These are scaffolding and evaluation resources, plus the open-model shortlist for
the voice and retrieval pipeline. They are US/generic and must never be used
as NZ product knowledge or as the advice voice. They exist to bootstrap structure,
test harnesses, and run cheap self-hostable tasks. Sourced from
/docs/strategy/knowledge-base-data-strategy.md §7, §8.
Public ML datasets (intent scaffolding + eval inspiration only)
The Bitext insurance chatbot dataset gives synthetic intents and QA pairs, useful to bootstrap an intent taxonomy that we then NZ-localise. The Snorkel multi-turn insurance underwriting dataset is good inspiration for the evaluation harness (agentic / tool-use traces), not for content. InsuranceQA variants are generic QA pairs (the original InsuranceQA benchmark is IBM's; the cited deccan-ai entry is a community re-host). None of these are NZ knowledge; use only to scaffold structure and test the pipeline.
Source: Bitext insurance chatbot dataset (https://huggingface.co/datasets/bitext/Bitext-insurance-llm-chatbot-training-dataset) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft Source: Snorkel Multi-Turn-Insurance-Underwriting (https://huggingface.co/datasets/snorkelai/Multi-Turn-Insurance-Underwriting) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft Source: InsuranceQA v2 (https://huggingface.co/datasets/deccan-ai/insuranceQA-v2) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft
RAG embeddings: BGE-M3
The retrieval embedding standard for this build: multilingual (incl. zh/ko), strong on retrieval, used in the §5 pipeline to embed every chunk before storing in pgvector. Lighter alternatives: multilingual-e5, paraphrase-multilingual-MiniLM.
Source: BGE-M3 (https://huggingface.co/BAAI/bge-m3) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft
Voice pipeline (diaspora languages) open models
Voice-out (TTS): CosyVoice 2 / Fun-CosyVoice 3 (Alibaba) is the standout for open
Mandarin + Cantonese + Korean with ~150ms streaming and zero-shot cloning; F5-TTS is
a strong multilingual code-switching alternative; Kokoro-82M is tiny/Apache-licensed
for cheap/edge. Voice-in (STT): SenseVoice-Small (Alibaba) beats Whisper on
Chinese/Cantonese; Qwen3-ASR covers 52 languages with language ID; Whisper large-v3
is the safe multilingual default. Self-host CosyVoice + SenseVoice on AU/NZ infra
for sensitive Chinese voice to keep audio off third-party endpoints (mirrors the
router's sensitive_pii residency gate).
Source: CosyVoice (https://github.com/FunAudioLLM/CosyVoice) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft Source: SenseVoice (https://huggingface.co/FunAudioLLM/SenseVoiceSmall) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft Source: Whisper large-v3 (https://huggingface.co/openai/whisper-large-v3) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft
Domain finance LLMs (cheap self-host tasks only, NOT the advice voice)
DragonLLM's "LLM Pro Finance" collection (a family of finance fine-tunes, not one model), FinGPT, and Bitext-Insurance fine-tunes are options for cheap, self-hostable extraction/classification or fallback, never the NZ advice brain. They are mostly English/Western and classification-oriented, and finance fine-tunes can hallucinate slightly more than their base. The professional broker voice comes from a frontier model + RAG on Your Generation's materials + prompting.
Source: DragonLLM, LLM Pro Finance (https://huggingface.co/DragonLLM) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft Source: FinGPT (https://github.com/AI4Finance-Foundation/FinGPT) · retrieved 2026-06-18 · rights: dataset-public · drives: information · status: draft