For the complete documentation index, see llms.txt. This page is also available as Markdown.
Depth
Depth is how much effort to spend on a request. It is the only quality dial, and it never names a model — internal models can change without breaking your integration.
| Value | Use when | Papers considered | Latency | Cost |
|---|---|---|---|---|
standard | interactive UI, short questions | fewer | fastest | lowest |
high_quality | technical questions needing careful reading | more | moderate | higher |
deep_review | agentic synthesis across the literature | most | minutes | highest |
| Value | max_papers ceiling | Latency p50 / p95 | Search credits |
|---|---|---|---|
standard | 20 | 6 s / 15 s | 5 |
high_quality | 50 | 20 s / 45 s | 15 |
deep_review | 200 | 90 s / 240 s | 60 |
max_papers above a tier's ceiling is clamped, not rejected. deep_review is available on searches
and chat messages, not on topic searches or extractions.
Choosing
Default to standard and escalate on demand — most user questions do not need a deep review, and
depth is the largest single driver of both cost and latency. For deep_review, stream the result or
use a webhook; do not block a request thread on it.
Related
Last updated