What Research can do
Research can discover scholarly papers and general web resources, evaluate selected public sources, and produce durable artifacts with evidence, relevance, recommendations, and limitations.
Discovery
Papers + web
The request can use one or both implemented research surfaces.
Extraction
80k characters
Content is accepted at or below the configured gate; larger bodies are rejected rather than truncated.
Output
One artifact
A completed source yields a structured, source-bound evidence record.
1. Find the right material
Discovery can combine `paper` and `web` resource kinds, a target source count, date and language settings, hostname allowlists/exclusions, and reviewed source-site preferences. Source-site preferences bias where the agent searches; `include_domains` is the hard retrieval constraint.
The agent searches, then explicitly saves positions from the current result list. That selection boundary stops a model from inventing a URL and keeps unsaved search results out of durable state.
Preset or explicit targets up to 1,000 sources.
Bounded search calls with explicit save cycles.
Thread-wide URL identity prevents prior sources from being rediscovered.
2. Read only what can become evidence
Selected public URLs are validated and retrieved one source at a time. A shared page-content cache can avoid repeat retrieval work, while an unavailable or oversized source is recorded as an isolated outcome instead of collapsing the full investigation.
The extractor receives the original question, source metadata, and one bounded normalized source body. It returns typed content tied back to the application-owned source identity, so the model cannot redirect an artifact to a different URL.
Public URL validation before cache or provider use.
Exactly 80,000 characters is accepted; 80,001 is rejected without model extraction.
Provider failures are classified so retryable work can be resumed.
3. Get an artifact you can inspect
| Artifact part | Why it matters |
|---|---|
| Summary | States what the retrieved source says without relying on model memory. |
| Relevance | Connects this particular source to the question that opened the thread. |
| Key claims and evidence | Preserves source-grounded findings that can be reviewed later. |
| Recommendations and limitations | Makes next steps and uncertainty explicit rather than silently smoothing them away. |
| Provenance | Keeps source URL and retrieval status attached to the artifact. |