Deep research

Deep research takes a question, searches the literature, and gives you back a ranked, deduplicated, clustered set of papers. Open it from the sidebar.

Run one

1. Open Deep Research from the sidebar and click New Research.
2. Write the question. Broad questions work well here; this is not a keyword search.
3. Pick a depth.
4. Click Start.

The sources

Choose the sources in the research configuration. Availability and defaults depend on the run and surface; not every database is enabled automatically.

The source selector includes arXiv, Semantic Scholar, PubMed, OpenAlex, and CrossRef. Builds with the Zenodo integration also offer Zenodo as an opt-in source of publication records. Select it before starting, or ask Kosmo to include it when proposing a research run.

Use the dedicated Zenodo agent search for datasets and software. A source selection does not grant access to restricted files or guarantee a PDF download. See Zenodo discovery and project publishing for the preview status, citations, and limits.

Depth

Depth sets how many papers to gather and how many search rounds to spend finding them.

Standard is the default and is the right call for a first pass at an unfamiliar area. Advanced options override the individual numbers if you need something the presets do not cover.

What happens while it runs

The progress panel names each stage as it goes:

1. Expanding query. Generating synonyms and MeSH terms, so a paper that uses your field's other word for the thing still turns up.
2. Searching sources. Querying the databases you selected.
3. Fetching papers. Retrieving metadata.
4. Deduplicating. The same paper from four databases collapses to one.
5. Filtering. Ranking by relevance to what you asked.
6. Clustering. Grouping the results into themes.

Adaptive Mode

Standard research runs the pipeline once. Adaptive Mode puts an agent in charge of it: it searches, reads what came back, works out what is missing, and searches again for the gap, up to the iteration budget. As it runs you see its goals, its coverage, and the gaps it has identified.

Adaptive Mode can also reach two databases the manual picker does not offer: DBLP for computer science, and Europe PMC, which covers biomedical literature along with bioRxiv and medRxiv preprints.

Use it when you do not yet know the shape of the literature. Use the standard pipeline when you do.

Advanced options

• Citation depth. How many hops out to follow from a found paper's references.
• Minimum relevance and minimum citations. Floors for what survives filtering.
• Utility model. A fast model used for scoring and classification inside the run, separate from the model doing the reasoning.

What you get out

The results panel gives you the papers with their statistics: source distribution, publications by year, themes, top authors, top venues, and average relevance. From there you can generate a literature review, a multi-section write-up with citations, or a knowledge graph of how the papers relate.

Sessions are durable. You can leave and reopen a run, pause, resume, steer, or cancel it. Reopening attaches to the persisted run instead of silently starting a second metered session. The sidebar lists past runs by status, date, paper count, or name.

Note: Deep research is metered against your plan separately from chat. The panel tells you when you have used the runs in your billing period.

Related

• The research graph
• Bibliography
• Choosing a model