Skip to main content
Topic control answers a different question from moderation or prompt injection. It asks whether your application should talk about a subject at all. That makes it a product-scope control, not a general safety detector. Anthale applies topic control from the policy you configure. The policy decides which allowed topics exist, which blocked topics override them, and whether a request on the input or output path should continue.

​
What topic control is for

Use topic control when the assistant should stay inside a known domain such as customer support, documentation help, product education, or internal operations. It is the right control when the main question is “should this assistant cover this subject” rather than “is this content harmful.”

​
How the rule model works

Topic control works through an allowed list and a blocked list. The allowed list defines the subjects the assistant may cover. The blocked list carves out exceptions or clearly prohibited subtopics. When both apply, the blocked list wins. If an allowed list exists, topics outside it are denied by default. A subtopic inherits from its parent unless it is explicitly blocked. Requests that are ambiguous should resolve conservatively toward restriction. For example, if the allowed list contains products and customer_support, and the blocked list contains products.pricing_strategy, then products.features remains allowed, products.pricing_strategy is blocked, and unrelated subjects outside the allowed list are denied. Topic control is not only a keyword filter. In practice, people ask about the same subject in several different ways, so topic policies work best when the topics are written as real domains of responsibility rather than as a pile of surface terms.

​
What it does not replace

Topic control does not replace Prompt Injection Protection, which handles authority manipulation, or Content Moderation, which handles disallowed content categories. A request can be in scope and still unsafe, or harmless and still out of scope.

​
When topic control is the right tool

Use topic control when the main product question is whether the assistant should cover a subject at all. If the issue is harmful content or attacker control, use moderation or prompt injection protection instead. Read Scope Escape when you need to reason about evasive out-of-scope requests.

​
Next steps

Next, read Scope Escape or go to Link Control.
Last modified on March 19, 2026