statistics, the list of crawled pages, and a test search.

Turning a website into something the AI can search happens in a few steps. The page statistics on the source detail screen (Pages crawled, Chunks indexed, Embeddings) reflect each of these steps.

1. Crawling. There There visits each page of your documentation and saves its content. This is the "Pages crawled" count.
2. Simplifying. Every page is converted to plain text and stripped of layout, navigation, and styling, so only the meaningful content remains. This is the same simplified text you see in the search results on the source detail screen.
3. Chunking. Each page is split into smaller sections called chunks. The AI retrieves the single most relevant chunk rather than a whole page, which leads to more precise answers. This is the "Chunks indexed" count.
4. Embedding. Each chunk is turned into an embedding, a numeric representation of its meaning. Embeddings let the AI find the right content based on meaning, even when a question uses different words than the documentation. This is the "Embeddings" count, which catches up to the chunk count once processing finishes.

When an agent or the AI answers a question, There There compares the question against every embedding and shows the closest matches. The similarity score (shown as a percentage) tells you how well a chunk matches the search: a higher percentage means a closer match in meaning. You can try this yourself with the test search on the source detail screen.

Keeping docs up to date

Documentation sources can be re-crawled to pick up changes. If your product documentation is updated frequently, re-crawl periodically to ensure the AI assistant always has the latest information. If a crawl fails, check the error message for details. Common issues include unreachable URLs or pages that block crawlers.

Articles

Articles are pieces of knowledge you write directly inside There There. Use them for information that does not live on a public website, like internal procedures, troubleshooting steps, or policies your team needs the AI to know about.

Articles are edited in a rich text editor and are indexed the same way as crawled documentation. The AI assistant searches them alongside your documentation sources.

Example articles

Not sure what to write? These are common articles teams add to a brain:

• Accessing past invoices and receipts. Step-by-step instructions for where customers find their billing history, how to download a PDF invoice, and who to contact for a VAT correction.
• Refund and cancellation policy. Your exact terms: the refund window, what qualifies, how long a refund takes to appear, and how to cancel a subscription.
• Resetting a forgotten password. The reset flow, what to do when the reset email does not arrive, and how to handle accounts tied to a single sign-on provider.
• Account and team deletion questions. What happens to data when an account is deleted, how to transfer ownership before leaving, and the retention period before data is permanently removed.
• Supported browsers and system requirements. The browsers and versions you officially support, plus any known issues and recommended workarounds.
• Escalating a production outage. An internal-only article describing how the AI should respond during an incident, including the status page link and when to hand off to a human.

Keep each article focused on a single topic. Several short, specific articles retrieve better than one long catch-all article.

MCP connections

MCP (Model Context Protocol) connections let you extend the AI assistant with tools from external services. For example, connecting your error tracking service lets the assistant look up recent errors when a customer reports a bug. See MCP Connections for the full guide.

Additional instructions

Each brain has an optional instructions field where you can add context for the AI. For example, you might write "This brain covers our billing product. When users mention invoices or payments, search here first." These instructions help the AI understand which brain is relevant to a given question.