Perplexity adds page screening to its Context API

Perplexity added Decider to its Context API on 19 November, allowing developers to ask yes-or-no questions about fetched pages and receive a score between zero and one. Its official announcement…

JJames Bennett2 min read

Perplexity added Decider to its Context API on 19 November, allowing developers to ask yes-or-no questions about fetched pages and receive a score between zero and one. Its official announcement presents screening for prompt injection before an agent reads a page as a use case.

For commerce teams, that could provide an additional check when an assistant retrieves supplier information, competitor product pages or other material the merchant does not control. The release adds page assessment to fetching; it does not establish that an external page is safe.

A score needs a policy

Our recommendation is to decide what happens at each threshold before putting a classifier into a live workflow. A suspicious page might be excluded, sent for review or used only in a restricted task. Record the question and result so a rejected source can be investigated.

A numeric score is not a guarantee, and the announcement does not give a universal threshold for merchant agents. Test ordinary product copy alongside pages that contain instructions directed at the assistant. Include ambiguous material such as quoted examples, customer reviews and technical documentation, where an instruction may be part of the content rather than an attempt to control your application.

Keep privileged actions behind separate checks

Screening source pages should not replace permission checks when an agent changes an order or accesses customer records. Treat fetched text as external content, restrict which tools the workflow can invoke and require the appropriate confirmation for consequential changes.

Measure false alarms as well as missed attacks. A filter that discards useful supplier pages can quietly reduce answer quality, while a low score should not grant a page authority over the app’s own instructions. Keep a route to human review when the source and the requested action conflict.

The Founder’s Inn has not benchmarked Decider’s detection performance. The linked cookbook supplied with the research returned a 404 during verification, so this report does not claim an exact request schema, question limit or production configuration. The official launch post and demonstration support the narrower feature description above.

ℹ️
Treat a page-screening score as one signal. Keep customer-data access and order changes subject to your application’s own permissions.
Perplexity Developers, Context API Decider announcement, 19 November 2026. Original announcement.

20 November update: current documentation and limits

The current criteria guide and prompt-injection cookbook are now accessible. They resolve the documentation gap noted at publication.

A request accepts one to five named yes-or-no questions, using the noul format. Scores run from zero to one. Crucially, scoring does not remove, reorder or block pages: the caller applies its own cutoff and routing policy. The API has no supported threshold field.

A successful page fetch does not guarantee a score. Check for returned criteria and handle scoring errors explicitly before sending content to your agent. The documentation also distinguishes fetching charges from additional scoring-token costs. These are implementation details verified from the current docs, not evidence of detection accuracy in a merchant deployment.

Filed underAI

More from AI

See all →