Context blocks

A context block is a short piece of information that accompanies every agent reply. Use it for facts that must always be available, such as company identity, audience, tone of voice, and universal rules.

The block belongs to the organization. It only affects an agent after it is attached in that agent’s Knowledge section and is in a state the runtime uses.

Context block or vector base?

	Context block	Vector base
When it enters the prompt	On each agent call, without a search	When the agent queries the base
Recommended size	Short and stable	Large or consulted on demand
Cost	Adds tokens to each call	Adds tokens when a query and result occur
Good for	Identity, tone, short policies, and shared rules	Manuals, catalogs, FAQs, and long documentation

Ask: does the agent need to know this in every reply? If so, use a context block. If the information is only needed in specific situations, use a vector base or an instruction in the agent’s prompt.

Open Contexts

1. In the admin panel, open Knowledge Bases.
2. Select the Contexts tab. Its current description is Short text that ships with every single reply.

The current URL is /admin/bases/context. The old /admin/context address redirects to it.

Create a block manually

1. On the Contexts tab, select New block. If the organization has no blocks yet, use Create your first block.
2. Fill in Name with up to 120 characters. This name also enters the prompt as the content heading; choose a clear label such as Tone of voice or Refund policy.
3. Write the Content in Markdown. Use Edit to write and Preview to check the formatting.
4. Select Save block.

Content accepts up to 200,000 characters. The interface shows a rough token estimate, calculated as one token per four characters, and warns above 2,000 tokens. Do not treat this number as an exact model token count.

A new manual block is created as Active, but is not attached to any agent yet.

Start from the website

When the organization has no blocks, Build from the website provides two ways to prepare drafts:

• Enter Company name and, optionally, Company website to research public sources and generate one block.
• Select I’d rather write it myself to create three drafts: What the company does, Who it is for, and What to never do.

In both cases, the result remains a Draft. Review the text and, when research was used, verify the sources shown in the editor before activating the block.

Caution

Build from the website uses an AI model and consumes credits. If the model call occurs and a later step fails, the attempt may still generate usage. The current screen does not show this warning before execution.

Understand the states

State in the interface	Enters the prompt?	What to do
Active, with no badge	Yes	Keep it while the content remains correct
Draft	No	Review it and use Turn on
Marked as outdated	Yes	Correct or confirm it; the mark does not stop its use
Off	No	Use Turn on to use it again

After 180 days without confirmation, the list asks for a review. Still valid updates the confirmation date but does not change the content or attachments.

Use Turn off to remove an active block from the prompt without deleting it. Use Delete block only when it is no longer needed: deletion also removes its agent attachments. The dialog reports how many agents are attached before confirmation.

Review suggestions and history

The See suggestions and history button opens Suggestions and history, which brings together pending proposals and recent block versions. History keeps the 20 most recent versions. Using Go back to this one on a version creates a new current version with the chosen content; it does not erase the rest of the history.

Review a suggestion before applying it. It may change a premise that every attached agent receives on future calls.

Attach it to an agent

1. Open Agents and edit the agent.
2. Go to Knowledge.
3. Under Context blocks, select the blocks you want.
4. Use the section’s save bar.

The runtime organizes valid blocks under Organization context in the system prompt, in attachment order. For each block, it includes the Name as a heading and then the Content. Draft and Off blocks are not sent; active and Marked as outdated blocks are sent.

Caution

The Suggested for all badge is only an indication in the current product. It does not attach the block to new agents automatically. Open each agent and confirm which blocks are selected.

The agent editor may list blocks in any state without showing their state, and its total estimate may include a selected draft or off block. To determine what actually enters the prompt, use the state table above.

See Native tools — overview to understand the relationship between bases, context blocks, and AutoLearn.

Current permissions

The Contexts route requires base read access. However, create and update controls may appear for someone who cannot complete them, because backend authorization still follows the legacy administrative role. If a visible action returns a permission error, ask an organization admin or owner to perform it.

What to write

Good candidates:

• Identity: what the company does, who it serves, and which market it operates in.
• Tone of voice: how the brand speaks, what it avoids promising, and what it does when it does not know.
• Short, universal rules: hours, a brief policy, or a restriction that applies to every interaction.

Avoid:

• catalogs, long tables, and manuals; use a vector base;
• instructions exclusive to one agent; put them in that agent’s prompt;
• unstable content without someone responsible for reviewing and confirming it;
• secrets, credentials, or unnecessary personal data.

Before activation, verify that the text is factual, short, understandable without an image or external context, and safe for every attached agent.