Evidence (PostHog, project modelschemas)#
ingest_failed, providerId nvidia, job models-poll: 44 events from 2026-11-16 20:51 UTC to 2026-11-17 07:33 UTC (every poll), error:
fetch failed: https://docs.api.nvidia.com/nim/reference/nvidia-embedding-2b-infer.md → 404 Not Found
Cause (upstream)#
NVIDIA's retrieval reference index (https://docs.api.nvidia.com/nim/reference/retrieval-apis.md, line 24) still lists:
| [nvidia / embed-qa-4](…/nvidia-embed-qa-4) | [Create embedding vector (embed-qa-4)](…/nvidia-embedding-2b-infer) |
The linked page (and nvidia-embed-qa-4, nvidia-embed-qa-4-infer, .md variants) now 404. inferUrlFromCell in src/server/providers/nvidia-openapi.ts follows that link, so the poll fails the fetch every 15 minutes.
Fix#
- Check whether
nvidia/embed-qa-4 is still in GET https://integrate.api.nvidia.com/v1/models. If it is not listed, an index row with no listing model should not be fetched at all (the index is stale, the listing is the authority).
- If it is still listed, the model has no published spec: report it once as a named, located failure (per fail-loud) and mark that row's schema
unavailable, keeping any stored schema — but do not throw for the whole provider, and do not guess another page.
- Add a test with this exact index row and a 404 for its page.
Checked:
nvidia/embed-qa-4IS still listed byGET https://integrate.api.nvidia.com/v1/models(2026-11-17), so take the second branch: the listed model's only spec page is gone.