update

src/interface.jl

@@ -266,7 +266,7 @@ function updatePlan(a::agentReflex)
   
 end
 
-function actor_mistral_openorca(a::agentReflex, taskevaluate="")
+function actor_mistral_openorca(a::agentReflex, taskrecap="")
   """
     general prompt format:
 
@@ -316,13 +316,24 @@ function actor_mistral_openorca(a::agentReflex, taskevaluate="")
   """
     <|system|>
     $(a.roles[a.role])
-    You have access to the following tools:
-    $toollines
-    Your earlier work:
-    $shorttermMemory
-    Your progress evaluation of the plan:
-    $taskevaluate
-    $(a.thinkingFormat[:actor])
+      <You have access to the following tools>
+      $toollines
+      </You have access to the following tools>
+      <Your plan>
+      $(a.memory[:shortterm]["Plan 1:"])
+      </Your plan>
+      <Your earlier work's recap>
+      $taskrecap
+      </Your earlier work's recap>
+      <Your job> 
+      Use the following format:
+      Thought: based on the plan and the recap of the plan, what to do? (pay attention to correct numeral calculation and commonsense).
+      Act: an action to take based on your thought, must be one of [{toolnames}]
+      Actinput: your input to the action based on your thought (pay attention to the tool's input)
+      Obs: observed result of the action
+
+      P.S.1 ask user one by one question.
+      </Your job>
     <|/s|>
     <|assistant|>
     Thought: 
@@ -336,11 +347,11 @@ function actor_mistral_openorca(a::agentReflex, taskevaluate="")
   chunkedtext = nothing
   latestTask = nothing
 
-  tempcounter = 0.0
+  tempcounter = 0.1
   while true  # while Thought or Act is empty, run actor again
     tempcounter += 0.1
     @show tempcounter
-    response = sendReceivePrompt(a, prompt, temperature=tempcounter)
+    response = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=tempcounter)
     response = splittext(response, ["Obs", "<|im_end|>"])
     #WORKING
     latestTask = shortMemLatestTask(a.memory[:shortterm]) +1
@@ -608,16 +619,16 @@ function actor(a::agentReflex)
 
   while true  # Actor loop
     # check whether the current task is completed, skip evaluation if memory has only "Plan 1:"
-    taskevaluation = ""
+    taskrecap = ""
     if length(keys(a.memory[:shortterm])) != 1
-      istaskcomplete, taskevaluation = checkTaskCompletion(a)
+      taskrecap = recap(a)
     end
     println("")
-    @show taskevaluation
+    @show taskrecap
     latestTask = shortMemLatestTask(a.memory[:shortterm]) +1
     println(">>> working")
     # work
-    toolname, toolinput, chunkedtext = actor_mistral_openorca(a, taskevaluation)
+    toolname, toolinput, chunkedtext = actor_mistral_openorca(a, taskrecap)
     println("")
     @show toolname
     @show toolinput
@@ -1098,6 +1109,39 @@ function checkTaskCompletion(a)
   return decision, response
 end
 
+function recap(a)
+  @show a.memory[:shortterm]["Plan 1:"]
+  # stimulus = a.memory[:shortterm]["user:"]
+  work = dictToString(a.memory[:shortterm])
+
+  prompt =
+  """
+    <|system|>
+      <Symbol meaning>
+      Plan: a plan
+      Thought: your thought
+      Act: the action you took
+      Actinput: the input to the action
+      Obs: the result of the action
+      </Symbol meaning>
+
+      <Your earlier work>
+      $work
+      </Your earlier work>
+      <Your job>
+      Recap: list all your observed results in detail
+      </Your job>
+
+      Let's think step by step.
+    </s|>
+    <|assistant|>
+    Recap: 
+  """
+  response = sendReceivePrompt(a, prompt, max_tokens=512, temperature=0.0)
+  response = split(response, "</s|>")[1]
+  
+  return response
+end
 
 """ Direct conversation is not an agent, messages does not pass through logic loop
   but goes directly to LLM.