v0.1.2 — Engineering Hardening

v0.1.2 — Engineering Hardening

Repository-level hardening of the completed Version-Aware RAG MVP:

• PEP 639 licensing: License-Expression: MIT, explicit root license inclusion,
  and setuptools ≥77 remove the former license metadata deprecation warnings.
• Canonical current-system architecture and a navigational documentation index,
  clearly separated from preserved historical design/experiment records.
• Seven frozen-evaluation constraints verified against committed neural-run
  environment records, without changing normal dependency ranges.
• Official GitHub Actions pinned to full commits; independent wheel/sdist
  clean-install CI acceptance with no model downloads, inference or API credentials.

No retrieval/generation algorithm changes. All application Python files,
benchmark/gold/review material, frozen runs/results, M3.1 contexts/outputs and
packaged runtime resources remain byte-identical to v0.1.1. No held-out retrieval
or paid generation was rerun. See the accompanying preservation report.

python -m pip install "version-aware-retriever[dense] @ https://github.com/meiming2019-web/version-aware-retriever/releases/download/v0.1.2/version_aware_retriever-0.1.2-py3-none-any.whl"

Ordinary answering still requires the unchanged external pinned dense-model cache.
Import/help/resource checks do not. Assets: wheel, sdist, SHA256SUMS and these notes.
The root MIT License and upstream notices are retained. v0.1.0 and v0.1.1 tags,
releases and historical notes remain untouched. Existing benchmark limitations,
negative results and pending corpus applicability review remain unchanged.

v0.1.1 — Licensing, CI, and Portfolio Polish

v0.1.1 — Licensing, CI, and Portfolio Polish

Final non-algorithmic polish release of the Version-Aware RAG MVP.

• Adds the MIT License for the project's original code and consistent package
  license metadata; upstream Pydantic notices remain unchanged.
• Adds credential-free GitHub Actions CI on Python 3.12 for tests, lint, format,
  strict typing, and offline acquisition/extraction verification.
• Improves the portfolio resume bullets, repository description/topics and CI badge.
• Updates package version and current installation links to 0.1.1.

No retrieval or generation algorithm, model, prompt, schema, setting, label or
corpus change. All frozen v0.1.0 experimental results remain unchanged: held-out
RRF Recall@5 .8111, MRR .8333, and three qualitative generation checks
with 3/3 decision matches and 3/3 reviewed-valid citations, not answer accuracy.
Cross-Encoder remains an ablation; retrieval UFPR remains 1. All 38 corpus
applicability assertions remain pending human review. No held-out retrieval or
paid generation was rerun. The v0.1.0 tag, release, assets and historical notes
remain available and untouched.

Install

Use Python 3.12 (tested; package requires ≥3.12):

python -m pip install "version-aware-retriever[dense] @ https://github.com/meiming2019-web/version-aware-retriever/releases/download/v0.1.1/version_aware_retriever-0.1.1-py3-none-any.whl"

The same pinned external dense-model cache is required for context-only retrieval;
weights are not bundled. See README for preparation and CLI commands.

Assets: wheel, source distribution, SHA256SUMS and these release notes. Both
distributions include the root MIT License. Wheel data remains runtime-only;
benchmark gold, reviews and results are excluded. Third-party source material
retains its original notices/licenses; model and provider terms remain separate.

v0.1.0 — Version-Aware RAG MVP

v0.1.0 — Version-Aware RAG MVP

First usable GitHub release of the bounded Pydantic migration RAG project.

• Frozen BM25 + pinned dense encoder → equal-weight RRF → Top 5.
• Grounded structured answers with source-linked citations or INSUFFICIENT_EVIDENCE.
• 34 evidence units; 16 held-out questions and 544 explicitly human-reviewed pairs.
• Held-out RRF Recall@3 .7833, Recall@5 .8111, MRR .8333.
• Cross-Encoder remains an ablation: less observed wrong-version exposure, but
  lower held-out Recall@3/5 and MRR. One negative pair does not establish broad correctness.
• Three frozen generation examples: two SUPPORTED and one INSUFFICIENT_EVIDENCE;
  decision matches 3/3, reviewed-valid citations 3/3, not answer accuracy.
• Installable wheel with immutable hash-checked corpus/provenance resources,
  original upstream notices and the version-aware-retriever console CLI.
• Deterministic offline tests, clean-install acceptance and safe TLS diagnostics.

Install

Use Python 3.12 (tested); package metadata requires ≥3.12:

python -m pip install "version-aware-retriever[dense] @ https://github.com/meiming2019-web/version-aware-retriever/releases/download/v0.1.0/version_aware_retriever-0.1.0-py3-none-any.whl"
python -m version_aware_retriever.dense prepare
version-aware-retriever --query "In Pydantic 2.5.3, how do I validate a dictionary into a model?" --context-only

Preparation downloads the fixed external snapshot; weights are not bundled.
Generation requires OPENAI_API_KEY and OPENAI_MODEL=gpt-5.6-sol. It sends the query
and five excerpts to OpenAI and may incur charges. No paid experiment is rerun
for this release. See README for source validation, reproduction and safe TLS setup.

Assets: wheel, source distribution, SHA256SUMS and these notes. Checksums cover
the wheel and sdist. Wheel data excludes gold/review/results and raw sources;
the sdist retains reproducibility data. Not published to PyPI.

Limits and licensing

Single ecosystem, small/shared corpus, single explicit human batch review of
AI-assisted annotations, one reviewed held-out wrong-version pair, uncalibrated
generation abstention and no production readiness claim. Retrieval UFPR remains 1.
All 38 corpus applicability assertions remain pending human review.

No root license was selected for this project's own code. Public release is not
an open-source license grant. Pydantic excerpts retain upstream MIT notices;
model/provider terms are separate. No frozen metrics, labels, prompts, outputs
or retrieval settings were changed for shipping.