Return solutions which are not grounded on search items

app.py

@@ -30,10 +30,14 @@ llm_router = Router(model_list=[
         "litellm_params": {
             "model": os.environ.get("LLM_MODEL"),
             "api_key": os.environ.get("LLM_API_KEY"),
-            "max_retries": 5
+            "rpm": 15,
+            "max_parallel_requests": 4,
+            "allowed_fails": 1,
+            "cooldown_time": 60,
+            "max_retries": 10,
         }
     }
-], cooldown_time=30)
+], num_retries=10, retry_after=30)
 
 # Jinja2 environment to load prompt templates
 prompt_env = Environment(loader=FileSystemLoader(
@@ -140,9 +144,10 @@ async def search_solutions(params: ReqGroupingResponse) -> SolutionSearchRespons
 
         # generate the completion in non-structured mode.
         # the googleSearch tool enables grounding gemini with google search
+        # this also forces gemini to perform a tool call
         req_completion = await llm_router.acompletion(model="chat", messages=[
             {"role": "user", "content": req_prompt}
-        ], tools=[{"googleSearch": {}}])
+        ], tools=[{"googleSearch": {}}], tool_choice="required")
 
         # ==================== structure the solution as a json ===================================
 

prompts/criticize.txt

@@ -1,4 +1,4 @@
-<role>You are an useful engineering assistant for innovation</role>
+<role>You are an expert system designer</role>
 <task>
 You are tasked with criticizing multiple solutions solving a set of requirements on different points, 
 namely the technical challenges of the solutions, the weaknesses and limitations of the proposed solutions.