Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: clamp the top_k a model passes to search_knowledge_base (#14222)

Roydon Sequeira772ac8b1f82b
diff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.py
index 564ae55efd6..0ecd2a13f07 100644
--- a/studio/backend/core/inference/tools.py
+++ b/studio/backend/core/inference/tools.py
@@ -15054,6 +15054,16 @@ def _scope_retrieval_kwargs(scope: dict) -> dict:
return {"mode": mode if mode in ("hybrid", "dense", "lexical") else "hybrid"}
+# The sidebar's Top K maximum. Uncapped, lexical binds top_k as an SQL LIMIT and dense fails past vec0's KNN limit.
+_MAX_KNOWLEDGE_SEARCH_TOP_K = 50
+
+
+def _kb_top_k(value) -> int | None:
+ """``value`` capped at the sidebar maximum, or None (the search default) when it is not a positive count."""
+ k = _opt_int(value)
+ return min(k, _MAX_KNOWLEDGE_SEARCH_TOP_K) if k is not None and k > 0 else None
+
+
def _search_knowledge_base(arguments: dict, rag_scope: dict | None) -> str:
"""Run the RAG search bound to the hidden per-request ``rag_scope`` (the model supplies only
``query``/``top_k``). Lazy import; missing sqlite-vec degrades to a friendly message."""
@@ -15070,7 +15080,7 @@ def _search_knowledge_base(arguments: dict, rag_scope: dict | None) -> str:
logger.warning("RAG tool unavailable: %s", exc)
return "Knowledge base search is unavailable on this server."
- top_k = _opt_int((arguments or {}).get("top_k") or scope.get("default_top_k"))
+ top_k = _kb_top_k((arguments or {}).get("top_k")) or _kb_top_k(scope.get("default_top_k"))
text, sources = search_knowledge_base_with_sources(
query = str(query),
scope_kb_id = scope.get("kb_id"),
diff --git a/studio/backend/tests/test_rag_retrieval.py b/studio/backend/tests/test_rag_retrieval.py
index 94c10739a33..0c1986d9f22 100644
--- a/studio/backend/tests/test_rag_retrieval.py
+++ b/studio/backend/tests/test_rag_retrieval.py
@@ -469,6 +469,26 @@ def test_scope_overrides_reach_retrieval(monkeypatch):
assert seen["mode"] == "hybrid"
+def test_model_top_k_is_clamped_in_every_mode(rag_conn, bow_embeddings, monkeypatch):
+ tools = _shared_setup_1(monkeypatch)
+ for i in range(60):
+ _add_doc(rag_conn, "kb_a", f"d{i}", f"f{i}.txt", f"h{i}", f"alpha note {i}")
+
+ def returned(top_k, mode):
+ out = tools._search_knowledge_base(
+ {"query": "alpha", "top_k": top_k}, {"kb_id": "a", "mode": mode, "default_top_k": 6}
+ )
+ return out.partition(tools.RAG_SOURCES_SENTINEL)[0].count("<chunk id=")
+
+ for mode in ("hybrid", "lexical", "dense"):
+ assert returned(3, mode) == 3, mode
+ # Not a count, so the sidebar's setting applies.
+ assert returned(-1, mode) == 6, mode
+ assert returned(-5, mode) == 6, mode
+ assert returned(2000, "lexical") == tools._MAX_KNOWLEDGE_SEARCH_TOP_K
+ assert returned(2000, "dense") == tools._MAX_KNOWLEDGE_SEARCH_TOP_K
+
+
def test_build_rag_autoinject_scope_overrides_env(monkeypatch):
tools = _shared_setup_1(monkeypatch)
seen: dict = {}
diff --git a/tests/utils/test_attention_native_routing.py b/tests/utils/test_attention_native_routing.py
index 711850290da..77e66e13678 100644
--- a/tests/utils/test_attention_native_routing.py
+++ b/tests/utils/test_attention_native_routing.py
@@ -102,12 +102,20 @@ def _dense(monkeypatch, dtype, mask):
def _gqa_falls_to_math(Q, K, masked):
# Expand K/V only where GQA leaves math (or nothing) while the expanded call has a faster kernel:
# before torch 2.14 the memory-efficient kernel rejects GQA, so masked fp32 GQA had only math.
- mask = torch.ones(Q.shape[0], 1, Q.shape[2], K.shape[2], dtype = torch.bool, device = Q.device) if masked else None
+ mask = (
+ torch.ones(Q.shape[0], 1, Q.shape[2], K.shape[2], dtype = torch.bool, device = Q.device)
+ if masked
+ else None
+ )
math = int(SDPBackend.MATH)
def choice(k, gqa):
try:
- return int(torch._fused_sdp_choice(Q, k, k, attn_mask = mask, is_causal = not masked, enable_gqa = gqa))
+ return int(
+ torch._fused_sdp_choice(
+ Q, k, k, attn_mask = mask, is_causal = not masked, enable_gqa = gqa
+ )
+ )
except RuntimeError: # no enabled backend takes this call
return None