pypi / nv-ingest-client
nv-ingest-client
What changed in nv-ingest-client 26, from its official release notes, upgrade guides and advisories. Also covers nv-ingest-api. Sources checked 2026-10-31.
6 breaking17 new APIs and options1 performance8 behavior and defaults1 tooling4 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initnv-ingest-client 2637 changes
Breaking changes 6
Package installs no longer support macOS Intel x86_64; Apple Silicon macOS, Windows x64, and Linux remain supported.
Helm replaces separate page-elements and table-structure NIM images with the combined object-detection image.
Legacy nv-ingest and compatibility pipeline CLI code paths are removed; use retriever ingest and the graph stage registry.
The minimum Ray version is raised to >=2.56.1 for CVE remediation.
First-class ingestion command and canonical replacement for the retired compatibility pipeline command.
An unacknowledged VectorDB write now fails instead of reporting completed with a positive row count.
New APIs and options 17
Adds an MCP tool for agentic retrieval when agentic service functionality is enabled.
Accepts a validated custom prompt for generic text generation.
Adds documented Markdown, JSON, and shell text inputs, plus inline text ingestion support.
Adds an optional embedding checkpoint for local Hugging Face inference and hosted or self-hosted embedding NIM use.
Adds endpoint wiring in service extraction workers, with distinct hosted and self-hosted HTTP contracts.
Enables agentic retrieval on the service query endpoint when set to true.
First-class query command with NIM URL flags, batch tuning, and LanceDB overwrite/append controls.
Runs an LLM-driven ReAct retrieval loop over the same LanceDB table as one-pass retrieval.
Selects a remote OpenAI-compatible NIM or NVIDIA-hosted endpoint for agentic retrieval.
Supports the Super-49B default LLM path and the Omni VLM-capable answer path.
Is supported across in-process and batch ingest modes.
Configures the worker acknowledgement timeout for VectorDB writes; defaults to 300 seconds.
Provides built-in summarization with either the default prompt or a custom prompt.
Adds a shared one-request-per-row text-generation operator for graph workflows.
Configures text-generation operators with supported model, endpoint, generation, prompt, retry, timeout, and worker settings.
Adds a shared one-request-per-row text-generation abstraction for graph workflows.
Adds service-mode chunking for TXT and HTML inputs.
Performance 1
Runs OCR crop batches across page rows to improve throughput.
Behavior and default changes 8
Each write commits rows independently; index maintenance is serialized and coalesced, and expired index-readiness waits leave committed rows queryable.
Changing a Helm NIM image repository or tag on an existing release cannot patch the modelPuller source; delete the NIMCache and its PVC, then upgrade.
The self-hosted default answer LLM is nvidia/llama-3.3-nemotron-super-49b-v1.5.
Helm defaults to Nemotron OCR v2, using the nvcr.io/nim/nvidia/nemotron-ocr-v2:2.0.1 image.
Helm maps rerank endpoint and model settings and auto-wires them when the rerankqa NIM is enabled, enabling reranking in split topology.
Helm's default VL embedding NIM image is upgraded to 2.3.0.
Helm's default VL reranking NIM image is upgraded to 2.3.0.
Nemotron 3 Nano Omni becomes the default image-captioning model for local and hosted paths; chart-classified PDF regions remain on layout and OCR.
Tooling 1
Development Docker Compose uses the combined object-detection image, with 2.0.0 tags unless NIM_*_TAG is overridden.
Fixes and other changes 4
Fixes local Hugging Face batch audio extraction hanging in interactive terminals when FFmpeg inherited the parent process's standard input.
Reports the package version instead of a stale package-version value, with Helm setting it from the running service image tag.
Agentic retrieval remains opt-in in service mode and requires a remote OpenAI-compatible LLM endpoint.
Inherits from TextGenerationTask.