The fleet embedding audit was revalidated against main 40f3a900b9ad0155062f03a47eeeac5cb49afab8: all 305 current structure records have pinned MiniLM 384D vectors, a PCA map and cosine neighbors, but no PaCMAP implementation. The visible map and SVG aria-label explicitly say PCA. Existing PaCMAP issue search returned no matches.
Coordinate with CultureBotAI/culturebotai-claw#426. Add the shared semantic-text space (BAAI/bge-large-en-v1.5 revision d4aa6901d3a41ba39fb536a557fa166f842b0e09 plus PaCMAP), while preserving the useful pinned MiniLM/PCA and cosine-neighbor view with accurate names. PaCMAP projects vectors and is a separate choice from the encoder; do not relabel MiniLM vectors as BGE.
Acceptance:
- Implement real PaCMAP with pinned runtime, deterministic input order/seed and complete vector-source/model/reducer provenance; a first PaCMAP canary can use existing committed vectors without neural inference.
- Introduce a versioned structure semantic adapter for the shared BGE view, with normalized vectors and keyed incremental cache reuse. Preserve unchanged vectors when source semantic hashes and model/adapter metadata are unchanged.
- Generate and publish full-current-corpus artifacts for the common view; keep source-space distinctions and existing useful views explicit in visible text, selector/links and accessibility labels.
- Validate expected model dimensions, complete IDs/source hashes, finite coordinates, source vector fingerprint and reducer parameters. Provide refresh/build/offline-verify recipes and tests; ordinary CI should not run neural inference or redownload weights.
- Verify rendered map pages and all record links, and keep neighbors defined in their original high-dimensional space.
This issue does not duplicate or take over #909 / PR #907, which address unchanged-vector churn in the cellular bud membrane PR. The underlying main refresh() still encodes every record; coordinate a durable incremental implementation without interfering with that active branch.
Implementation sites: scripts/build_text_embedding_map.py, data/embeddings/, src/cellstructuremech/templates/embedding_map.html, scripts/render_pages.py, justfile, pyproject.toml, scripts/run_qc.py and tests/test_text_embedding_map.py.
The fleet embedding audit was revalidated against main
40f3a900b9ad0155062f03a47eeeac5cb49afab8: all 305 current structure records have pinned MiniLM 384D vectors, a PCA map and cosine neighbors, but no PaCMAP implementation. The visible map and SVG aria-label explicitly say PCA. Existing PaCMAP issue search returned no matches.Coordinate with CultureBotAI/culturebotai-claw#426. Add the shared semantic-text space (
BAAI/bge-large-en-v1.5revisiond4aa6901d3a41ba39fb536a557fa166f842b0e09plus PaCMAP), while preserving the useful pinned MiniLM/PCA and cosine-neighbor view with accurate names. PaCMAP projects vectors and is a separate choice from the encoder; do not relabel MiniLM vectors as BGE.Acceptance:
This issue does not duplicate or take over #909 / PR #907, which address unchanged-vector churn in the cellular bud membrane PR. The underlying main
refresh()still encodes every record; coordinate a durable incremental implementation without interfering with that active branch.Implementation sites:
scripts/build_text_embedding_map.py,data/embeddings/,src/cellstructuremech/templates/embedding_map.html,scripts/render_pages.py,justfile,pyproject.toml,scripts/run_qc.pyandtests/test_text_embedding_map.py.