update

src/interface.jl

@@ -3,7 +3,7 @@ module interface
 
 export agentReact, agentReflex,
         addNewMessage, clearMessage, removeLatestMsg, conversation, writeEvaluationGuideline,
-        grading, analyze, selfReflext
+        grading, analyze, selfReflext, actor_mistral_openorca2
 
 using JSON3, DataStructures, Dates, UUIDs, HTTP
 using CommUtils, GeneralUtils
@@ -319,7 +319,8 @@ function planner_mistral_openorca(a::agentReflex, usermsg::String)
   return prompt
 end
 
-function actor_mistral_openorca(a::agentReflex, usermsg::T) where {T<:AbstractString}
+#WORKING try use Thought/Act/ActInput/Obs loop because some time step 2 depend on step 1
+function actor_mistral_openorca(a::agentReflex, usermsg, plan, step)
   """
     general prompt format:
 
@@ -359,11 +360,13 @@ function actor_mistral_openorca(a::agentReflex, usermsg::T) where {T<:AbstractSt
     {usermsg}
     <|im_end|>
     <|im_start|>assistant
-
+    {plan}
   """
   
   prompt = replace(prompt, "{role}" => a.roles[a.role])
   prompt = replace(prompt, "{thinkingFormat}" => a.thinkingFormat[:actor])
+  prompt = replace(prompt, "{step}" => step)
+  prompt = replace(prompt, "{plan}" => plan)
   toolnames = ""
   toollines = ""
   for (toolname, v) in a.tools
@@ -634,22 +637,22 @@ function work(a::agentReflex, usermsg::String)
       respond = sendReceivePrompt(a, prompt)
       plan = split(respond, "<|im_end|>")[1]
       plan = split(plan, "Response:")[1]
-      _plan = replace(plan, "Plan:"=>"Plan $(a.attempt):")
-      logmsg = "assistant: $_plan\n"
+      plan = replace(plan, "Plan:"=>"Plan $(a.attempt):")
+      logmsg = "assistant: $plan\n"
       a.memory[:shortterm] *= logmsg
       a.thoughtlog *= logmsg
-      actorstate, msgToUser = actor(a, plan)
+      actorstate, msgToUser = actor(a, usermsg, plan)
 
       if actorstate == "chatbox"
         respond = msgToUser
         break
       elseif actorstate == "all steps done"
         println("all steps done")
-        #WORKING give answer to the question
+
         respond = formulateRespond(a, a.memory[:shortterm])
         a.memory[:shortterm] *= "Respond: $respond\n"
 
-        #TODO evaluate. if score < 8/10 try again.
+        # evaluate. if score < 8/10 try again.
         headerToDetect = ["user:", "assistant:", ]
         headers = detectCharacters(a.memory[:shortterm], headerToDetect)
         chunkedtext = chunktext(a.memory[:shortterm], headers)
@@ -664,9 +667,10 @@ function work(a::agentReflex, usermsg::String)
           a.thoughtlog = ""
           break
         else  # self evaluate and reflect then try again
-          report = analyze(a, a.memory[:shortterm])
-          @show report
-          lessonwithcontext = selfReflext(a, report)
+          analysis = analyze(a, a.memory[:shortterm])
+          @show analysis
+
+          lessonwithcontext = selfReflext(a, analysis)
           @show lessonwithcontext
           a.memory[:shortterm] = ""
           #TODO add lesson and context into longterm memory
@@ -676,7 +680,7 @@ function work(a::agentReflex, usermsg::String)
           @show chunkedtext
           push!(a.memory[:longterm], Dict(:context=>chunkedtext["Context:"],
                                           :lesson=>chunkedtext["Lesson:"]))
-          error(">>>>>>>>>>")
+          error(">>>>>>>>>")
         end
       else
         error("undefied condition, actorstate $actorstate $(@__LINE__)")
@@ -719,7 +723,7 @@ end
     msgToUser = "message from assistant to user"
 
 """
-function actor(a::agentReflex, plan::T) where {T<:AbstractString}
+function actor(a::agentReflex, usermsg, plan)
   actorState = nothing
   msgToUser = nothing
 
@@ -731,28 +735,39 @@ function actor(a::agentReflex, plan::T) where {T<:AbstractString}
     a.step += 1
     @show a.step
     if a.step <= totalsteps
-      stepdetail = extractStepFromPlan(a, plan, a.step) 
-      prompt = actor_mistral_openorca(a, stepdetail)
+      # WORKING in step 2, I need to use a.memory[:shortterm] as input to actor()
+      prompt = actor_mistral_openorca(a, usermsg, plan, a.step)
       @show prompt
       respond = sendReceivePrompt(a, prompt)
-      respond = split(respond, "<|im_end|>")[1]
       @show respond
-      headerToDetect = ["Question:", "Plan:", "Thought:", "Act:", "ActInput:", "Obs:", "...", "Answer:", 
-                      "Conclusion:", "Summary:"]
-      headers = detectCharacters(respond, headerToDetect)
 
+      # some time LLM not generate a number after headers but I want it
+      if occursin("Act:", respond)
+        headerToDetect = ["Question:", "Plan:", "Thought:", 
+                          "Act:", "ActInput:", "Obs:", "...", 
+                          "Answer:", "Conclusion:", "Summary:"]
+        headers = detectCharacters(respond, headerToDetect)
+        respond = addStepNumber(respond, headers, a.step)
+      end
+      
+      respond = split(respond, "Obs $(a.step):")[1]
+      
       # add to memory
-      _respond = addStepNumber(respond, headers, a.step)
-      a.memory[:shortterm] *= _respond
-      a.thoughtlog *= _respond
-
+      a.memory[:shortterm] *= respond
+      a.thoughtlog *= respond
+      headerToDetect = ["Question $(a.step):", "Plan $(a.step):", "Thought $(a.step):", 
+                        "Act $(a.step):", "ActInput $(a.step):", "Obs $(a.step):", "...", 
+                        "Answer $(a.step):", "Conclusion $(a.step):", "Summary $(a.step):"]
+      headers = detectCharacters(respond, headerToDetect)
       chunkedtext = chunktext(respond, headers)
-      toolname = toolNameBeingCalled(chunkedtext["Act:"], a.tools)
-      toolinput = chunkedtext["ActInput:"]
+      @show chunkedtext
+      toolname = toolNameBeingCalled(chunkedtext["Act $(a.step):"], a.tools)
+      toolinput = chunkedtext["ActInput $(a.step):"]
       @show toolname
       @show toolinput
-
+      error(">>>>>>>>>")
       if toolname == "chatbox"  # chat with user
+        #TODO donot use chatbox to respond to user
         respond = toolinput
         msgToUser = respond
         actorState = toolname
@@ -925,8 +940,7 @@ function analyze(a, shorttermMemory::T)  where {T<:AbstractString}
     1. What happened?
     2. List all relationships, each with cause and effect .
     3. Look at each relationship, figure out why it behaved that way.
-    4. Do relationships behaved differently than your expectation? If yes, why?
-    5. What could you do to improve the respond?
+    4. What could you do to improve the respond?
     <|im_end|>
   """
 
@@ -935,7 +949,6 @@ function analyze(a, shorttermMemory::T)  where {T<:AbstractString}
 end
 
 
-
 """ Write a lesson drawn from evaluation.
 
 Args:
@@ -968,7 +981,7 @@ julia> report =
 julia> lesson = selfReflext(agent, report)
 ```
 """
-function selfReflext(a, report::T) where {T<:AbstractString}
+function selfReflext(a, analysis::T) where {T<:AbstractString}
   prompt =
   """
     <|im_start|>system