Skip to content

feat(examples): package ABI course deep projects#214

Closed
svonava wants to merge 1 commit into
mainfrom
feat/1963-deep-course-pack
Closed

feat(examples): package ABI course deep projects#214
svonava wants to merge 1 commit into
mainfrom
feat/1963-deep-course-pack

Conversation

@svonava

@svonava svonava commented Jul 18, 2026

Copy link
Copy Markdown
Contributor

Part of sie-internal#1963.

Complements #213's minimal managed-cloud smoke with the remaining deep-example course packaging. It intentionally does not claim managed US-production readiness before the parallel model-catalog work is frozen and measured.

What changed

  • Select retrieval-ablation as the third project from code evidence: it has qrels plus NDCG/Recall evaluation and isolates reranking as one controlled change. Use document-ocr as the bounded fixture/OCR stage of vision-doc-rag rather than a separate retrieval project.
  • Add examples/abi-course-deep with a machine-readable three-project manifest, explicit catalog/model-revision gates, expected output contracts, change/measurement prompts, and intentionally null credit/budget fields.
  • Add a 12-document/six-query synthetic retrieval fixture and a dependency-free offline contract plus live dense-versus-rerank runner. Live mode requires explicit catalog-frozen model IDs and reads the key only from SIE_API_KEY.
  • Add a deterministic vision corpus preparer that SHA-256 verifies and copies exactly three existing Apache-2.0 synthetic document images; no PDF or dataset download is needed.
  • Add ten fast offline contract tests for sample bounds, qrels, fixture hashes, output shapes, unresolved catalog/budget gates, deterministic preparation, and credential-shaped literals.
  • Commit frozen Python 3.12 locks for the course pack and contract-review example; the latter previously ignored uv.lock despite being the principal student example.
  • Link all three source examples to the bounded course path and mark their current model IDs as candidates pending catalog freeze and live validation.

Safety and release gates

  • No real API key is logged, persisted, or passed on the command line.
  • No AWS, Modal, managed SIE, or external dataset calls were made.
  • Managed model IDs/revisions, cold/warm latency, request IDs, usage, settled credits, and per-student budget remain explicit null/live_measurement_required gates until #1943 and the US-production exercise are complete.

Validation

  • uv sync --project examples/abi-course-deep --frozen
  • uv run --project examples/abi-course-deep --frozen python -m unittest discover -s examples/abi-course-deep/tests (10 passed)
  • uv run --project examples/abi-course-deep --frozen ruff check examples/abi-course-deep
  • uv run --project examples/abi-course-deep --frozen ruff format --check examples/abi-course-deep
  • uv sync --project examples/contract-review-agent --frozen
  • uv run --project examples/contract-review-agent --frozen make-sample (3 contracts, one scan, 12 obligations)
  • uv run --project examples/contract-review-agent --frozen review --list
  • Python compile checks for the course pack, contract-review, and vision-doc-rag scripts
  • git diff --check

@svonava

svonava commented Jul 18, 2026

Copy link
Copy Markdown
Contributor Author

@coderabbitai review

@svonava
svonava force-pushed the feat/1963-deep-course-pack branch from d87b3c2 to 4c59d6d Compare July 18, 2026 06:42
@svonava

svonava commented Jul 18, 2026

Copy link
Copy Markdown
Contributor Author

Superseded by #213. We are improving the canonical quickstart, retrieval-ablation, vision-doc-rag, and contract-review examples directly instead of maintaining an ABI-course-specific package or manifest. The useful bounded/offline paths were reimplemented and validated in those canonical examples; this PR will not be merged.

@svonava svonava closed this Jul 18, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant