fix(knowledge): one active embedding model resolver, model-aware cache keys, and search refuses rather than embed in another space (#18143, #18185) by mrveiss · Pull Request #18230 · mrveiss/AutoBot-AI

refactor(kb): one resolver for the active embedding model, shared by the generator and every cache key (#18143)

mrveiss

committed

3c4ed3c

refactor(knowledge): chunker exposes a public embedding API; no caller reaches its private hooks (#18185)

mrveiss

committed

b56a0e8

fix(knowledge): the embedding fallback uses the canonical model or refuses, never a different model's vectors (#18185)

mrveiss

committed

8ac0f3e

fix(search): an embedding refusal reaches the API as a failed search with its reason (#18185)

mrveiss

committed

ab49250

fix(knowledge): the L2 embedding cache key carries the active model, so a model change never serves another model's vectors (#18143)

mrveiss

committed

354d119

fix(search): an embedding refusal is a 503 with its reason, never a 200 with empty results (#18185)

mrveiss

committed

c535ad4

fix(search): in-process callers surface an embedding refusal instead of reporting no matches (#18185)

mrveiss

committed

7ca73d4

fix(npu): the client sends the worker the model name the worker loads, so no request falls into the mock path (#18143)

mrveiss

committed

3097e24

fix(search): query embeddings never come from the accelerator's unified space, and a refused batch query keeps its reason (#18185)

mrveiss

committed

4c9b84c

revert(search): NPU query embeddings stay in the index's space; the three-space mismatch is tracked separately (#18185)

mrveiss

committed

60e1366

fix(search): when the index-space embedding is unavailable the search refuses, never emits a vector in another space (#18185)

mrveiss

committed

2cf5fa9

fix(search): a query embeds in its index's recorded space or the search refuses, and the refusal reaches the caller (#18185)

mrveiss

committed

7d9177b

Navigation Menu

Repository navigation

fix(knowledge): one active embedding model resolver, model-aware cache keys, and search refuses rather than embed in another space (#18143, #18185) - #18230

fix(knowledge): one active embedding model resolver, model-aware cache keys, and search refuses rather than embed in another space (#18143, #18185)#18230
mrveiss merged 12 commits into
mainfrom
issue-18143-embedding-model-resolver

Commits on Oct 11, 2026

refactor(kb): one resolver for the active embedding model, shared by the generator and every cache key (#18143)

refactor(knowledge): chunker exposes a public embedding API; no caller reaches its private hooks (#18185)

fix(knowledge): the embedding fallback uses the canonical model or refuses, never a different model's vectors (#18185)

fix(search): an embedding refusal reaches the API as a failed search with its reason (#18185)

fix(knowledge): the L2 embedding cache key carries the active model, so a model change never serves another model's vectors (#18143)

fix(search): an embedding refusal is a 503 with its reason, never a 200 with empty results (#18185)

fix(search): in-process callers surface an embedding refusal instead of reporting no matches (#18185)

fix(npu): the client sends the worker the model name the worker loads, so no request falls into the mock path (#18143)

fix(search): query embeddings never come from the accelerator's unified space, and a refused batch query keeps its reason (#18185)

revert(search): NPU query embeddings stay in the index's space; the three-space mismatch is tracked separately (#18185)

fix(search): when the index-space embedding is unavailable the search refuses, never emits a vector in another space (#18185)

fix(search): a query embeds in its index's recorded space or the search refuses, and the refusal reaches the caller (#18185)