SPARKIT
← Back to blog

Reports can now include figures from the papers they cite

A SPARKIT report has always been a cited, source-backed answer. Now it can also show the evidence. When a report cites an open-access paper that contains a figure directly relevant to your question, SPARKIT pulls that figure and embeds it inline — right next to the paragraph it illustrates — with a caption crediting the source and its license.

It runs on every research report. But it is deliberately conservative, and most of this post is about when a figure appears, because "sometimes there isn't one" is by design.

How it works

After the agent finishes writing, a separate best-effort pass looks at the sources the report actually cites. For each one it can, it:

  1. Resolves the paper to a DOI — PubMed, DOI, PMC, and Semantic Scholar links all resolve.
  2. Checks whether the paper is open access, and under what license.
  3. Extracts the figures from the open-access full text.
  4. Scores each figure against your question with a vision model, keeping only those that clearly illustrate the answer.
  5. Embeds the survivors inline, next to the most relevant paragraph, with attribution.

Because the pass is best-effort, it never delays or blocks a report. If any step fails, the report is simply returned without a figure.

When a figure is included

A figure appears only when all four of these hold:

  • It comes from a paper the report cites. Figures are only ever pulled from the report's own sources — every embedded figure traces back to a citation in the same report.
  • That paper is open access under a permissive license — specifically CC-BY, CC-BY-SA, CC0, or public domain. This is the load-bearing condition. Paywalled papers, and open-but-restricted licenses such as CC-BY-NC-ND, are excluded: those licenses don't grant the right to reproduce the figure, so we don't.
  • The figure is genuinely relevant to your question. Every candidate is scored for how directly it illustrates the answer, and anything below a firm bar is dropped. We would rather show nothing than pad a report with a loosely-related chart.
  • The figure is legible. Thumbnails, logos, and decorative graphics are filtered out.

Reports carry at most a few figures, and many carry none.

Why some reports have no figures

Usually it's the license condition. Entire fields publish mainly in subscription or CC-BY-NC-ND journals. A question about single-cell sequencing in a particular cancer might cite excellent work in Nature or Cell — none of it CC-BY, so none of its figures can be reused. Questions grounded in open-access-heavy literature — much of genetics, epidemiology, and microbiology lives in PLOS, BMC, and the PMC open-access subset — surface figures far more often.

This is intentional. A figure shows up only when we are allowed to show it and it earns its place. "No figure" means one of those two tests wasn't met — not that the answer is weaker for it. The citations are all still there, and every figure that does appear carries its source title, its citation number, and its license, so you can audit it exactly the way you'd audit any cited claim.

How to think about the deep-research tool landscape

Perplexity, ChatGPT Deep Research, Gemini Deep Research, Claude Research, Elicit, SPARKIT — the deep-research category has converged on the basics. The remaining differences are about audience, deployment surface, and how each tool treats citations. A map of who builds what for whom, and where SPARKIT fits.

Read post →