docs: OCR v2 polish on multimodal-extraction (Helm/local HF)

[kheiss-uwzoo]

Summary

Docs-only polish for multimodal-extraction.md: scannable Helm / NVCF hosted / local Hugging Face OCR v2 guidance and a clearer chart/captioning note.

Merge after #2295. Hosted-default docs assume CPU actors default to nemotron-ocr-v2 on build.nvidia.com once that eng PR lands.

Does not change releasenotes.md. The 26.05 section correctly records what 26.5.0 shipped (HF v2, Helm v1); #2239 merged after that release.

Changes

• multimodal-extraction.md: Split OCR deployment guidance (Helm / NVCF hosted / local HF); chart pdfium-path callout preserved from main; Omni caption routing note

PR scope check

• Base: main
• Files changed: 1 — docs/docs/extraction/multimodal-extraction.md
• Red flags: none

Test plan

• Spot-check OCR and chart sections in multimodal extraction
• Confirm 26.05 release notes unchanged on main
• Merge feat: switch hosted OCR CPU actor defaults to Nemotron OCR v2 #2295 first, then rebase if needed before merge

[greptile-apps]

Greptile Summary

Documentation-only polish for multimodal-extraction.md that splits the OCR deployment section into three scannable sub-entries (Helm/NIM, NVCF hosted, local Hugging Face) and clarifies that chart-labeled PDF regions bypass the Omni caption stage.

• OCR section restructured: Single paragraph replaced by three bold-headed deployment variants; the new anchor link to prerequisites-support-matrix.md#nemotron-ocr-v2-language-mode resolves correctly to the existing "OCR artifacts (Helm vs local Hugging Face)" section.
• Chart/captioning note tightened: Explicit statement added that chart-labeled regions remain on the layout-and-OCR path and are not routed through the Omni caption stage.
• Merge dependency: The NVCF default-endpoint claim (line 81) describes behavior that lands in PR feat: switch hosted OCR CPU actor defaults to Nemotron OCR v2 #2295; the PR description acknowledges this ordering requirement, but there is no enforcement mechanism preventing an out-of-order merge.

Confidence Score: 4/5

Safe to merge only after PR #2295 lands; merging in isolation publishes the NVCF default-endpoint description before the backing code change exists.

The single changed file is documentation. Every internal anchor and cross-reference validates. One section — the NVCF hosted-default claim — documents behavior that the current codebase does not yet implement; the CPU actor sets the OCR URL to an empty string when ocr_invoke_url is omitted, so remote OCR does not fire. The PR description explicitly states this PR must follow #2295, but nothing in the merge process enforces that ordering.

docs/docs/extraction/multimodal-extraction.md — specifically the NVCF hosted-inference paragraph (line 81) which describes a default that does not exist in the current codebase.

Important Files Changed

Filename	Overview
docs/docs/extraction/multimodal-extraction.md	Documentation polish splitting OCR guidance into Helm/NIM, NVCF hosted, and local HF sections; anchor link to prerequisites-support-matrix.md validates correctly, but the NVCF default claim documents future behavior contingent on PR #2295 merging first.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[OCR / Scanned Document] --> B{Deployment mode?}
    B -->|Helm / NIM| C[Nemotron OCR v2 NIM\nnvcr.io/nim/nvidia/nemotron-ocr-v2:1.4.0]
    B -->|NVCF hosted\nno ocr_invoke_url set| D[build.nvidia.com\nNemotron OCR v2 endpoint\n⚠ requires PR 2295]
    B -->|Local Hugging Face| E[Nemotron OCR v2\nmultilingual mode]
    C --> F[OCR Artifacts]
    D --> F
    E --> F
    G[Chart-labeled PDF region] -->|NOT routed| H[Omni Caption Stage]
    G -->|stays on| I[Layout-and-OCR path]

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A[OCR / Scanned Document] --> B{Deployment mode?}
    B -->|Helm / NIM| C[Nemotron OCR v2 NIM\nnvcr.io/nim/nvidia/nemotron-ocr-v2:1.4.0]
    B -->|NVCF hosted\nno ocr_invoke_url set| D[build.nvidia.com\nNemotron OCR v2 endpoint\n⚠ requires PR 2295]
    B -->|Local Hugging Face| E[Nemotron OCR v2\nmultilingual mode]
    C --> F[OCR Artifacts]
    D --> F
    E --> F
    G[Chart-labeled PDF region] -->|NOT routed| H[Omni Caption Stage]
    G -->|stays on| I[Layout-and-OCR path]

Loading

_{Reviews (10): Last reviewed commit: "docs: add NVCF hosted OCR v2 default to ..." | Re-trigger Greptile}

[charlesbluca]

Should this live in a new release notes section? As-is, this would be changing the description of OCR defaults for 26.5.0, but #2239 didn't land until after this release