Skip to content

Releases

v0.3.5 — Portfolio readiness

Choose a tag to compare

@kuotunyu kuotunyu released this 30 Aug 09:54

v0.3.5 — Portfolio readiness

Highlights

  • Reworked the private Streamlit demo into a reviewer-first BYOK journey with clear capability, privacy, cost, citation, and refusal boundaries.
  • Added a ten-case deterministic offline portfolio regression. All six answerable source contracts and all ten routing/retrieval-stage contracts pass with zero provider calls.
  • Added a content-free article-level freshness artifact for 15 labor instruments and 884 non-deleted articles. The manual audit reports named law/source and article-label changes without publishing legal text.
  • Preserved the v0.3.4 Qdrant candidate pair and the existing retrieval configuration: structure-aware/fixed collections, hybrid BM25+dense retrieval, reranker, RRF k=60, and threshold 0.03.

Evidence boundary

The 40-question formal benchmark, 60-case reliability suite, and archived provider judgments are unchanged. The compact portfolio regression demonstrates source rank, deterministic routes, retrieval-stage refusal, and correct passage to the later citation-completeness boundary; it does not execute or score generated prose. The official artifact is content-free and calls no Gemini, OpenAI, or paid hardware.

The article snapshot contains only law names, article labels, and SHA-256 fingerprints. It contains no legal text, chapters, URLs, credentials, endpoints, provider payloads, or local paths. Freshness remains an attended manual operation: no scheduler, heartbeat, cron job, or automatic writer exists.

Private demo and cost posture

  • Hugging Face Space remains private, on one free cpu-basic replica, with no persistent paid storage or accelerator.
  • Visitors provide their own Gemini or OpenAI API key. The Space has no owner LLM key and cross-provider fallback is disabled.
  • Qdrant remains Free Tier. Runtime access is collection-scoped and read-only; temporary writer and transition keys are revoked after attended maintenance.
  • The private Space deployment receipt is bound before the immutable tag is created. The public repository does not disclose its URL, endpoint, key suffixes, session tokens, or account identity.

Known limitations

  • The system covers the audited 15-instrument corpus; it is not a general legal database or legal advice service.
  • The 0.03 reranker threshold is calibrated evidence, not a universal answerability classifier. The reliability suite retains one measured direct false refusal among 40 answerable stress cases.
  • Two semantically near unanswerable portfolio cases correctly reach the later LLM/citation-completeness boundary; the offline artifact does not claim to have judged a final model refusal.
  • The full corpus, model weights, private indexes, raw provider runs, generated answers, and judge reasons are outside the public source release.

Verification and rollback

Run uv run python scripts/verify_release.py to recompute the public evidence and publication boundary without loading models, contacting providers, or starting Qdrant. A failed private deployment stays private and is rolled back by restoring the previous Space revision and collection base; old collections are not deleted automatically. No paid tier is enabled as a fallback.

v0.3.4 - Wage Arrears Retrieval Hardening

Choose a tag to compare

@kuotunyu kuotunyu released this 29 Aug 21:25
3ec5ade

Highlights

  • Adds deterministic Traditional-Chinese wage-arrears query expansion for retrieval while preserving the original user question for answer generation.
  • Covers ten positive and ten collision regression routes, including retirement, pension, insurance, attendance, benefits, overtime and severance wording collisions.
  • Keeps the audited 15-law, 884-article corpus snapshot and existing retrieval thresholds unchanged.
  • Updates the public release manifest, design evidence, bilingual documentation and regression verification for package version 0.3.4.

Verification

  • Targeted wage-arrears regression: 20 of 20 routes passed; positive Hit at 1 and Hit at 5 were both 10 of 10.
  • Release verifier: pass with exact public allowlist and zero privacy findings.
  • No provider calls are required by the release verification path.

Immutable lineage and post-tag note

The existing annotated v0.3.4 tag remains immutable at commit 3ec5ade and was not moved or recreated.

After the tag was prepared, create-only blue-green Qdrant maintenance hardening was merged to main through PR #12 at main commit cd4aa1c. That post-tag work adds dry-run lineage checks, fixed 400/80 and 481/884/1024 shape contracts, local-only model loading, redacted create-new receipts, and manual Free-tier cutover/rollback guidance. It is intentionally documented as post-tag main evidence and is not claimed as part of the immutable v0.3.4 source archive.

PR #12: #12

v0.3.2 — Provider Evidence

Choose a tag to compare

@kuotunyu kuotunyu released this 29 Aug 06:47

Publishes the completed privacy-safe provider safety cross-check for Gemini gemini-3.5-flash-lite and OpenAI gpt-5.6-luna: 5 completed requests per provider, citation success 1.0 for both, refusal accuracy 0.8 for Gemini and 1.0 for OpenAI, and total observed cost US$0.0049034. This release also hardens the release verifier against sensitive-value reflection, ineligible or nondeterministic question substitution, malformed provider metadata, date drift, and duplicated documentation-claim drift. This is a bounded operational cross-check, not formal model-quality evidence; the formal evaluation baseline remains v0.1.0. Raw provider payloads, prompts, answers, credentials, private paths, and personal data are not published.

v0.3.1 — Reliability and Provenance

Choose a tag to compare

@kuotunyu kuotunyu released this 28 Aug 22:23

v0.3.1

Reliability and provenance release for the Traditional Chinese Taiwan labor-law Hybrid RAG demo.

Highlights

  • complete 15-law / 884-article corpus snapshot contract and provenance audit
  • 60-question reliability stress set plus independently recomputed 40-row formal guard
  • provider budget preflight and a strict US$5-per-provider evaluation cap
  • immutable embedding and reranker model snapshots pinned to full Hugging Face commit SHAs
  • Transformers 5.16.1 with dependency auditing free of known published vulnerabilities
  • public Hugging Face Space remains visitor-BYOK; owner Gemini/OpenAI keys are not installed

Verified

  • GitHub tag CI passed all release gates
  • 438 local tests passed on the merged main tree
  • release verifier status: pass; public/privacy scan issues: 0
  • free Hugging Face CPU Basic deployment is running

Paid Gemini/OpenAI cross-provider evaluation remains explicitly marked pending_credentials until dedicated project keys are provided. No provider calls were made for this release.

v0.1.0 - Source-only portfolio release

Choose a tag to compare

@kuotunyu kuotunyu released this 24 Aug 05:36

Source-only portfolio release

This release freezes the filtered public snapshot of the Traditional-Chinese hybrid RAG portfolio project.

Evidence and scope boundaries

  • The reported evaluation is limited to a fixed 40-question set: 30 answerable and 10 unanswerable questions. It is not a universal benchmark.
  • Retrieval metrics and refusal counts are independently recomputable from committed privacy-reduced traces. Faithfulness and relevancy are re-aggregations of archived provider evidence and are not independently reproducible without the excluded answers, judge reasons, and provider responses.
  • One answerable, out-of-set colloquial narrative was observed to be falsely refused by the reranker threshold. This is a documented failure boundary, not an estimated population error rate.
  • The repository includes no complete generated answers, judge reasons, provider responses, private evaluation runs, prompts, full indexes, credentials, or full corpus.
  • No claim is made that the system can answer all Taiwan labor-law questions or all conversational phrasings.
  • This portfolio artifact is not legal advice and is not presented as a production legal service.
  • Only two attributed sample JSON files are redistributed under Taiwan's Open Government Data License (OGDL) terms. The MIT license covers the software, not those OGDL samples.

GitHub's automatically generated source archives are the only release downloads. No wheel, source distribution, corpus, model, index, container image, or provider artifact is attached.