Add clausefinder_3gpp example: clause-cited 3GPP spec Q&A with an abstention gate

[chinthave657]

Description

A telecom-domain RAG example over the public GSMA/3GPP corpus, demonstrating two patterns not covered by existing examples:

1. Clause-aware domain chunking — specs are chunked at clause-heading boundaries with breadcrumb trails, 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. Calibrated abstention— below a cosine relevance floor the retrieval tool returns an explicit ABSTAIN and the agent tells the user the indexed specs don't cover the question, rather than answering from prior knowledge. The default threshold is calibrated against the built index (in-corpus questions score ≥0.50, out-of-domain ≤0.36).

All-NVIDIA endpoints: nvidia/nemotron-3-nano-30b-a3b (agent, native tool calling) + nvidia/nemotron-3-embed-1b (embeddings) via build.nvidia.com — no local GPU. No spec text ships in the repo; users download the corpus themselves from the public GSMA dataset.

Distilled from my larger open-source project ClauseFinder (hybrid retrieval + deterministic citation validation + NeMo Guardrails on the same corpus).

Testing

• Offline pytest suite (7 tests): chunker (incl. letter-suffixed clauses like 5.3.5.13b), breadcrumbs, index round-trip + corruption detection, cosine search, citation formatting, abstention message.
• End-to-end with nat run against nvidia-nat 1.8: in-corpus question returns a clause-cited answer; out-of-corpus question ("Bluetooth Class 1 power") abstains. Real outputs are in the example README.
• ruff + yapf clean against the repo config; SPDX headers on all files.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.