NVIDIA / NVIDIA/NeMo-Agent-Toolkit-Examples

New example proposal: 3GPP spec Q&A with clause citations and an abstention gate (telecom RAG)

Open
#32 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
32
Forks
23
PR merge metrics
No merged PRs in 30d

Description

Hi — I'd like to contribute a telecom-domain example and wanted to check the fit before opening the PR (per CONTRIBUTING).

What it demonstrates: a ReAct agent (native tool calling) over a custom retrieval function with two behaviors I haven't seen in the existing examples: (1) domain-aware chunking — 3GPP specs from the public GSMA/3GPP HF dataset are chunked at clause boundaries so every retrieved passage carries a citable reference like [TS 38.331 §5.3.5.3], and the agent cites clauses in its answers; (2) a calibrated abstention gate — below a cosine relevance floor the tool returns an explicit ABSTAIN and the agent tells the user the indexed specs don't cover the question, instead of answering from prior knowledge.

Stack: nvidia/nemotron-3-nano-30b-a3b (agent) + nvidia/nemotron-3-embed-1b (embeddings), both via build.nvidia.com — no local GPU. Corpus is downloaded by the user from the public GSMA dataset; no spec text ships in the repo.

Status: built and tested against nvidia-nat 1.8 — offline pytest suite for the chunker/index/abstention, and both the cited-answer and abstain paths verified end-to-end with nat run. Follows the nat workflow create layout (nat_clausefinder_3gpp module, configs symlink, pyproject with nat.components entry point).

It's a distillation of a larger open-source project of mine (https://github.com/chinthave657/clausefinder — full hybrid retrieval + citation validator + NeMo Guardrails on the same corpus), scoped down to one clean teachable pattern.

Happy to adjust scope or naming — PR is ready to open whenever.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with CONTRIBUTING and the existing nat workflow create layout, then inspect the described nat_clausefinder_3gpp module, configs symlink, and pyproject entry point. Run the offline pytest suite and the cited-answer and abstain paths with nat run; done means the example follows the repository layout and all listed checks pass without shipping spec text.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
ai, search
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
55/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.