Skip to main content

​
What you get

The retrieval pattern behind a good “ask a research question” feature: one user question becomes many angles, each angle its own search, and the results merge into a single summary the user can trace back to papers.

​
Who it’s for

Product engineers building research answers into an application — the difference between this and a single search is the difference between a demo and a product.

​
How Phylo runs it

Phylo’s agents make parallel calls across multiple angles of a question, gather the relevant papers, and distill the results into a summary for the user. The workflow starts wherever the researcher is:
“It’s very robust, with almost zero errors. It gives you genuinely useful results, and it embeds into a high-throughput agentic workflow easily.” — Kexin Huang, Co-founder & CEO, Phylo

​
Generating good angles

This is where the quality lives. A weak decomposition produces five rephrasings of the same query; a strong one covers genuinely different retrieval paths.
1

Vary the axis, not just the wording

Cover the mechanism, the population, the outcome, the competing explanation, and the methodology. Five paraphrases of one angle retrieve nearly the same papers and give false confidence in coverage.
2

Include the null and the contradiction

Add an angle phrased toward the opposite finding. Searching only for confirmation returns confirmation, and users notice when your product never disagrees with them.
3

Use the field's terminology, not the user's

The user’s phrasing is rarely what the literature indexes under. Translate first — a reconnaissance search to learn the vocabulary, then the real angles.
4

Bound the concurrency to your tier

“Parallel” here means angles, not simultaneous calls: there is a one-query-per-second limit on every plan except Enterprise, so issue the searches sequentially. See rate limits.
5

Merge on DOI and keep the angle that found it

A paper surfaced by several independent angles is a stronger signal than one found by a single query — that count is worth surfacing in the answer.

​
Distilling without losing the thread

The synthesis step is where grounding usually breaks:
  • Group by finding, not by paper. A paper-by-paper summary is a reading list; grouping by finding with the supporting papers attached is an answer.
  • Attach a citation to every claim and drop sentences you cannot support. Make “the literature does not settle this” a real output.
  • Show which angles were searched. It tells the user what was and was not covered, and it is the cheapest trust-building thing you can render.
  • Report disagreement. Where angles returned conflicting evidence, say so rather than averaging it into a confident middle.

​
What to check before you trust it

  • Inspect the generated angles on real questions. If they are paraphrases, fix the decomposition prompt — no amount of retrieval quality compensates.
  • Test a question with a genuinely contested answer. The output should surface the conflict, not resolve it silently.
  • Watch merge behaviour on near-duplicates. Preprint and published versions of the same work can carry different DOIs.
  • Measure it. Use a benchmark to check whether more angles actually improve recall, or just cost more.

Screen thousands of target-disease pairs for evidence

The batch counterpart, on the same infrastructure.

Add a citable evidence layer to an agent you already ship

Wiring this into an agent you have already built.

Best practices

Fan-out and merge, rate limits, and the five primitives.

All use cases

Browse the gallery by persona.