✳flâneur — a map of the web's best reading

AI Search Insights | Exa

exa.ai · 5,270 words · saved by 2 readers

At Exa, we've built our own search engine from the ground up. We developed a distributed crawling/parsing system, trained custom embedding and reranking models, designed a new vector database -- all in order to build the best search engine to serve AI applications. When we evaluate Exa search on a variety of benchmarks, we consistently perform state of the art compared to other search APIs. Using Exa retrieval in your application should improve performance on downstream tasks. But what does "best" actually mean? How do you evaluate web search quality? In this post, we'll share both our evaluation results and our philosophy behind evaluating web search. We can evaluate search result quality by having LLM graders score the relevance and quality of results returned for different query sets. We run queries through each search engine, then pass each (query, result) pair to an LLM grader, which independently evaluates the relevance and quality of the result for the query, outputting a score

Our AI Research: How We Evaluate Semantic Search Technology | Exa Blog Introducing Exa Agent Introducing Exa Agent: frontier web research at a fraction of the cost. Read more Products Search Contents Deep Agent Monitors Resources Careers We're hiring Case Studies Partners Demos Contact us Research Brand Pricing Blog Docs About Contact sales Sign up How we do evals at Exa Michael Fine May 30, 2025 Evaluating the Best Search Engine At Exa, we've built our own search engine from the ground up. We developed a distributed crawling/parsing system, trained custom embedding and reranking models, desig

Explore this link on the map →

saved by

related reading