Search by concept
When the agent goes looking for evidence, it searches by biological meaning — the concepts, entities, and relationships in your question — not by literal word matching. That lets it pull the right papers from the life-science literature even when they describe your idea in completely different terms.
#Beyond keyword matching
A keyword search is a string-matching exercise: a paper only surfaces if it contains the exact words you typed. That fails constantly in biology, where the same idea is written a dozen different ways. A study about the p53 pathway may never use the phrase tumor suppressor; a paper on programmed cell death may never say apoptosis. Keyword search treats those as unrelated and quietly drops the most relevant work.
The agent instead builds a semantic representation of your question and matches it against the meaning of each paper's title, abstract, and metadata. Two documents that discuss the same biology score as similar even when they share almost no vocabulary. In practice this means the agent finds papers that discuss the same concept, regardless of which terminology the authors happened to choose.
| You ask about | Keyword search misses | Concept search still finds |
|---|---|---|
| tumor suppressor | papers that only say "p53", "RB1", or "PTEN" | the specific genes that act as tumor suppressors |
| programmed cell death | papers phrased only as "apoptosis" or "pyroptosis" | the mechanistic literature under either name |
| T-cell activation | papers written around "lymphocyte co-stimulation" | the shared immunological concept |
#Understanding synonyms and relationships
Because matching happens at the level of meaning, the agent absorbs the messy naming conventions of biology automatically. You do not have to enumerate every alias for a gene, pathway, or disease — the search behaves as if you had.
- Synonyms and aliases — gene symbols, protein names, and their historical or alternative labels are treated as the same entity.
- Abbreviations — expanded forms and their acronyms (e.g. epidermal growth factor receptor and
EGFR) are linked. - Relationships between concepts — the search reflects how entities relate, so a query about a pathway can surface work on its upstream regulators, downstream effectors, and associated diseases.
This is what makes the agent good at surfacing non-obvious connections: because it reasons over relationships rather than exact strings, it can bridge a query in one subfield to relevant evidence described in the vocabulary of another.
#How relevance is scored
Every candidate paper is ranked by how closely its meaning aligns with your question, not by how many times your words appear in it. Neural language models trained on a large corpus of scientific literature produce that relevance score, so a conceptually perfect match with different wording can — and often does — outrank a paper that merely repeats your exact phrase.
- Interpretation — your natural-language query is parsed into a semantic representation of the concepts and relationships it describes.
- Conceptual matching — that representation is compared against each paper's title, abstract, and metadata to measure genuine topical overlap.
- Ranking — results are ordered by relevance so the most conceptually on-point papers rise to the top, even across differing terminology.
- Refinement — the agent narrows the set with constraints such as publication window, open-access availability, and citation signals to keep the evidence base focused and defensible.
The goal is not to return the largest possible pile of hits. The agent uses relevance ranking to distill millions of publications into a core, statistically-backed evidence set — the handful of papers most worth reading in full and cross-checking.
#What the search covers
The agent searches across an index of 40M+ papers spanning the life-science literature, enriched with citation graphs and structured metadata. That breadth lets it reach beyond the obvious top hits to foundational and adjacent work that keyword search would never connect.
- Titles, abstracts, and metadata are matched semantically to find conceptually relevant papers.
- Citation relationships help the agent trace a topic back to the foundational studies it rests on and outward to the work that builds on it.
- Recency and impact signals — publication dates and citation counts — are available to weight or filter results toward recent or influential work.
Semantic search is where the agent decides what to read. From there it moves on to reading those papers in full, extracting findings, and screening them for integrity — so the quality of this concept-level retrieval shapes everything downstream.