iapp/tswap-wellness-benchmark

TSWAP — Thai Wellness Benchmark & Evaluation Release

Data and evaluation logs released with the paper "TSWAP: A Multilingual Retrieval-Augmented Thai Wellness Advisor" (Pornthep Ukosaramig, Digital Touch Point Co., Ltd.; Kobkrit Viriyayudhakorn, iApp Technology Co., Ltd.).

TSWAP is a deployed eight-language wellness advisor grounded, via RAG, in a verified knowledge base of Thai traditional medicine (แพทย์แผนไทย / สมุนไพรไทย) and certified wellness providers — live at https://chat.thaismartwellnessai.com. To our knowledge this is the first public benchmark/evaluation release for Thai traditional-medicine and wellness NLP.

Contents

Folder	What it is
benchmark/	The 50-question Thai retrieval golden set (questions + gold document IDs, scored by Recall@K; deployed system: Recall@5 = 0.88). Being finalized from the production repository — see Status below.
qa-logs/	Production QA campaign logs (Aug 2026, tested on the live web client): 236 second-round PASS cases and 56 tracked ODD/FAIL cases, graded PASS/ODD/FAIL against per-case expected behavior across 6 groups (content, colloquial geospatial search, multilingual, context, safety, input robustness).
frontier-probe/	The frontier no-retrieval probe: 71 real questions (probe_set.json), run/grading scripts, and complete per-answer logs for both runs (probe_results_run2 is the one reported in the paper; probe_results_run1_truncated is archived provenance — do not cite its numbers). Model: gemini-2.5-flash, Vertex AI, thinking disabled.
harness/	Evaluation-harness scaffold (production API client, ablation runner, benchmark schema) for the paper's proposed full answer-quality benchmark protocol.

Status

The qa-logs/, frontier-probe/, and harness/ artifacts are complete. The golden-set files in benchmark/ are being copied from the production repository; until then the retrieval numbers are documented in the paper and reproducible against the live index with the included harness.

Privacy

No real user data is included. Apparent personal identifiers inside adversarial QA test inputs (e.g. a citizen-ID / credit-card number in TC_MB_SAFE_20) are synthetic test values (the card number is the canonical Visa test number). UAT results are reported in aggregate only, with informed consent from all participants.

Reproducing the probe

cd frontier-probe
python3 run_frontier_probe.py     # Vertex AI credentials required; resumable
python3 grade_frontier_probe.py   # scripted surface checks -> graded.tsv

License & citation

Data: CC BY 4.0 · Harness code: Apache-2.0 · Code mirror: https://github.com/kobkrit/tswap-wellness-benchmark

This research project is financially supported by the Program Management Unit for Competitiveness (PMUC) under grant number C05F680144.

@misc{tswap2026,
  title  = {TSWAP: A Multilingual Retrieval-Augmented Thai Wellness Advisor},
  author = {Ukosaramig, Pornthep and Viriyayudhakorn, Kobkrit},
  year   = {2026},
  note   = {arXiv preprint; data: https://huggingface.co/datasets/iapp/tswap-wellness-benchmark}
}