fix(core): include LiteLLM roles in embedding identity

Signed-off-by: phernandez <paul@basicmachines.co>

src/basic_memory/repository/litellm_provider.py

@@ -123,6 +123,20 @@ class LiteLLMEmbeddingProvider(EmbeddingProvider):
             attrs["forward_dimensions"] = self.forward_dimensions
         return attrs
 
+    def identity_key(self) -> str:
+        """Return the embedding semantics that should invalidate stored vectors."""
+        document_input_type = self.document_input_type or "-"
+        query_input_type = self.query_input_type or "-"
+        forward_dimensions = str(
+            _should_forward_dimensions(self.model_name, self.forward_dimensions)
+        ).lower()
+        return (
+            f"{self.model_name}:{self.dimensions}:"
+            f"document_input_type={document_input_type}:"
+            f"query_input_type={query_input_type}:"
+            f"forward_dimensions={forward_dimensions}"
+        )
+
     async def _embed(self, texts: list[str], *, input_type: str | None) -> list[list[float]]:
         if not texts:
             return []

src/basic_memory/repository/search_repository_base.py

@@ -586,11 +586,19 @@ class SearchRepositoryBase(ABC):
     def _embedding_model_key(self) -> str:
         """Build a stable model identity for vector invalidation checks."""
         assert self._embedding_provider is not None
-        return (
-            f"{type(self._embedding_provider).__name__}:"
-            f"{self._embedding_provider.model_name}:"
-            f"{self._embedding_provider.dimensions}"
-        )
+        provider = self._embedding_provider
+
+        provider_identity = f"{provider.model_name}:{provider.dimensions}"
+        from basic_memory.repository.litellm_provider import LiteLLMEmbeddingProvider
+
+        if isinstance(provider, LiteLLMEmbeddingProvider):
+            # Trigger: LiteLLM can change request semantics without changing model/dimensions.
+            # Why: asymmetric providers use role-specific document/query params, and
+            # dimension forwarding changes provider-side output-size behavior.
+            # Outcome: reindex treats those semantic config changes as stale vectors.
+            provider_identity = provider.identity_key()
+
+        return f"{type(provider).__name__}:{provider_identity}"
 
     def _plan_entity_vector_shard(
         self,

tests/repository/test_sqlite_vector_search_repository.py

@@ -11,6 +11,7 @@ from sqlalchemy import text
 
 from basic_memory import db
 from basic_memory.config import BasicMemoryConfig, DatabaseBackend
+from basic_memory.repository.litellm_provider import LiteLLMEmbeddingProvider
 from basic_memory.repository.search_index_row import SearchIndexRow
 from basic_memory.repository.sqlite_search_repository import SQLiteSearchRepository
 from basic_memory.schemas.search import SearchItemType, SearchRetrievalMode
@@ -296,6 +297,30 @@ async def test_sqlite_vector_sync_skips_unchanged_and_reembeds_changed_content(s
     assert model_changed_result.embedding_jobs_total == model_changed_result.chunks_total
 
 
+def test_sqlite_embedding_model_key_includes_litellm_role_settings():
+    """LiteLLM role changes should invalidate previously embedded document chunks."""
+    repo = _make_sqlite_repo_for_unit_tests()
+    repo._embedding_provider = LiteLLMEmbeddingProvider(
+        model_name="nvidia_nim/nvidia/embed-qa-4",
+        dimensions=1024,
+        document_input_type="passage",
+        query_input_type="query",
+    )
+    passage_query_key = repo._embedding_model_key()
+
+    repo._embedding_provider = LiteLLMEmbeddingProvider(
+        model_name="nvidia_nim/nvidia/embed-qa-4",
+        dimensions=1024,
+        document_input_type="document",
+        query_input_type="query",
+    )
+    document_query_key = repo._embedding_model_key()
+
+    assert passage_query_key != document_query_key
+    assert "document_input_type=passage" in passage_query_key
+    assert "query_input_type=query" in passage_query_key
+
+
 @pytest.mark.asyncio
 async def test_sqlite_prepare_window_uses_shared_reads_and_serialized_write_scope(monkeypatch):
     """SQLite should batch read-side prepare work but serialize write-side mutations."""