fix(embeddings): every embedding model resolves from the index's recorded space or the active model; unrecorded docs collections keep working and propose a stamp (#18143)

[mrveiss]

Thinking Path

#18143 was reopened on its AC1, "one resolver for the generator and every cache key": several pipelines still hard-coded nomic-embed-text. Swapping every literal to active_embedding_model() would reintroduce the bug #18230 fixed: a pipeline with stored vectors would query in a different space after an active-model change. So each embedding-model literal was classified:

1. Index-bound (embeds into or queries a stored index): its space is recorded in knowledge/index_embedding_space.py, and the generator and cache key resolve from that table.
2. Active (fresh embeddings, nothing stored): active_embedding_model().
3. Not a model choice (pricing tables, catalogue lists, defaults): allowlisted, each with a reason.

What Changed

• Index-bound: the pattern analyzer, cross-language patterns (detector and endpoint) and the code-analytics collections (analytics_infrastructure via a CollectionEmbeddingSpace mixin) resolve their model and cache key from the table (make_key(space.space_id, …)).
  • Ten collections are recorded, each nomic-embed-text/768, taken from its own build path.
  • A collection missing from the table refuses (EmbeddingUnavailableError).
  • The cache keys change, for example nomic-embed-text becomes nomic-embed-text/768, so old entries miss once.
• Active: NPU embeddings and the accelerator use active_npu_worker_model() (new, in autobot_shared/embedding_model.py). The memory-graph config reads active_embedding_model() at call time. The cognition seeder embeds with the active model, records the (model, dim) it creates, and refuses a collection recorded in a different space. The doc indexer's hard-coded fallback default is removed, so an empty model now fails loudly.
• Docs search (services/knowledge/doc_searcher.py) uses the space the collection records. For a collection that records none (the live autobot_docs records none, verified read-only), the owner chose on 2026-10-11:
  • It reads one stored vector's dimension. If it equals the active model's dimension, search keeps working, with a once-per-process warning, and files a Umbrella: SLM Data hygiene page — detect and safely clean up orphaned data across services #17038 proposal to stamp the collection's space (a pending one is reused). Otherwise it refuses.
  • The stamp executor (collection_space_stamp_action.py, registered with its proposer, so the bug(approvals-seam): the orphan-storage deletion gate has a detector and an executor but no proposer — the registered post-approval action is unreachable #17315 guard passes) runs on approval. It re-checks the dimension, refuses on a mismatch, merges the model and dim into the existing metadata, reads it back, and writes an approval comment and an audit entry.
• One HNSW reader (knowledge/backends/hnsw_params.py): a ChromaDB modify drops hnsw:* from metadata but keeps it in configuration. So the helper reads configuration["hnsw"] first, with the metadata fallback. knowledge/index.py uses it, and the in-memory backend mirrors real ChromaDB.
• Guard (repo_tests/embedding_model_literal_guard_18143_test.py): no bare embedding-model literal in production code outside the table, the SSOT defaults and a reasoned allowlist. Contrast fixtures, a reach floor, and stale entries fail.

Behaviour changes

• A changed active model no longer moves the query space of the pattern, cross-language or code-analytics pipelines.
• CrossLanguagePatternDetector and _init_infrastructure no longer take an embedding_model argument.
• Docs search on an unrecorded collection with a matching dimension keeps working, warns, and proposes a stamp. With a mismatched dimension, it refuses.

Verification

• Services, knowledge backends, codebase analytics, the new tests, the literal guard and the size ratchet: 3644 passed. The HNSW and stamp follow-up set: 156 passed.
• Hand check, not a repo test (chromadb is stubbed in the test conftest): against a temporary PersistentClient on the pinned chromadb, the existing metadata keys survived modify_metadata; hnsw:space left metadata but stayed in configuration.hnsw; queries still returned cosine distances.

Not done, stated rather than implied

• The analyzer's inline content digest: waits for autobot_shared/content_digest.py, which comes from the kb: two independent semantic chunkers — the ingest pipeline uses neither the embedding-based one nor its metadata #14485/canonical(knowledge): two functions named content_hash return different values — full digest vs 16-char truncation #15722 branch.
• fix(code-index): the autobot_code collection can hold 384-d MiniLM and 768-d nomic vectors side by side — one space per index #18319: the autobot_code collection can hold 384-d and 768-d vectors side by side. It is left unrecorded, so any vector read refuses.

Closes #18143

Model Used

Claude Opus 5.5 (master session: classification decision, host check, PR). Implementation by a Claude Sonnet subagent.

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Next included review available in 1 minute.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

• Configuration used: Repository: mrveiss/AutoBot-AI/.coderabbit.yaml
• Review profile: ASSERTIVE
• Plan: Advanced
• Run ID: 2c369f0a-132d-4922-b645-ce369ab2750c

---

📥 Commits

Reviewing files that changed from the base of the PR and between 6c68345 and c7771ad.

---

⛔ Files ignored due to path filters (2)

• repo_tests/python_file_size_ratchet_baseline.py is excluded by !repo_tests/python_file_size_ratchet_baseline.py
• scripts/python_file_size_known_large.py is excluded by !scripts/python_file_size_known_large.py

---

📒 Files selected for processing (26)

• autobot-backend/ai_hardware_accelerator.py
• autobot-backend/api/approval_gates.py
• autobot-backend/api/codebase_analytics/endpoints/cross_language_patterns.py
• autobot-backend/api/codebase_analytics/endpoints/cross_language_scoping_test.py
• autobot-backend/api/codebase_analytics/npu_embeddings.py
• autobot-backend/autobot_memory_graph/core.py
• autobot-backend/code_intelligence/analytics_infrastructure.py
• autobot-backend/code_intelligence/cross_language_patterns/detector.py
• autobot-backend/code_intelligence/pattern_analysis/analyzer.py
• autobot-backend/code_intelligence/pattern_analysis/embedding_test.py
• autobot-backend/initialization/lifespan.py
• autobot-backend/knowledge/backends/base.py
• autobot-backend/knowledge/backends/chromadb_adapter.py
• autobot-backend/knowledge/backends/hnsw_params.py
• autobot-backend/knowledge/backends/memory_adapter.py
• autobot-backend/knowledge/index.py
• autobot-backend/knowledge/index_embedding_space.py
• autobot-backend/services/knowledge/cognition_seeder.py
• autobot-backend/services/knowledge/collection_space_proposal.py
• autobot-backend/services/knowledge/collection_space_stamp_action.py
• autobot-backend/services/knowledge/doc_indexer.py
• autobot-backend/services/knowledge/doc_searcher.py
• autobot-backend/tests/test_index_bound_embedding_space_18143.py
• autobot_shared/embedding_model.py
• autobot_shared/paths.py
• repo_tests/embedding_model_literal_guard_18143_test.py

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No new hardcoded values of either class — ssot and other both block.

Known backlog in pipeline-scripts/hardcoded_values_baseline.txt is suppressed and tracked in #14371.