fix: track active embed provider to prevent dimension mismatch on query

backend/rag.py

@@ -50,6 +50,7 @@ class RAGEngine:
         self.top_k = top_k
         self.llm_model = llm_model
         self.embed_provider = embed_provider
+        self._active_embed_provider = embed_provider  # tracks what was actually used
         self.chunks: list[str] = []
         self.chunk_embeddings: np.ndarray | None = None
         self._ingest_progress: dict = {"state": "idle", "embedded": 0, "total": 0}
@@ -65,8 +66,8 @@ class RAGEngine:
             self._genai_client = None
 
     def _embed_texts(self, texts: list[str], task_type: str = "RETRIEVAL_DOCUMENT") -> np.ndarray:
-        """Embed a batch of texts using the configured provider."""
-        if self.embed_provider == "gemini":
+        """Embed using the active provider (locked in after first successful batch)."""
+        if self._active_embed_provider == "gemini":
             return self._embed_gemini(texts, task_type)
         return self._embed_hf(texts)
 
@@ -82,7 +83,8 @@ class RAGEngine:
                 emb = emb.reshape(1, -1)
             return emb
         except Exception as e:
-            log.warning("Gemini embedding failed (%s), falling back to HF", e)
+            log.warning("Gemini embedding failed (%s), switching to HF for this session", e)
+            self._active_embed_provider = "hf"  # lock to HF for all remaining batches + queries
             return self._embed_hf(texts)
 
     def _embed_hf(self, texts: list[str]) -> np.ndarray:
@@ -128,6 +130,7 @@ class RAGEngine:
         )
         self.chunks = splitter.split_text(text)
         self.chunk_embeddings = None
+        self._active_embed_provider = self.embed_provider  # reset for each fresh ingestion
         self._ingest_generation += 1
         total = len(self.chunks)
         self._ingest_progress = {"state": "running", "embedded": 0, "total": total}