Perplexity adds source-linked text extraction to Decider Perplexity adds source-linked text extraction to Decider

What changes when the answer is selected text?

The launch expands a family designed for classification and scoring. Perplexity describes uses such as finding names, identifying personal information and extracting answers from documents. These are narrower jobs than composing a response. A source-linked field gives the consuming application two things to inspect: the selected content and the location from which it came. That makes it possible to display a highlight next to a populated form, saving a reviewer another search for the supporting sentence.

What does the software receive?

The model card defines each result as text with start and end character positions, plus a score. The end position is exclusive. No match produces an empty list. Locate accepts text, not images, and mixing it with decision questions requires separate computation passes. An application therefore receives a structured extraction result, not an assurance that a single combined request has the same latency as one classification request.

How would a simple extraction check work?

Consider this invented shipping note: Ship to Oslo on Friday. Counting characters from zero, Oslo begins at position 8 and ends before position 12. A hypothetical extraction returning those positions can be checked against the original string without asking another model. If it instead supplies Paris, the substring check fails immediately. If it correctly selects Friday when the requested field is the destination, the substring check passes while the task still fails. The example separates faithful copying from choosing the correct information. Neither result is a test of Decider; both illustrate checks a document pipeline can make.

What does the smaller model change?

The Flash model card describes a smaller version distilled from the larger model, with the same request interface and extraction capability. Its stated weight-memory requirement is about 8.5 GiB in BF16, plus working memory. The larger model card lists about 49 GiB for weights. That changes the starting hardware requirement for local trials. It does not include the rest of an application, input-processing memory or concurrent requests. A shared interface can simplify testing both sizes against the same saved examples without making their error patterns identical.

What remains an application decision?

In the shipping example, an empty result needs a different path from a confidently selected destination. A reviewer could inspect the original note, while the application leaves the destination unset. A useful evaluation would also include two cities, a corrected destination and a document with no destination. These cases ask whether the selected field answers the requested question, not merely whether it exists somewhere in the document. The distinction matters most when an extracted value feeds another action: highlighting evidence and accepting a business value are separate steps.