Perplexity Decisions API adds web grounding, boosting accuracy from 51% to 86% on PWS-Decisions

• Perplexity’s Decisions API now supports web grounding, allowing the model to search the web before answering to handle facts that change after training.
• On the PWS-Decisions benchmark, accuracy rose from 51% to 86% with typical responses in under 400 ms, outperforming the published leaderboard for public-set accuracy.
• In a separate test of 328 true-or-false claims, web grounding increased correct answers from 56.4% to 97.3%, cutting errors from 143 to 9, though median latency rose from 139 ms to 366 ms.
• Users enable the feature by adding "mode": "web_grounding" to the request, with each completed search costing $0.001 in addition to input tokens.

Implementation Details

• The API gateway runs Fast Search queries in parallel for each question, merging results and appending evidence (title, URL, page text) to the state before the model call.
• State can be text, JSON, images, or a URL; if a URL is provided, the search query carries that URL so the page itself serves as evidence.
• The response includes a search_results field showing the pages the model read, while the request and response shape remain otherwise unchanged.

Availability and Limits

• A request using Web Grounding can include up to 100 questions.
• Search evidence counts toward input tokens, and the model answers from available evidence if a search fails or times out.