Fix Weak Models Concurrency Issue (#811)

backend/danswer/configs/app_configs.py

@@ -199,6 +199,8 @@ DOC_TIME_DECAY = float(
     os.environ.get("DOC_TIME_DECAY") or 0.5  # Hits limit at 2 years by default
 )
 FAVOR_RECENT_DECAY_MULTIPLIER = 2
+# Currently this next one is not configurable via env
+DISABLE_LLM_QUERY_ANSWERABILITY = QA_PROMPT_OVERRIDE == "weak"
 DISABLE_LLM_FILTER_EXTRACTION = (
     os.environ.get("DISABLE_LLM_FILTER_EXTRACTION", "").lower() == "true"
 )

backend/danswer/secondary_llm_flows/query_validation.py

@@ -1,6 +1,7 @@
 import re
 from collections.abc import Iterator
 
+from danswer.configs.app_configs import DISABLE_LLM_QUERY_ANSWERABILITY
 from danswer.direct_qa.interfaces import DanswerAnswerPiece
 from danswer.direct_qa.interfaces import StreamingError
 from danswer.llm.factory import get_default_llm
@@ -52,7 +53,18 @@ def get_query_answerability(user_query: str) -> tuple[str, bool]:
     return reasoning, answerable
 
 
-def stream_query_answerability(user_query: str) -> Iterator[str]:
+def stream_query_answerability(
+    user_query: str, skip_check: bool = DISABLE_LLM_QUERY_ANSWERABILITY
+) -> Iterator[str]:
+    if skip_check:
+        yield get_json_line(
+            QueryValidationResponse(
+                reasoning="Query Answerability Eval feature is turned off",
+                answerable=True,
+            ).dict()
+        )
+        return
+
     messages = get_query_validation_messages(user_query)
     filled_llm_prompt = dict_based_prompt_to_langchain_prompt(messages)
     try:

backend/danswer/server/chat/search_backend.py

@@ -104,6 +104,7 @@ def query_validation(
 def stream_query_validation(
     new_message_request: NewMessageRequest, _: User = Depends(current_user)
 ) -> StreamingResponse:
+    # Note if weak model prompt is chosen, this check does not occur
     query = new_message_request.query
     return StreamingResponse(
         stream_query_answerability(query), media_type="application/json"