PERPLEXITY / SEARCH & RETRIEVAL
Better evidence. Faster answers.
Hybrid retrieval and ranking systems that select relevant evidence for production RAG pipelines.
Explore the work
My contribution
Combined staged candidate filtering, embedding-based scoring, and selective cross-encoder reranking. Evaluated releases with retrieval benchmarks, LLM graders, human judgments, and A/B tests.
The work supported pipelines serving 200 million daily queries and contributed to 93.0% overall SimpleQA accuracy. Native INT8 quantization reduced embedding storage by 75%, with retrieval quality validated against full-precision baselines.
30.2% lower p50 latency