Perplexity has upgraded its Search API with significant improvements to snippet extraction and new evaluation capabilities, the company said in a technical update.

The AI-powered search company focused on snippet quality optimization across two key dimensions: relevance and content size. Better snippets directly impact downstream answer accuracy and reduce token costs for developers using large language models.

Span-level snippet evaluation

The company built a new evaluation system that identifies and labels specific segments within source documents based on their relationship to search queries. The system categorizes content spans as "vital" for inclusion, "irrelevant" for exclusion, or duplicates.

This span-level approach enables more precise measurement of snippet quality, tracking both correctly included and correctly omitted content. The improvements allow Perplexity to generate smaller, more relevant snippets while handling structured data formats including tables, nested lists, and dynamically rendered content.

Internal evaluations revealed that smaller content budgets produced more accurate results after the quality improvements. The company adjusted default configurations to reduce response payload size and latency while delivering more useful content per result.

SEAL benchmark performance

Perplexity introduced results from the SEAL benchmark, which tests retrieval systems on time-sensitive questions where correct answers change over time. The benchmark requires real-time index freshness and smart extraction from continuously updated sources.

Using Claude Sonnet 4.5 on the February 22 SEAL release, Perplexity's scores increased while other providers declined on SEAL-Hard questions. The company has integrated SEAL into its open-source search_evals framework alongside existing benchmarks.

New API features

The Search API now supports up to five queries in a single request, with results grouped per query in submission order. This reduces round trips for applications running parallel searches or agents decomposing complex questions into multiple retrieval tasks.

Expanded filtering options include language filtering by ISO 639-1 code and regional search by ISO country code, complementing existing domain and recency filters. These can be combined for precise result scoping.

The Python SDK now provides native Search API support alongside existing Agent API and Sonar API integration. Full documentation is available at docs.perplexity.ai.