align workflow snippets with create_ingestor API

[kheiss-uwzoo]

Summary

• Migrate extraction examples from Ingestor() to create_ingestor(run_mode="batch") and fix broken Python fences in audio-video.md.
• Show ASRParams.audio_endpoints in the Helm Parakeet example so batch graph ingest reaches the in-cluster NIM (audio:50051); note that endpoint auto-wiring applies to the retriever service, not graph ingest (Greptile).
• Correct custom-metadata.md vdb_upload parameters (vdb_kwargs, sidecar meta triplet) and document remote lancedb_uri for service mode.
• Update GraphIngestor.vdb_upload guidance and document ingest() return types (ray.data.Dataset in batch; pandas.DataFrame in inprocess); rename the workflow snippet variable to result (Greptile).
• Add Python API guide cross-links, update Helm anchor, and apply NVIDIA style fixes (refer to, through, such as).

Test plan

• python skills/scripts/validate_code_blocks.py docs/docs/extraction/ (code fences parse)
• Spot-check rendered pages: audio-video, custom-metadata, embedding, workflow-document-ingestion
• Confirm internal links resolve in docs build

[greptile-apps]

Greptile Summary

This PR updates six extraction documentation pages to align code snippets with the create_ingestor(run_mode="batch") API, replacing the legacy Ingestor() constructor throughout. It also corrects vdb_upload parameter names in custom-metadata.md, adds the missing ASRParams.audio_endpoints wiring in the Helm Parakeet example, fixes a broken Python fence in audio-video.md, and applies NVIDIA style guide fixes (refer to, through, such as).

• API migration: all Ingestor() calls across embedding.md, audio-video.md, and workflow-document-ingestion.md are replaced with create_ingestor(run_mode="batch"), with from nemo_retriever import create_ingestor imports added where missing.
• custom-metadata.md corrections: vdb_upload now uses vdb_kwargs={"lancedb_uri": ..., "table_name": ...} with the sidecar meta triplet (meta_dataframe, meta_source_field, meta_fields), and documents that service mode requires a remote lancedb_uri.
• workflow-document-ingestion.md: renames chunks → result and corrects the return-type comment to distinguish ray.data.Dataset (batch) from pandas.DataFrame (inprocess).

Confidence Score: 5/5

Documentation-only changes that correct API names, parameter signatures, and return-type comments — no production code paths are modified.

All code examples are syntactically correct and consistent with the GraphIngestor implementation (vdb_upload is confirmed implemented). The only finding is a single leftover Ingestor reference in vdbs.md that is inconsistent with the updated line in the same file, but it does not affect runtime behavior.

docs/docs/extraction/vdbs.md — line 80 still names the old Ingestor class while line 129 was updated to GraphIngestor.vdb_upload.

Important Files Changed

Filename	Overview
docs/docs/extraction/audio-video.md	Migrates both Helm and hosted-inference snippets from Ingestor() to create_ingestor(run_mode="batch"); adds missing audio_endpoints for in-cluster Parakeet; fixes broken Python fence; adds results = ingestor.ingest() call; applies NVIDIA style changes.
docs/docs/extraction/custom-metadata.md	Corrects vdb_upload parameter names to vdb_kwargs, meta_dataframe, meta_source_field, and meta_fields; adds guidance on remote lancedb_uri for service mode; simplifies the On this page TOC to match the consolidated section structure.
docs/docs/extraction/embedding.md	Replaces all three Ingestor() calls with create_ingestor(run_mode="batch") and adds missing imports; minor style and punctuation fixes; removes duplicate link from related topics.
docs/docs/extraction/workflow-document-ingestion.md	Renames chunks variable to result; corrects the return-type comment to distinguish ray.data.Dataset (batch) from pandas.DataFrame (inprocess); updates description of GraphIngestor.vdb_upload to reflect that it is now implemented.
docs/docs/extraction/vdbs.md	Updates VDB upload wording to use create_ingestor(...) / GraphIngestor.vdb_upload; adds Python API guide cross-link; one sentence at line 80 still names the old Ingestor class while line 129 in the same file now correctly says GraphIngestor.vdb_upload.
docs/docs/extraction/nimclient.md	Adds a single cross-reference sentence pointing readers to the Python API guide for ingest and pipeline APIs used in NimClient UDFs.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["create_ingestor(run_mode=...)"] --> B["GraphIngestor (batch / inprocess)"]
    A --> C["ServiceIngestor (service)"]
    B --> D[".files(...)"]
    D --> E[".extract(...)"]
    E --> F[".embed(...)"]
    F --> G[".vdb_upload(...) now documented"]
    G --> H[".ingest() → ray.data.Dataset (batch) or pandas.DataFrame (inprocess)"]
    C --> D2[".files(...)"]
    D2 --> E2[".extract(...)"]
    E2 --> F2[".embed(...)"]
    F2 --> G2[".vdb_upload(vdb_kwargs={lancedb_uri, table_name}, meta_*)"]
    G2 --> H2[".ingest_async().result()"]

Loading

Prompt To Fix All With AI

Fix the following 1 code review issue. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 1
docs/docs/extraction/vdbs.md:80
This sentence still refers to the old `Ingestor` class while line 129 of the same file (updated by this PR) uses `GraphIngestor.vdb_upload`. The stale name creates a contradiction inside one document, which could confuse readers following the migration guidance.

```suggestion
When using `GraphIngestor.vdb_upload`, pass `vdb_op="lancedb"` or a `LanceDB` instance so uploads target LanceDB. If you omit `vdb_op`, the library still defaults the string argument to `"milvus"` for backward compatibility, which is not the LanceDB operator—always pass `vdb_op="lancedb"` when you intend LanceDB.
```

_{Reviews (3): Last reviewed commit: "Merge branch 'main' into kheiss/snippets" | Re-trigger Greptile}