Restructure metrics collection so they occur at run-time

src/server.py

@@ -881,9 +881,6 @@ class WebServer:
             return
 
         logger.info(f"{agent_type}.process_message: {message.status} {f'...{message.response[-20:]}' if len(message.response) > 20 else message.response}")
-        if message.metadata["eval_count"]:
-            agent.metrics.tokens_prompt.labels(agent=agent.agent_type).inc(message.metadata["prompt_eval_count"])
-            agent.metrics.tokens_eval.labels(agent=agent.agent_type).inc(message.metadata["eval_count"])
         message.status = "done"
         yield message
         return

src/utils/agents/base.py

@@ -263,11 +263,11 @@ class Agent(BaseModel, ABC):
             for response in llm.chat(
                 model=model, 
                 messages=messages, 
-                stream=True, 
                 options={
                     **message.metadata["options"],
                     # "temperature": 0.5,
-                }
+                },
+                stream=True,
             ):
                 # logger.info(f"LLM::Tools: {'done' if response.done else 'processing'} - {response.message}")
                 message.status = "streaming"
@@ -278,6 +278,7 @@ class Agent(BaseModel, ABC):
                     yield message
 
                 if response.done:
+                    self.collect_metrics(response)
                     message.metadata["eval_count"] += response.eval_count
                     message.metadata["eval_duration"] += response.eval_duration
                     message.metadata["prompt_eval_count"] += response.prompt_eval_count
@@ -290,6 +291,10 @@ class Agent(BaseModel, ABC):
             message.metadata["timers"]["llm_with_tools"] = f"{(end_time - start_time):.4f}"
             return
             
+    def collect_metrics(self, response):
+        self.metrics.tokens_prompt.labels(agent=self.agent_type).inc(response.prompt_eval_count)
+        self.metrics.tokens_eval.labels(agent=self.agent_type).inc(response.eval_count)
+
     async def generate_llm_response(self, llm: Any, model: str, message: Message, temperature = 0.7) -> AsyncGenerator[Message, None]:
         logger.info(f"{self.agent_type} - {inspect.stack()[0].function}")
         
@@ -354,6 +359,7 @@ class Agent(BaseModel, ABC):
                     },
                     stream=False  # No need to stream the probe
                 )
+                self.collect_metrics(response)
 
                 end_time = time.perf_counter()
                 message.metadata["timers"]["tool_check"] = f"{(end_time - start_time):.4f}"
@@ -382,6 +388,7 @@ class Agent(BaseModel, ABC):
                     },
                     stream=False
                 )
+                self.collect_metrics(response)
 
                 end_time = time.perf_counter()
                 message.metadata["timers"]["non_streaming"] = f"{(end_time - start_time):.4f}"
@@ -441,6 +448,7 @@ class Agent(BaseModel, ABC):
                     yield message
 
                 if response.done:
+                    self.collect_metrics(response)
                     message.metadata["eval_count"] += response.eval_count
                     message.metadata["eval_duration"] += response.eval_duration
                     message.metadata["prompt_eval_count"] += response.prompt_eval_count

src/utils/agents/job_description.py

@@ -544,6 +544,7 @@ class JobDescription(Agent):
            message.chunk = ""
 
       if response.done:
+          self.collect_metrics(response)
           message.metadata["eval_count"] += response.eval_count
           message.metadata["eval_duration"] += response.eval_duration
           message.metadata["prompt_eval_count"] += response.prompt_eval_count