How to Build a Research Agent with Exa Find Similar and Deep Research

November 9, 2026

The first Exa query is easy: describe what you want, get pages. The second question is harder. What do you do when you found one excellent source and need the other nineteen that a human researcher would have found by clicking through citations?
That is what find_similar is for, and it changes how you build a research agent. Instead of guessing at query phrasings, you anchor on a document and fan out from it.
This guide covers three patterns: seed-and-expand, multi-hop deep research, and the deduplication step everyone forgets.

What does find_similar actually do?

It takes a URL and returns pages with similar embeddings. No query text required. Set exclude_source_domain=True unless you want the same site’s coverage, which for news and blogs is often exactly what you do want.

res = exa.find_similar(url="https://example.com/post", num_results=20)

In practice the first five results are usually excellent and results 15–20 drift. Two things tighten that drift: pass include_domains to restrict sources, and cut at ten results unless you have a dedup step that can handle noise.

How do you do seed-and-expand without blowing your budget?

Start with one neural search on the topic, take the top result, then run find_similar on it. Repeat for two seeds. Now you have around forty candidate pages from three starting points, which is more than most research tasks need.
The budget maths: 3 searches plus 2 find_similar calls, all without contents, costs a handful of credits. The expensive part is always the contents fetch. That is why the pipeline order matters:

  1. Search and fan out with URLs only.

  2. Deduplicate by normalised URL and by title similarity.

  3. Score candidates cheaply, for example by domain trust and publish date.

  4. Fetch contents for the top eight only.
    Reversing steps 3 and 4 is the single most common way teams 5x their Exa bill. Fetching text for forty results and then discarding thirty-five is pure waste.

When should you use Exa’s deep research instead?

Deep research accepts a question and runs its own multi-step loop: search, read, refine, search again. It returns a written answer with sources rather than a list of links.
Use it when:

  • The question needs three or more hops (“which companies in this space raised money after their founders left Google”).

  • You want prose you can hand to a user without post-processing.

  • Latency tolerance is minutes, not seconds.
    Do not use it for link lists or single-fact lookups. It is slower and pricier than a direct search, and for a one-hop question the extra hops mostly add confidence, not information.

How do you deduplicate results across searches?

Aggregators, syndication and republished press releases mean the same text arrives under five URLs. Dedup at two levels:

  • URL normalisation. Strip utm_*, trailing slashes, www., and tracking fragments. Then compare.

  • Title and shingle similarity. Normalise the title to lowercase alphanumerics and compare; for near-duplicates, take the first 200 characters of the extracted text and check overlap.
    Skipping this produces a research report that cites the same TechCrunch article five times. Users notice faster than you would think.

What is the cheapest way to rank candidates before fetching?

Three signals, all available without a content fetch:

  • published_date — anything older than your window is out.

  • Domain tier — a static allow/deny list of 30 domains catches most of the noise.

  • Score from the search call — useful for ordering, not for filtering, because scores are not comparable across queries.
    A deterministic ranker beats an LLM ranker here. Saving fifty LLM calls to sort URLs is a rounding error on cost and a large win on latency.

What limits should you expect?

  • Index freshness. New pages take hours to a day to appear. A monitor built on Exa is not a real-time alert system.

  • Paywalled pages. text returns what is publicly reachable, so a WSJ article may come back as a stub.

  • PDFs. Exa extracts text, but table-heavy PDFs lose structure. Do not build a numbers pipeline on PDF extraction alone.

  • Rate limits. Call it in sequence with a small concurrency window. Parallel fan-out at 50 requests triggers 429s that cost more in retries than you saved.

Who should skip this approach?

If your users search within one known corpus, use that corpus’s own search. Exa indexes the open web; pointing it at an internal wiki returns unrelated public pages that happen to share vocabulary. Same if the research question is answerable by a database query: a graph fan-out over the internet is a strange way to look up last quarter’s revenue.

Quick Summary

  • find_similar(url=...) expands one good source into twenty without writing a query, which is the cheapest way to get a balanced source set.

  • Fan out without contents first, deduplicate, rank with static signals, then fetch text for eight results, not forty.

  • Deep research suits three-plus-hop questions and returns prose with sources; it is overkill for one-fact lookups.

  • Deduplicate by normalised URL and title shingles, or your report will cite the same syndicated article five times.

  • Exa’s index lags hours to a day on fresh pages, so do not build real-time monitoring on it.

Tags: