Skip to content

Add remote NIM request payload capture - #2508

Draft
charlesbluca wants to merge 3 commits into
NVIDIA:mainfrom
charlesbluca:nim-request-payloads
Draft

Add remote NIM request payload capture#2508
charlesbluca wants to merge 3 commits into
NVIDIA:mainfrom
charlesbluca:nim-request-payloads

Conversation

@charlesbluca

@charlesbluca charlesbluca commented Aug 17, 2026

Copy link
Copy Markdown
Collaborator

Description

Adds opt-in persistence of final outbound remote-NIM request payloads for library and service deployments.

The shared capture layer records HTTP JSON requests for extraction, embedding, and reranking, plus Triton gRPC input tensors as compressed NumPy archives. It supports filtering by operation/stage, configurable best-effort or required write behavior, endpoint sanitization, and local or fsspec-compatible storage.

Python users can configure InferenceCaptureConfig on create_ingestor, GraphIngestor, or Retriever. Helm service configuration is available through serviceConfig.inferenceCapture, including VectorDB query-embedding capture. Documentation describes configuration, replay artifacts, and sensitive-data handling.

Checklist

  • I am familiar with the Contributing Guidelines.
  • New or existing tests cover these changes.
  • The documentation is up to date with these changes.

Validation:

  • pytest -q nemo_retriever/tests/test_inference_capture.py nemo_retriever/tests/test_nim_tracing.py nemo_retriever/tests/test_service_query_rerank.py nemo_retriever/tests/test_service_pipeline_spec.py nemo_retriever/tests/test_service_pipeline_tracing.py (117 passed)
  • helm template with capture enabled
  • python -m mkdocs build --strict --config-file mkdocs.yml
  • Live self-hosted-NIM JP20 extraction, query, and replay verification

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant