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

## Executive summary

Exa introduces Exa Snapshot, a feature enabling version-controlled web search, allowing users to search the web as it existed on any past date. The presentation details how modern agentic systems perform searches differently from humans, requiring robust search infrastructure. The core value proposition is enabling sophisticated backtesting of predictive signals and financial models by eliminating the bias of live, real-time data, thereby preventing overfitting in agentic workflows.

## Key takeaways

- Agentic Search vs. Human Search: Agents are highly verbose and execute many searches to fully understand a task space, unlike humans who might give up after a few failed queries. Exa is designed to support these intensive, multi-search agentic tasks.
- Exa's Core Principles: Exa focuses on three core areas: quality (providing a new benchmark for assessing search tool performance separately from inference), cost-effectiveness (using dynamic highlights to return precise snippets, saving tokens), and low latency.
- Version-Controlled Search with Exa Snapshot: Exa Snapshot provides access to 400 billion+ web snapshots, allowing users to search by date. This is crucial for backtesting, enabling the simulation of what information was available at a prior point in time, which is vital for financial and predictive modeling.
- Backtesting and Overfitting Prevention: By using Snapshot, users can train models or agents on historical data (e.g., 1 day, 7 days, or 1 month before an event) to test if a feature or signal could have been reliably predicted, thereby preventing overfitting to current market information.

## Technical details

- Exa Architecture & Search Capabilities: Exa provides its own crawling infrastructure, database, embedding model, and retrieval architecture to support agentic web search. They offer ExaAgent Ultra for very long-running deep research tasks.
- Dynamic Highlights for Cost Efficiency: Exa implemented dynamic highlights, which return exact right snippets across all search results instead of full pages or single pages. This allows for higher quality results while using approximately 95% fewer tokens, significantly reducing downstream inference costs for agents.
- Version Search API: The system supports two modes: 1) Finding a URL as it was at or before a given date, and 2) Version Search, which simulates executing a search query at a prior point in time, returning only results available up to that date.
- Predictive Signal Generation: Users can use Snapshot to backtest complex features (e.g., adverse media scans, historical share price dips, or SEC filings) by setting up evaluation harnesses to check if a model or agent could have found a discrepancy before the market was aware of it.

## Practical implications

- Build engineers can integrate version-controlled search into data pipelines to enrich training datasets, ensuring that predictive features are sourced from a reliable historical record.
- The ability to backtest agentic features (like adverse media scans) allows teams to validate model performance against historical web data, significantly reducing the risk of overfitting to current information.
- The focus on token efficiency (dynamic highlights) is critical for optimizing the cost and throughput of large-scale agentic workflows.

## Topics

Agentic Workflow, Web Search, Version Control, Backtesting, Predictive Modeling, Information Retrieval, Ryan Ahern — Exa, AIE CODE summit

Source: https://www.youtube.com/watch?v=vS2IA9pq0Q4