Deep Verify
An AI agent drives the second pass — it plans queries, reads results, and decides when it's confident, instead of following a fixed pipeline.
What is Deep Verify?
Deep hands the citation to an AI agent. The agent reads your citation, decides what to search for, runs queries against academic databases, judges the candidates, and either commits a verdict or keeps looking. Use it when Fast leaves you with an unsure or not-found result.
The default. Runs a precise structured search across academic databases — no AI agent in the loop. Cheap, fast, and great for clean reference lists.
An AI agent drives the search end-to-end: it picks the queries, reads the results, and judges each candidate. Best for citations that Fast can't confirm.
When to use Deep
Reach for Deep when Fast returns Unsure or Not Found, or when you already know the citation is tricky and want the AI agent on the case from the start.
- 1
In-text citations
Sentences that describe a cited work without a full bibliographic entry (e.g. "Smith et al. proposed X"). Fast's parser can't help — the agent handles them natively by reasoning over the description.
- 2
Ambiguous or short references
Citations missing critical fields (e.g. only an author and year, no title). The agent improvises queries from whatever signal it has, and broadens the search if the first attempt is inconclusive.
- 3
Likely-hallucinated references
Citations from AI-generated text where you suspect the source may not exist. The agent looks harder before giving up — and is stricter about confirming, so fabricated references don't sneak through.
- 4
Critical references
When you need extra confidence — for example before a submission deadline. Sending the citation through the agent buys you a second, more thorough opinion.
Examples
Three concrete cases where Fast struggles and the AI agent earns its credits.
In-text citation
Fast can't parse this — there's no bibliographic entry, just a descriptive sentence. Result: skipped or Unsure.
The agent reads the sentence, extracts "transformer", "self-attention", "Smith 2017" as search signals, looks up the matching paper, and confirms the match.
Truncated reference
Fast has nothing to match against — no title, no venue, no co-authors. Result: Not Found or low-confidence Unsure.
The agent improvises: it tries "Wang 2018" alone, then broadens with related terms from surrounding context, judges each candidate against what little it knows, and either commits a match or honestly returns Not Found.
Likely-hallucinated reference
Fast may surface a partial title-similarity hit and call it Unsure, leaving you to figure out whether the paper is real.
The agent runs multiple queries, finds no matching record across academic databases, and rules the citation Inauthentic — clearly flagging it as fabricated.
How to trigger Deep
Two ways: pick "Deep" from the dropdown next to the Verify button to run the entire batch through the AI agent, or click the per-citation Deep Verify button on a result card to hand just that one to the agent.
[1] Smith, J. (2020). On the foundations of citation verification.
On the foundations of citation verification
The per-citation button only appears on results that finished Fast — clicking it hands that single citation to the AI agent and re-runs it.
How the agent works
Deep runs an AI agent that drives the search itself — it decides what to query, reads results, and stops when it's confident.
- The agent reads your citation and plans its own search strategy — author + year, title keywords, venue, or paraphrased topic — instead of following a fixed query template.
- It queries academic databases through tool calls, inspects the returned candidates, and decides whether to commit a verdict or run another search.
- Once the agent reaches a confident match (or rules the citation out), it stops — no extra calls, no extra credits.
- If the agent can't confirm anything, you'll see Not Found — and you can request Suggested Replacements (see the next guide).
Cost
Deep costs roughly 5× the credits of Fast per citation, because the AI agent runs multiple searches and uses a more capable model to judge each batch of candidates. The exact rate is shown on the dropdown description.
Tip: run Fast on the whole batch first; only invoke the agent on the problem citations using the per-result Deep Verify button. That keeps the cost low.
Best Practices
Two workflows cover almost every case — pick the one that matches your input.
Run the entire text through Fast first. It's cheap, fast, and clears the easy ones in one pass. Then look at the cards that came back Unsure or Not Found and click the per-result Deep Verify button to hand only those individual citations to the AI agent. This is the recommended workflow for most papers and reference lists.
For complex inputs — paragraphs full of in-text references, mixed-format citations, sources with little structure, or text where Fast keeps coming back inconclusive — pick Deep from the dropdown for the whole batch. The AI agent is the only mode that can handle these end-to-end.