Version-Controlled Search: The Web as It Was — Ryan Ahern, Exa

Highlights

• Exa offers version-controlled web search through Exa Snapshot, which indexes over 400 billion snapshots of the web searchable by specific dates.

• Dynamic highlights extract precise relevant snippets across all search results, reducing token usage by 95% compared to returning full-page text.

• Version search enables querying the internet as it existed at or before a specific past date to prevent future information leakage in model training.

• Financial and risk modeling teams backtest prediction algorithms and agentic features against historical web states rather than waiting months for live production feedback.

Timeline

1. 00:00:00 - 06:51:00

   Agentic Web Search and Exa Infrastructure

   Modern search tasks rely heavily on agents that issue frequent, detailed queries to gather comprehensive information. Exa provides customized retrieval architectures for diverse sectors like hedge funds, insurers, and venture capital firms. By combining proprietary crawling infrastructure with advanced snippet extraction, the platform minimizes context window token consumption and lowers overall inference costs for downstream applications.

   • AI agents execute verbose queries and perform multiple search iterations differently than human users.
   • Exa maintains its own crawling infrastructure, embedding models, and retrieval architecture.
   • Dynamic highlights return precise text snippets to reduce inference token costs by 95%.
   • Exa Agent Ultra provides long-running deep research capabilities built on proprietary search harnesses.
2. 06:51:00 - 11:42:00

   Exa Snapshot and Version-Controlled Search

   Continuous freshness checks enable the systematic capture of web pages across time. Users can retrieve specific URLs as they appeared on a given date or execute semantic searches restricted to historical indexes. This temporal isolation prevents future outcomes from contaminating past data analysis, allowing developers to audit information discovery prior to widespread public reporting.

   • Exa Snapshot provides access to over 400 billion versioned snapshots of web URLs indexed with timestamps.
   • Version search executes queries against the web state as it existed at or before a specified historical date.
   • Historical financial reporting discrepancies, such as OpenAI's revenue run rate estimates, illustrate how early data can be isolated.
   • Backtesting allows developers to verify whether models could have predicted market events using only information available prior to public announcements.
3. 11:42:00 - 19:04:00

   Backtesting Models and Future Architectures

   Systems across industries utilize versioned web data across three primary integration archetypes ranging from traditional machine learning enrichments to fully agent-driven simulations. Backtesting complex semantic queries, such as shareholder calls or multi-variable corporate distress indicators, verifies model generalizability. Testing against historical web states provides immediate performance validation without relying solely on real-time deployment cycles.

   • Traditional machine learning models integrate unstructured versioned web data like software pricing and terms of service.
   • Agentic features incorporate automated adverse media scans and equity research directly into existing model pipelines.
   • Workshops on company acquisition predictions demonstrate how complex semantic features avoid overfitting during backtesting.
   • Version-controlled search eliminates the need to wait months for live production feedback to validate predictive agentic setups.