embed_query: repo_type param picks the matching query prefix

app.py

@@ -585,10 +585,21 @@ async def embeddings_batch(req: EmbBatchReq):
 
 
 @app.get("/embed_query")
-@cache(ttl=CACHE_TTL, key="embq:{text}")
-async def embed_query_route(text: str = Query(min_length=3, max_length=200)):
-    """Embed a free-text topic phrase with the serving query encoder."""
-    return {"embedding": embed_query(f"Instruct: {DS_TASK}\nQuery:{text}")}
+@cache(ttl=CACHE_TTL, key="embq:{repo_type}:{text}")
+async def embed_query_route(
+    text: str = Query(min_length=3, max_length=200),
+    repo_type: str = Query(default="datasets"),
+):
+    """Embed a free-text topic phrase with the serving query encoder.
+
+    repo_type picks the query prefix the corpus was embedded against
+    (datasets use the instruct prompt, models the search_query prefix).
+    """
+    if repo_type not in ("datasets", "models"):
+        raise HTTPException(status_code=422, detail="repo_type must be datasets|models")
+    prompt = (f"Instruct: {DS_TASK}\nQuery:{text}" if repo_type == "datasets"
+              else f"search_query: {text}")
+    return {"embedding": embed_query(prompt)}
 
 
 class VectorSearchReq(BaseModel):