Last updated: 2026-04-01
This file is the implementation/release engineering log.
It records what changed in code and operational behavior across releases.
User-facing release highlights are tracked in NEWS.md.
It is intentionally different from:
DEVELOPMENT_CONTINUITY.md(living handover + design principles + decision continuity).
In short:
DEVELOPMENT_CONTINUITY.md= "how to continue development safely".IMPLEMENTATION_NOTES.md= "what was implemented and shipped".NEWS.md= "user-facing release highlights".
- Package focus is embedding/vectorization plus distance/scoring workflows.
- Backends are provider-pluggable (
hf,openai,tei) through a shared backend config/dispatch interface. - OpenAI Batch workflow is explicit async:
- submit -> status -> collect
- pending state is expected and non-fatal.
- Demo structure uses provider subfolders:
demos/openalexdemos/openai
- pkgdown output/deploy target is
_site/. - CI includes PR checks matrix and manual triggers.
- Documentation sync to align with implemented API and current repository layout.
- Clarified ridge/reference-area behavior in docs (
distance_ridge()+score_ridge()). - Updated vignette paths and internal/exported function descriptions.
- Version bump to
0.1.3. - Included repository cleanup scope for stale
inst/qdrant functions/*helpers.
- Introduced stronger OpenAI demo/tutorial flow and async handling guidance.
- Added two-phase demo workflow support:
- render does not hard-fail when batch is still pending,
- finalize step performs status/collect/compare later.
- Added persistent direct-vs-batch comparison outputs under:
project/openai_batch_comparison/label=<label>/.
- Improved tutorial narrative consistency across demos.
- Consolidated OpenAI batch helper/module naming and structure.
- Added/updated CI workflows for PR check suites with caching and manual trigger support.
- Added pkgdown GH Pages workflow alignment and
_sitedestination usage. - Documentation cleanup to remove stale scope references and align README/package messaging with embedding + distance computation focus.
- Added
DEVELOPMENT_CONTINUITY.mdas required pre-commit continuity artifact.
- OpenAI batch collection may complete after demo rendering; this is expected async behavior.
- Batch/direct embedding comparisons can show tiny numeric differences due to floating-point serialization and runtime-level numerical variance.
- These small differences are not typically semantic regressions when cosine similarity remains effectively 1.0.
When behavior changes, update both files with non-overlapping intent:
- Update
IMPLEMENTATION_NOTES.mdfor shipped implementation scope. - Update
DEVELOPMENT_CONTINUITY.mdfor design rationale, decisions, and handover continuity. - Update
NEWS.mdfor user-facing release notes. - Ensure README/vignettes match.