Perplexity Ships Query-Aware Compression to Cut Retrieval Tokens by 70%.

Perplexity has shipped a production-ready approach to retrieval-augmented generation that dramatically cuts the amount of context fed to its answer models. The company announced on LinkedIn that its query-aware compression system reduces context tokens by up to 70% while simultaneously improving answer quality.

The technique addresses a core challenge in RAG systems: not all retrieved content is equally useful. Perplexity's approach culls ads, navigation elements, metadata, and other low-signal text before handing context to the answer model. The result is a 63% increase in vital content per snippet, effectively raising the signal-to-noise ratio for each query.

Perplexity reports achieving a 50x compression ratio on SimpleQA, a benchmark for factual question answering, while maintaining performance levels comparable to frontier models. The company's contribution isn't context compression itself, which has been explored in RAG research, but making it query-aware, citation-preserving, and fast enough to run in production orchestration.

Query-aware compression means the system tailors what it keeps based on the specific question, rather than applying a generic filter. Citation preservation ensures that even after aggressive pruning, the model can still attribute information to specific sources, a requirement for Perplexity's answer format.

"Better context is better than more context," the company wrote in its announcement. The shift reflects a broader trend in AI systems away from simply cramming more data into context windows and toward smarter preprocessing.

For developers building RAG applications with tools like Supabase or orchestration platforms like Make, the technique offers a potential blueprint. Perplexity's focus on speed is critical: compression that takes too long to run defeats the purpose in latency-sensitive search.

The move comes as Perplexity continues to expand its infrastructure capabilities. The company recently introduced ROSE, a custom inference engine for faster GPU model serving, and added Finance Search to its Agent API for live data retrieval. Query-aware compression appears to be another piece of Perplexity's broader effort to make retrieval-augmented generation practical at scale.

The announcement highlights a key tension in building AI-powered search: larger context windows from models like Claude or Gemini don't eliminate the need for better retrieval. In fact, they may make compression more valuable. Reducing tokens cuts both latency and cost, especially when running inference on expensive frontier models.

Perplexity's emphasis on maintaining citation quality is also notable. Many compression techniques sacrifice traceability for speed, but Perplexity's system is designed to preserve source attribution throughout the pipeline. For applications where verifiability matters, that's not a nice-to-have, it's a requirement.

The company has published technical details in a research blog post, signaling that the approach may influence how other teams tackle RAG pipelines. Whether this becomes a standard layer in retrieval systems, or remains specific to Perplexity's architecture, depends on how easily the technique generalizes. But for now, it's a clear signal that RAG optimization is moving beyond embeddings and reranking into smarter, query-aware preprocessing.