Perplexity Decider 1.3 pulls passages from documents and shows where they came from

Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12 with a feature that returns a passage from an input document and shows where it starts and ends. Called Locate, it gives an application the text it was looking for and the positions needed to highlight it. A support system could classify a ticket and give a reviewer the sentence that explains why a customer is asking for a refund.

Earlier Decider versions scored supplied answers or assigned ratings on a predefined scale. In v1.3, a developer can describe what to find in ordinary language and receive the relevant wording from the input. A reviewer can open the document at that spot to read what surrounds it. Perplexity says the feature returns an empty list when it finds no match.

Classify a document, then surface its wording

Consider a request that asks whether a ticket is about delivery or returns, and asks for the customer's reason for returning an item. The classification result supplies probabilities for the choices, while the extraction supplies the passage with the stated reason. The same approach could mark the line containing a total on an invoice so a reviewer can see whether the model picked the total or another figure in the table. The positions refer to the text supplied to the model.

Perplexity's service can handle extraction, classification, and scoring questions about the same input in one request. The model returns values such as probabilities or passages in an application-readable form. On OpenRouter the extraction feature is exposed through a separate route from the one for decisions. OpenRouter lists support for returning an extracted string or a list of strings from text, with the source span and a confidence value.

Perplexity reported average scores of 69.0 for Decider v1.3 and 63.6 for Flash across 17 extraction tasks. The tests scored the answers in different ways. Entity tasks required an exact match for the passage boundaries, while questions answered from passages gave partial credit for overlapping text. A separate set tested whether the model filled the right fields. The resulting averages describe those tests. A different invoice or a set of customer messages can present different boundaries and layouts.

Model weights and hosted access

The models are available through Perplexity's Decisions API and OpenRouter, and Perplexity posted the weights on Hugging Face under the Apache 2.0 license. The published files let developers operate their own deployment. The model cards say that v1.3 expects a new input format. A deployment carried over from the v1.1 or v1.2 generation needs to prepare its requests for that format to reproduce the newer model's intended behavior.

Hugging Face lists 26 billion parameters for the larger release and 5 billion for Flash. OpenRouter gives both the same extraction response format. The Hugging Face cards describe sending images along with a request, whereas OpenRouter specifies text for its extraction route for these models. For cases that call for human review, where a passage starts and ends matters to the reader. A single word may leave out a condition in the sentence that changes the meaning, even when the model correctly points to that word in the original.

Separately from the extraction benchmark, Perplexity reported an analysis of one million requests from the first 18 days of its Decisions API. It used requests from users who consented to research, removed duplicates, and masked personal information. The company reported 93.98 percent agreement between Decider v1.3 and answers from the large language models it used as judges. That figure measures agreement with those models on the sampled requests, which lacked human reference labels.

OpenRouter's model pages displayed free input and output pricing for both variants when checked on November 14. Those listings cover OpenRouter's hosted service. Self-hosted copies run on the operator's own equipment, and OpenRouter's extraction listing specifies text input and its separate extraction route. The model pages also state that the free endpoints are rate limited.