Exa Snapshot gives web search a cutoff date

Most search APIs expose the current web, which can contaminate AI evaluations when answers appear online after a task was created. Exa has launched Snapshot, a search capability that limits queries and page fetches to versions stored by a specified point in time.

Exa says its index contains more than 400 billion webpage snapshots spanning two decades. Snapshot is available as a research preview through the existing /search and /contents endpoints. The former discovers pages; the latter retrieves archived versions of known URLs.

When browsing becomes answer lookup

A benchmark written in June may have public solutions by September, including papers, pull requests, issue threads, and blog posts. An agent with web access can retrieve those answers during training or evaluation, allowing the reward signal or grader to credit retrieval of leaked material as successful task completion.

Choosing a cutoff before the task was published removes later page versions from the available evidence and reduces that source of contamination. It cannot prove independent problem-solving because an agent may have memorized the answer or obtained it through another channel.

Fixed cutoffs also limit drift caused by edited pages. Current retrieval signals may still change which eligible pages appear or how they rank, so teams requiring identical evaluation fixtures should persist the responses or URL manifests alongside the cutoff.

One timestamp sets the boundary

The Python SDK accepts an ISO 8601 timestamp through snapshot_as_of inside the contents options. With exa-py installed and EXA_API_KEY configured, a recent public-tier request looks like this:

from datetime import datetime, timedelta, timezone
from exa_py import Exa

cutoff = (
    datetime.now(timezone.utc) - timedelta(days=30)
).isoformat()

exa = Exa()
result = exa.search(
    "latest stable Python release notes",
    num_results=3,
    contents={
        "snapshot_as_of": cutoff,
        "highlights": True,
    },
)

The rolling example keeps the timestamp within the public tier’s five-month archive window. Raw JSON requests use the camel-case field snapshotAsOf. Benchmarks should store a fixed timestamp and confirm that the account’s archive access will continue to cover it.

Current ranking selects archived pages

On /search, Exa’s current retrieval system discovers and ranks candidate URLs. A URL qualifies only when Exa has stored a version at or before the requested timestamp, and the service returns the latest eligible version it holds. Historical result order may therefore differ from the search results shown on the cutoff date.

Every content-derived field comes from the archived version, including the title, author, publication date, text, highlights, and summaries. Snapshot availability depends on Exa’s crawl history, so page changes made between stored captures may be absent.

The preview caps history and traffic

Because the public window rolls forward, a fixed benchmark cutoff will eventually fall outside standard access. Evaluations tied to older events or long-running studies need deeper archive access before that happens.

Four uses for a time-bounded web

Point-in-time retrieval supports workflows where later publications, edits, or disclosures would invalidate the result:

• Agent training and evaluation: Reinforcement-learning runs can restrict browsing agents to information available when each task was created, reducing leakage into rewards and benchmark scores.
• Financial backtesting: Researchers can test web-derived signals using pages available by a historical date, matching the point-in-time discipline already applied to prices and company fundamentals.
• Documentation and policy audits: Teams can fetch the same URL at separate cutoffs and run their own diff across documentation, pricing pages, policies, or filings.
• Historical agent tests: Developers can rerun an agent against an earlier content boundary without exposing it to subsequent page updates.

The API docs provide the supported request formats and current compatibility details.

Historical research needs sampling care

Two decades of archived pages could also support research into the web before widespread LLM-generated publishing. Such studies still need to account for crawl coverage, missing versions, publication provenance, and the selection effects introduced by current retrieval signals.

Search-based studies inherit those modern ranking choices when assembling a historical sample. Researchers working from known URL collections can use /contents with fixed dates, while open-ended studies should record the query, cutoff, search mode, run date, and returned URLs.