Skip to content

feat(embeddings): bind task roles and publish bounded BGE/Jina CPU comparison - #107

Merged
phense merged 11 commits into
mainfrom
feat/embedding-comparison-83
Nov 4, 2026
Merged

phense merged 11 commits into
mainfrom
feat/embedding-comparison-83

Conversation

@phense

@phense phense commented Nov 4, 2026

Copy link
Copy Markdown
Owner

Optional task-bound embedding profiles now keep document builds and duplicate checks in the document vector space, queries in the query space, and require both role probes before activation. Empty prefixes preserve existing identity and request bytes. Existing schema 039 data, access boundaries and selected generations are unchanged.

Publish a reproducible CPU comparison of BGE with the pinned Jina v5 omni Nano/Small text paths, including failed attempts, independent DE/EN questions, paired ranking intervals and populated exact-vector recovery. Retain BGE: Nano is faster on this VM but has mixed independent retrieval results, and Small exceeds the fixed candidate budgets. This PR does not adopt Jina or change production configuration.

Refs #83. The original broader acceptance remains open; native measurements and replay do not establish live serving parity or target-host adoption readiness.

Validation

  • Full suite: 2926 passed, 12 skipped, one known warning, exit 0, 3805.29 s. Command: .venv/bin/python -m pytest -q --tb=short -p no:cacheprovider --durations=20 -x.
  • 168 focused provider/benchmark/release tests passed, including archived populated local/central upgrades. Final documentation checks: 17 passed.
  • The suite captured runtime, scripts, tests, SQL and release metadata at 63266ae; all those inputs are byte-identical at the PR head. Subsequent changes only correct documentation presentation and synchronize/archive the completed task-prefix OpenSpec delta. The all-file unchanged flag is intentionally false; the final documentation is separately checked.
  • Populated BGE 1024/Nano 768 replay completed: strict source/target restore; 326 original rows across 55 tables retained; grants, roles and memberships preserved; interrupted/resumed builds, activation failure, dimension/revision/profile refusal, scope checks, post-selection writes and catch-up/rollback exercised. Both owned databases removed. Embeddings were replayed from exact measured vectors; CLI parser/dispatch and MCP tool dispatch ran in-process, without full client transports.
  • Independent complete bounded reviews: 0 Critical / 0 High / 0 Medium. Ranking counts, paired bootstrap intervals and all 74 source/archive hashes were independently reproduced. Non-blocking supervisor/exporter follow-ups are tracked as LOW116/117/118 in Low level bugs #75.
  • Version 0.6.32 → 0.6.33; wheel built, 39 SQL files byte-identical and lock checked. No SQL migration or automatic vector rewrite.

Evidence and limits

See the comparison and reproducible commands for raw observations, model/source manifests, denominators, errata and failed/unattempted runs. The corpus is synthetic/public. Jina results are native text-path observations, not Ollama endpoint parity or a license approval. Small is not treated as a completed candidate, and the smaller Nano vector analysis is not an activated index generation.

@phense
phense merged commit a25192d into main Nov 4, 2026
2 checks passed
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