Text-span extraction comes to Perplexity’s Decision models

Version 1.3 of Perplexity’s Decider models adds locate to extract exact passages, alongside classification and probability scores.

Responses from Perplexity’s decision models can now identify specific passages in the text being analyzed. The locate feature, added to Decider v1.3 and Decider Flash v1.3, returns selected excerpts, their start and end positions, and an associated score. Instructions can specify a type of information, such as a person’s name, or describe the desired passage more precisely. If no match is identified, the result is an empty list.

The update extends a family of models designed to select and evaluate rather than write free-form responses. An application supplies content and questions with possible answers: whether a web search is necessary, how to classify a document, or which level to assign on a rubric. Results take the form of probabilities that software can act on. Extraction requests can also identify contract signatories, skills mentioned in a job posting, or personal information.

The larger model is adapted from Qwen3.8-27B; the Flash version uses Qwen3.5-4B and learns through distillation to reproduce the larger model’s response distributions. The provider positions the larger model for nuanced judgments and detailed rubrics, and Flash for high-volume classification and routing. This distinction describes a trade-off between resources and quality, without guaranteeing that the same choice will suit every type of content.

Extraction adds a mechanism separate from ordinary decisions. The team trained a component to identify passage boundaries, using LoRA adapters and a dedicated locate marker. The original weights and classification component remain frozen during this stage. Because merging the adapters degraded performance on some decisions, they are kept separate and activated for extraction requests.

The published code and model card specify the feature’s limits: it requires nonempty text, does not accept images, and handles no more than 128 extraction questions per request. Visual classification capabilities therefore do not automatically extend to locate. A request can combine extraction and classification, but the published implementation runs separate forward passes for these two operations. A single request does not mean a single pass through the model.

The weights and code are distributed under the Apache 2.0 license. Local deployment still requires infrastructure: the larger model’s documentation specifies approximately 49 GiB for the BF16 weights alone, plus working memory on a CUDA GPU. Access is also available through the Decisions API and OpenRouter, which currently lists free variants subject to usage limits.

On the latter platform, extraction uses a dedicated endpoint, currently in alpha. Requested fields replace classification questions, with responses consisting of text or lists of text. The request format therefore differs from yes/no decisions, multiple-choice questions, and rubric scores, even when calling the same model.

To evaluate locate, the team assembled 17 out-of-domain tasks covering contracts, résumés, structured documents, scientific reports, and support tickets. Named-entity recognition and personal-information detection evaluations require extracted boundaries to match the references exactly. Extractive question answering instead gives partial credit for overlapping characters. Structured documents are evaluated by the proportion of correctly extracted fields.

The reported average score reaches 69.0 for the larger model and 63.6 for Flash, compared with 43.3 for Liquid d1, 41.5 for Clairic, and 33.6 for GLiNER large v2.5. This average combines the primary metrics of different tasks; it does not represent a general rate of 69% correct extractions. The comparison is conducted by Perplexity and involves models of different sizes, including an undisclosed size for Clairic.

Another internal study covers one million production requests, comprising more than 2.3 million questions, collected with explicit user permission and with personal information masked. Because this traffic lacks verified reference answers, language models generate the answers used for comparison. Agreement reaches 93.98%, compared with 93.68% for Decider v1.2. This result measures agreement with those judges, rather than accuracy established through human annotation.

Probabilities also require application-specific tuning. The documentation recommends calibrating thresholds on representative examples because the cost of a false positive differs across moderation, routing, and document filtering. A faithfully reproduced excerpt can still be the wrong selection: finding words in a source does not verify the truth of what they assert.

Among the internal uses described, the system evaluates pages collected for the search index, identifying error pages, anti-bot barriers, and content considered unhelpful. Another setup examines agent activity logs to find incidents missed by security rules or unjustified alerts. Proposed changes to those rules remain subject to approval by a member of the security team.