Update deprecating LLM nemotron-3-nano-30b-a3b → nemotron-3.5-lightning-30b-a3b (+ thinking:false for strict-JSON) - #72
Merged
david-duu merged 1 commit intoAug 19, 2026
Conversation
…ing-30b-a3b The hosted nemotron-3-nano-30b-a3b endpoint is deprecating. Swap the summarization, chat, and notification LLMs to the successor nemotron-3.5-lightning-30b-a3b, and set extra_body.chat_template_kwargs.thinking: false on each (the successor is a reasoning model whose chain-of-thought otherwise floods content and breaks the strict-JSON contract / graph extraction). Validated end-to-end on an L4: grounded graph-RAG probe passes with a passing negative control. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
karthikt-nvidia
marked this pull request as ready for review
August 19, 2026 00:38
david-duu
merged commit Aug 19, 2026
f83fe2f
into
NVIDIA:dev/streaming-data-to-rag
1 of 2 checks passed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What & why
nvidia/nemotron-3-nano-30b-a3b— the hosted LLM used by the summarization, chat, and notification paths inconfig/config.yaml— is being deprecated as a hosted endpoint. This updates all three references to its successornvidia/nemotron-3.5-lightning-30b-a3b, which the hosted gateway currently serves (verified live:/v1/modelslists it and a completion returns HTTP 200 echoing the model id). No self-hosted container is needed.The non-obvious part: disable "thinking"
nemotron-3.5-lightningis a reasoning model — on long / agentic / RAG prompts it emits chain-of-thought into the responsecontent. With this blueprint's strict-JSON contract andmax_tokens: 2048, that overflows the budget, the JSON is truncated → parse failure → the pipeline falls back to a canned "cannot answer," and graph-extraction during ingest fails intermittently.Fix:
extra_body.chat_template_kwargs.thinking: falseon each LLM block. With it,contentis clean JSON and reasoning is separated (reasoning_content: null); verified empirically against the endpoint. This resolved both the probe failure and the ingest-side graph-extraction failures.Validated on real hardware
Deployed the full blueprint on an NVIDIA L4 (23 GB) — 14 containers (ASR + embedding + reranker NIMs + the graph-RAG stack) — ingested live FM-radio transcripts, then ran the graph-RAG probe:
fm-radio-ch0/ch2with timestamps from the run; 3 clean JSON parses, 0 failures.Notes for maintainers
ctx_ragimage (COPY config/), so this needs an image rebuild to take effect — editing config alone silently keeps the old model.max_tokens: 2048is a little tight for a reasoning model even with thinking off (one truncated-JSON retry self-recovered) — worth raising for headroom.Prepared and validated by an automated blueprint-refresh CI (deploy → swap → re-validate on a GPU). Please review before merging.