Skip to navigation
Configure DocsFabModels and Routes

Probe a Model Route in DocsFab

Dev
View as Markdown

Check an exact model and serving route before drafting a composition. The default probe is offline; --live sends model requests.

Prerequisites

Install DocsFab, identify the exact served model ID, and configure the route’s credential before a live probe.

Steps

Use the command and checks below.

Probe One Route

Replace MODEL_ID with the exact model ID that the route serves:

docsfab profile probe --route nvidia-inference-api \
  --model MODEL_ID
docsfab profile probe --route nvidia-inference-api \
  --model MODEL_ID --live --json

The default, offline invocation does not spend. It prints that live inference was not requested, plus a credential-free locator. Add --live to run the request-policy probe and the endpoint compatibility probe. That spend gate matches docsfab doctor --live: no --live means no model request.

--live uses the same probe stack as packages/runtime/scripts/model-matrix/probe-models.mjs. The stack lives in src/profile/probe-core.js, so the installed package ships it. Do not run a second HTTP probe.

Observation fields are the request-policy verdict, the locator, and the endpoint compatibility report. They never include credentials, model reply text, or Hub Metadata payloads. When the provider refuses a request, the report quotes the provider’s own message after DocsFab removes credentials, URLs, and account, function, and request identifiers.

Use an exact Hub or OpenAI model id. Prefix matchers fail closed with a stable error:

profile probe rejects prefix matchers; use an exact Hub or OpenAI model id.

The following OpenAI-compatible chat-completions routes are supported today:

  • nvidia-inference-api
  • nvidia-api-catalog
  • openai-api

Interpret The Verdict

The following table describes each verdict and its draft.

VerdictMeaningDraft
harness-compatibleText arrived in content.Omit requestPolicy when the route marks the policy optional.
reasoning-split-controllablechat_template_kwargs.thinking=false restored content.Include the measured request policy.
reasoning-split-uncontrolledThe split remains.Draft refuses.
unreachableEndpoint, credential, or model problem.Draft refuses.

A proxy refusal is not an unreachable verdict. When a proxy setting or the proxy itself refuses any probe request, profile probe --live fails with the refusal’s outbound-proxy-* code and next action, even when another sample answered. For the codes, refer to Proxy Failures.

The NVIDIA Inference API route marks the installed request policy optional. Harness-compatible models on that route, including OpenAI SKUs served through the hub, omit the Nemotron control. The NVIDIA API Catalog route requires a measured policy, so draft accepts it only for reasoning-split-controllable.

Use these topics for the supporting contracts and next actions.

Refer to Read The Endpoint Compatibility Verdict.

Refer to Authoritative Sources.

Refer to Draft A Composition Record.