docs(nemo-retriever): document Nemotron Parse hosted vs self-hosted contract selection - #2397
Conversation
… NIM API divergence
The hosted NVIDIA Build endpoint (model nvidia/nemotron-parse) currently
expects a different request contract than the self-hosted nemotron-parse-v1.2
NIM the library targets, so extract_method="nemotron_parse" against the hosted
Build endpoint can fail with HTTP 400 ("model does not support text input").
Document the divergence and workaround (self-hosted Helm NIM, local HF, or
default pdfium) in the support matrix NVCF endpoints section, and add a matching
troubleshooting entry for the error symptom.
Greptile SummaryThis docs-only PR documents the hosted NVIDIA Build vs self-hosted Nemotron Parse request-contract selection that landed in #2398, and globally renames
|
| Filename | Overview |
|---|---|
| docs/docs/extraction/prerequisites-support-matrix.md | Renames extract_method to method in table/notes, and adds a new hosted Build vs self-hosted NIM callout; one sentence uses undefined "endpoint lists" terminology that may confuse readers. |
| docs/docs/extraction/troubleshoot.md | Renames extract_method to method in one existing entry, and adds a new Nemotron Parse model/endpoint mismatch troubleshooting section. |
| docs/docs/extraction/customize-extend.md | Single-line rename of extract_method to method; consistent with the rest of the PR. |
| docs/docs/extraction/faq.md | Single-line rename of extract_method to method; consistent. |
| docs/docs/extraction/multimodal-extraction.md | Renames extract_method to method in three locations; all consistent. |
| docs/docs/extraction/overview.md | Renames extract_method to method in two locations; consistent. |
| nemo_retriever/README.md | Single-line rename of extract_method to method; consistent. |
| nemo_retriever/helm/README.md | Single-line rename of extract_method to method in the Helm value description; consistent. |
Flowchart
%%{init: {'theme': 'neutral'}}%%
flowchart TD
A[User sets method=nemotron_parse] --> B{nemotron_parse_invoke_url set?}
B -- No --> C[Use default hosted Build URL]
B -- Yes --> D{Points to Build endpoint?}
D -- Yes --> C
D -- No --> E[Use self-hosted chat endpoint]
C --> F{nemotron_parse_model set?}
F -- No --> G[Auto-select nvidia/nemotron-parse]
F -- Yes --> H{Model matches Build contract?}
H -- Yes --> I[Use image-only tool-call contract]
H -- No versioned v1.2 model --> J[HTTP 400 Content cannot be a plain string]
G --> I
E --> K{nemotron_parse_model set?}
K -- No --> L[Auto-select nvidia/nemotron-parse-v1.2]
K -- Yes --> M[Use specified model]
L --> N[Use tagged text-prompt contract]
M --> N
Reviews (4): Last reviewed commit: "docs(nemo-retriever): standardize extrac..." | Re-trigger Greptile
|
Holding this as draft pending Charles's engineering compat changes so we handle each |
charlesbluca
left a comment
There was a problem hiding this comment.
Thanks for holding this until #2398 landed. The hosted and self-hosted deployments still use different request contracts, but #2398 added library-side compatibility, making hosted Build a supported path. The inline comments identify the places that still describe the old failure mode, along with one invalid public parameter name.
The PR title and description also need to be refreshed: they currently state that hosted extraction fails and that dual-API support remains future work. They should instead describe the supported contract-selection behavior and any configuration guidance retained by this PR.
Once the docs and PR metadata reflect #2398, I did not find another content blocker in this diff; the anchors and cross-links are consistent.
…ontract selection After NVIDIA#2398, the library supports both NVIDIA Build and self-hosted nemotron-parse contracts. Refresh the support-matrix note and troubleshooting entry to describe automatic contract selection and model/endpoint mismatch guidance instead of the pre-NVIDIA#2398 failure mode. Use method="nemotron_parse" for the public ExtractParams parameter.
Normalize remaining extract_method="nemotron_parse" references to method="nemotron_parse" in the support matrix and troubleshoot pages. ExtractParams exposes method; .extract() rejects extract_method.
|
Thanks for cleaning up the touched pages, and apologies—I was not clear that I meant the remaining Nemotron Parse/general extraction-selector references across the customer-facing docs as well. These can be updated directly from
One caveat: |
…r docs Replace remaining Nemotron Parse and general extraction-selector extract_method references with method across the touched customer-facing pages. Leave extract_method="audio" in the package README unchanged; audio uses .extract_audio() rather than .extract(method=...).
Summary
Documents Nemotron Parse hosted Build vs self-hosted request-contract selection after library-side compatibility landed in #2398.
Hosted NVIDIA Build and self-hosted
nemotron-parse-v1.2still use different request contracts, but the library now selects the matching contract automatically. Hosted Build is a supported path.Changes
prerequisites-support-matrix.md— Default NVCF endpoints: annotate thenemotron-parserow and add a configuration note covering automatic model/contract selection, when to setnemotron_parse_invoke_url/nemotron_parse_model, and the mixed-endpoint rule. Uses publicmethod="nemotron_parse"(notextract_method).troubleshoot.md— model/endpoint mismatch entry for the residual HTTP 400 / text-input failure when an incompatible tagged orv1.2model is sent to Build; resolution is omitnemotron_parse_modelor setnvidia/nemotron-parse.Scope
Docs-only. Runtime contract selection is already in #2398.
Self-review
method=for public API guidance added by this PRcheck-nrl-doc-leakage.ps1)