docs(nemo-retriever): 26.07 RC readiness corrections for NeMo Retriever Library

[kheiss-uwzoo]

Description

Fix customer-facing RC documentation issues:

• Live RAG imports (NVBugs 6492587): Import LiteLLMClient and LLMJudge from the supported public API, nemo_retriever.models.llm. The documented nemo_retriever.llm module is not included in current nemo-retriever wheels and fails with ModuleNotFoundError.
• Local GPU install: Align the CUDA 13 override command with the frozen [local] dependency metadata in nemo_retriever/pyproject.toml: torch==2.11.0 and torchvision==0.26.0.
• Nemotron Parse installer wording: Make uv pip install nemo-retriever[nemotron-parse] the primary command in the extraction overview and support matrix, consistent with the UV-first README, while explicitly identifying plain pip install as equivalent for users who do not use UV.
• Remote NIM platform support: Clarify that supported local GPU inference requires Linux and CUDA, while the base package for remote NIM inference can also be installed on Windows x64 and macOS x64/ARM64.
• Statement-style headings: Rename non-FAQ question-style headings to statement titles per NVIDIA Template Library guidance (overview H1, root index H1, LanceDB section, MkDocs nav). FAQ headings stay as questions. Existing heading anchors are preserved.

The branch was rebased onto current main. Earlier NVAIE clarification and obsolete QA-link cleanup commits were dropped because equivalent updates are already present on main. Heading renames were folded in from #2409 (now closed).

Validation

• DORI code-block validation: passed (13 README blocks checked, no issues)
• DORI documentation/link validation: passed (no broken links)
• pre-commit run for changed documentation files: passed
• NRL extraction page-role and link-CTA leakage checks: passed
• README Torch/Torchvision pins match nemo_retriever/pyproject.toml
• Full strict MkDocs build is blocked by the pre-existing missing THIRD_PARTY_LICENSES.md snippet referenced by docs/docs/license.md; the failure is unrelated to this diff

Checklist

• I am familiar with the Contributing Guidelines.
• The documentation is up to date with these changes.
• Changed-file validation passes.

[greptile-apps]

Greptile Summary

This is a documentation-only PR that fixes several customer-facing issues in the NeMo Retriever Library 26.07 RC docs. The changes are accurate and internally consistent based on verification against the actual module structure and pyproject.toml.

• Import path correction: LiteLLMClient and LLMJudge are updated to import from nemo_retriever.models.llm — verified correct against the package's __init__.py, which lazy-loads both from nemo_retriever.models.llm.clients.*.
• Torch/torchvision pins: The README now shows torch==2.11.0 and torchvision==0.26.0 with index https://download.pytorch.org/whl/cu130, matching the [local] extra in nemo_retriever/pyproject.toml exactly.
• Heading renames and platform scope: Question-style H1/H2 headings are converted to statement titles with explicit anchor IDs preserved; platform support (Linux-only for local GPU, Windows/macOS for remote NIM) is clarified consistently across six docs pages and the README.

Confidence Score: 4/5

Safe to merge for documentation purposes; however the companion code fix for the same broken import is missing from this PR and would leave the tools/evaluation re-export path still broken for users of that module.

All documentation changes are technically accurate — import paths, torch/torchvision pins, CUDA index URL, and heading anchors are verified against the actual source tree. The one gap is that the lazy-import table in tools/evaluation/init.py still routes LiteLLMClient and LLMJudge through the non-existent nemo_retriever.llm.clients module, meaning the fix is incomplete: README users get the right path but programmatic consumers of the evaluation module still hit ModuleNotFoundError.

Files Needing Attention: nemo_retriever/src/nemo_retriever/tools/evaluation/init.py — lazy-import paths for LiteLLMClient and LLMJudge still reference the non-existent nemo_retriever.llm.clients module.

Important Files Changed

Filename	Overview
nemo_retriever/README.md	Fixes LiteLLMClient/LLMJudge import paths, aligns torch/torchvision CUDA 13 override command with pyproject.toml pins, and restructures intro to separate local GPU vs. remote NIM paths. All technical content verified correct.
docs/docs/extraction/overview.md	H1 renamed from question style to statement style with anchor preserved; nemotron-parse note updated to show uv pip install as primary with pip install as equivalent alternative.
docs/docs/extraction/prerequisites-support-matrix.md	Adds platform summary banner, rewrites Linux/CUDA requirements to distinguish local GPU vs. remote NIM paths, converts some MkDocs admonitions to blockquotes (one admonition at line 116 remains in !!! style, creating a minor formatting inconsistency).
docs/docs/extraction/vdbs.md	"Why LanceDB?" section heading and TOC entry renamed to "LanceDB Overview"; the explicit anchor id #why-lancedb is preserved on the heading, so the TOC link continues to resolve correctly.
docs/docs/index.md	Root index H1 renamed to statement style; explicit anchor #what-is-nvidia-nemo-retriever added to preserve any existing inbound links.
docs/mkdocs.yml	Nav label for extraction/overview.md updated to "NeMo Retriever Library Overview" to match the renamed H1; no other nav or plugin configuration touched.
docs/docs/extraction/multimodal-extraction.md	Four cross-reference link labels updated from "What is NeMo Retriever Library?" to "NeMo Retriever Library Overview"; no content changes.
docs/docs/extraction/audio-video.md	Two cross-reference link labels updated to match renamed overview heading; no content changes.
docs/docs/extraction/getting-started-about.md	Step 2 in the getting-started flow now notes that remote NIM workflows support Windows x64 and macOS; final link updated to new overview heading text.
docs/docs/extraction/troubleshoot.md	One cross-reference link label updated; troubleshoot pip install example intentionally left as plain pip install since the PR treats it as equivalent to uv pip install.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[User installs nemo-retriever] --> B{Inference path?}
    B --> C["Local GPU (Linux only)\nuv pip install nemo-retriever[local]"]
    B --> D["Remote NIM\nuv pip install nemo-retriever\n(Linux / Windows x64 / macOS x64/ARM64)"]
    C --> E["Override torch/torchvision\ntorch==2.11.0, torchvision==0.26.0\n-i https://download.pytorch.org/whl/cu130"]
    C --> F["Optional: Nemotron Parse\nuv pip install nemo-retriever[nemotron-parse]"]
    D --> G["from nemo_retriever.models.llm import LiteLLMClient, LLMJudge"]
    E --> H[Run GPU-accelerated ingestion pipeline]
    F --> H

Loading

_{Reviews (10): Last reviewed commit: "docs(extraction): use GitHub-friendly bl..." | Re-trigger Greptile}