module interface
    

export agentReact, agentReflex,
        addNewMessage, clearMessage, removeLatestMsg, conversation, directconversation, 
        writeEvaluationGuideline, grading, analyze, selfReflext, 
        formulateUserResponse, extractinfo, updateEnvState, chat_mistral_openorca, 
        recap, readKeywordMemory

using JSON3, DataStructures, Dates, UUIDs, HTTP, Random
using GeneralUtils
using ..type, ..utils, ..llmfunction

# ------------------------------------------------------------------------------------------------ #
#                                        pythoncall setting                                        #
# ------------------------------------------------------------------------------------------------ #
# Ref: https://github.com/JuliaPy/PythonCall.jl/issues/252
# by setting the following variables, PythonCall.jl will use:
#     1. system's python and packages installed by system (via apt install)
# or  2. conda python and packages installed by conda
# if these setting are not set (comment out), PythonCall will use its own python and packages that
# installed by CondaPkg.jl (from env_preparation.jl)
# ENV["JULIA_CONDAPKG_BACKEND"] = "Null"      # set condapkg backend = none
# systemPython = split(read(`which python`, String), "\n")[1]     # system's python path
# ENV["JULIA_PYTHONCALL_EXE"] = systemPython  # find python location with $> which python  ex. raw"/root/conda/bin/python"

# using PythonCall
# const py_agents = PythonCall.pynew()
# const py_llms = PythonCall.pynew()
# function __init__()
#     # PythonCall.pycopy!(py_cv2, pyimport("cv2"))

#     # equivalent to     from urllib.request import urlopen    in python
#     PythonCall.pycopy!(py_agents, pyimport("langchain.agents"))
#     PythonCall.pycopy!(py_llms, pyimport("langchain.llms"))
# end

# ---------------------------------------------- 100 --------------------------------------------- #

""" Add new message to agent.

  Arguments:

  Return:

```jldoctest
julia> addNewMessage(agent1, "user", "Where should I go to buy snacks")
```
"""
function addNewMessage(a::T1, role::String, content::T2) where {T1<:agent, T2<:AbstractString}
  if role ∉ a.availableRole   # guard against typo
    error("role is not in agent.availableRole $(@__LINE__)")
  end

  # check whether user messages exceed limit
  userMsg = 0
  for i in a.messages
    if i[:role] == "user"
      userMsg += 1
    end
  end
  messageleft = 0

  if userMsg > a.maxUserMsg # delete all conversation
    clearMessage(a)
    messageleft = a.maxUserMsg
  else
    userMsg += 1
    d = Dict(:role=> role, :content=> content, :timestamp=> Dates.now())
    push!(a.messages, d)
    messageleft = a.maxUserMsg - userMsg
  end

  return messageleft
end

function clearMessage(a::T)  where {T<:agent}
  for i in eachindex(a.messages)
    if length(a.messages) > 0 
      pop!(a.messages)
    else
      break
    end
  end
  
  a.memory = newAgentMemory()

  # @show a.messages
  # @show a.memory
end

function removeLatestMsg(a::T)  where {T<:agent}
  if length(a.messages) > 1
    pop!(a.messages)
  end
end

function chat_mistral_openorca(a::agentReflex)
  """
    general prompt format:

    "
    <|system|>
    {role}
    {tools}
    {thinkingFormat}
    {context}
    <|im_end|>
    <|im_start|>user
    {usermsg}
    <|im_end|>
    <|im_start|>assistant

    "

    Note:
      {context} = 
        "
        {earlierConversation}
        {env state}
        {shortterm memory}
        {longterm memory}
        "
  """

  conversation = messagesToString(a.messages)

  aboutYourself =
  """
    Your name is $(a.name)
    $(a.roles[a.role])  
  """

  prompt = 
  """
    <|system|>
      <About yourself>
      $aboutYourself
      </About yourself>
    </s>
    $conversation
    <|assistant|>
  """
  
  response = sendReceivePrompt(a, prompt, a.config[:text2text][:mqtttopic], 
                              timeout=180, stopword=["<|", "</"])
  response = split(response, "<|")[1]
  response = split(response, "</")[1]

  return response
end


function planner_mistral_openorca(a::agentReflex)
  """
    general prompt format:

    "
    <|system|>
    {role}
    {tools}
    {thinkingFormat}
    <|im_end|>
    {context}
    <|im_start|>user
    {usermsg}
    <|im_end|>
    <|im_start|>assistant

    "

    Note:
      {context} = 
        "
        {earlierConversation}
        {env state}
        {shortterm memory}
        {longterm memory}
        "
  """

  conversation = messagesToString(a.messages)
  toollines = ""
  for (toolname, v) in a.tools
    if toolname ∉ [""]
      # toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
      toolline = "$toolname: $(v[:description])\n"
      toollines *= toolline
    end
  end

  # skip objective and plan because LLM is going to generate new plan
  shorttermMemory = dictToString(a.memory[:shortterm], skiplist=["Objective:", "Plan 1:"])

  aboutYourself =
  """
    Your name is $(a.agentName)
    $(a.roles[a.role])
    $(a.roleSpecificInstruction[a.role])
  """
  
  # assistant_plan_prompt =
  #   """
  #   <s>
  #     <|system|>
  #     <About yourself>
  #     $aboutYourself
  #     </About yourself>
  #     <You have access to the following tools>
  #     $toollines
  #     </You have access to the following tools
  #     <Your earlier work>
  #     $shorttermMemory
  #     </Your earlier work>
  #     <Your job>
  #     Plan: first you should always think about your conversation with the user and your earlier work thoroughly then extract and devise a complete, task by task, plan to achieve your objective (pay attention to correct numeral calculation and commonsense).
  #     Keyword memory: using JSON format, list all variables in the plan you need to find out
  #     P.S.1 each task of the plan should be a single action.
  #     </Your job>
  #     <Example>
  #     Plan: 
  #     1. Ask the user about how many miles per day they drive
  #     2. Ask the user about what stuff they usually carry with
  #     3. Ask the user about preferred type of car they want to buy (sedan, sport, SUV, etc)
  #     4. Ask the user about their price range
  #     5. Use inventory tool to find cars that match the user's preferences and are within their price range
  #     6. Use finalanswer tool to present the recommended car to the user.
  #     Keyword memory: {"mile per day": null, "carry item": null, "car type": null, "price range": null}
  #     </Example>
  #   </s>
  #   $conversation
  #   <|assistant|>
  #   Plan: 
  #   """

  assistant_plan_prompt =
    """
    <|im_start|>system
      <About yourself>
      $aboutYourself
      </About yourself>
      <You have access to the following tools>
      $toollines
      </You have access to the following tools
      <Your earlier work>
      $shorttermMemory
      </Your earlier work>
      <Your job>
      Plan: first you should always think about your conversation with the user and your earlier work thoroughly then extract and devise a complete, task by task, plan to achieve your objective (pay attention to correct numeral calculation and commonsense).
      Keyword memory: using JSON format, list all variables in the plan you need to find out
      P.S.1 each task of the plan should be a single action.
      </Your job>
      <Example>
      Plan: 
      1. Ask the user about how many miles per day they drive
      2. Ask the user about what stuff they usually carry with
      3. Ask the user about preferred type of car they want to buy (sedan, sport, SUV, etc)
      4. Ask the user about their price range
      5. Use inventory tool to find cars that match the user's preferences and are within their price range
      6. Use finalanswer tool to present the recommended car to the user.
      Keyword memory: {"mile per day": null, "carry item": null, "car type": null, "price range": null}
      </Example>
    <|im_end|>
    $conversation
    <|im_start|>assistant
    Plan: 
    """


  response = sendReceivePrompt(a, assistant_plan_prompt, max_tokens=1024, temperature=0.1,
                          timeout=180, stopword=["<|user|>", "</"])
  response = split(response, "<|")[1]
  response = split(response, "</")[1]
  response = split(response, "\n\n")[1]
  @show response = "Plan: " * response

  headerToDetect = ["Plan:", "Keyword memory:",]
  headers = detectCharacters(response, headerToDetect)
  @show headers
  chunkedtext = chunktext(response, headers)
  a.memory[:c] = chunkedtext

  plan = chunkedtext["Plan:"]
  keywordmem = JSON3.read(chunkedtext["Keyword memory:"])   # use JSON to change from json string to dict

  for (k, v) in keywordmem
    a.memory[:keyword][String(k)] = v    
  end

  return plan
end

""" Update the current plan.
"""
function updatePlan(a::agentReflex)
  # conversation = messagesToString_nomark(a.messages)
  
  toollines = ""
  for (toolname, v) in a.tools
    if toolname ∉ ["askbox"]
      toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
      toollines *= toolline
    end
  end

  work = dictToString(a.memory[:shortterm])

  prompt =
  """
  <s>
    <|system|>
    $(a.roles[a.role])
    Request the user’s input for the following info before you could search wine stock:
      - occasion
      - food pair ask the user
      - user's personal taste of wine
      - ambient temperature at the serving location
      - wine price range
      - wines we have in stock (use tools to get the info)
    You provide a personalized recommendation of up to two wines based on the user's info above, and you describe the benefits of each wine in detail.
    
    You have access to the following tools:
    $toollines

    Your work:
    $work

    Your job is to update the plan using available info from your work.
    P.S. do not update if no info available.

    For example:
    Plan: 1. Ask the user for their food type.
    Obs: It will be Thai dishes.
    Updated plan: 1. Ask the user for their food type (Thai dishes).
    </s>
    Updated plan: 
  """

  result = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=0.1)
  @show updatedPlan = result
  a.memory[:shortterm]["Plan 1:"] = result
  
end


function selfAwareness(a::agentReflex)

  getonlykeys = ["Actinput", "Obs"]
  worknoplan = similar(a.memory[:shortterm])
  worklength = length(a.memory[:shortterm])
  for (i, (k, v)) in enumerate(a.memory[:shortterm])
    if i >= worklength - 1
      count = 0
      for i in getonlykeys
        if occursin(i, k)
          count += 1
        end
      end
      if count != 0
        worknoplan[k] = v 
      end
    end
  end

  println("")
  @show worknoplan

  work = dictToString(worknoplan)

  aboutYourself =
  """
    Your name is $(a.agentName)
    $(a.roles[a.role])

  """

  prompt = 
  """
  <s>
    <|system|>
      <About yourself>
      $aboutYourself
      $(a.roleSpecificInstruction[a.role])
      </About yourself>
      <Your job> 
      Use the following format strictly:
      Info extraction: break the latest observed result down to the smallest unit of information then repeat every important word
      Info mapping: based on extracted info, explicitly state what each info could possibly match which keyword memory's key
      Info matching: using JSON format, state every info relates to what key in my memory matches which value (Donot use nested JSON)
      </Your job>
    </|system|>
    <Example>
      <Your earlier work>
      The user wants to buy an 4WD electric SUV car under 20000 dollars.
      </Your earlier work>
      <Your keyword memory>
      {\"car type\": null, \"powertrain type\": null, \"price\": null, \"color\": null, \"financing\": null, \"drive type\": null}
      </Your keyword memory>
      <|assistant|>
        Info extraction:
          - The user is buying a SUV car.
          - The user prefers 4WD drive type
          - The user prefers electric powertrain
          - price is under 20000 dollars
        Info mapping: 
          - "SUV" could matches "car type" key
          - "electric" could matches "powertrain type" key
          - "under 20000 dollars" could match "price" key
          - "4WD" could matches "drive type" key
        Info matching: {\"car type\": \"SUV\", \"powertrain type\": \"electric\", \"price\": \"under 20000\", \"color\": null, \"financing\": null, \"drive type\": "4WD"}
      </|assistant|>    
    </Example>
  </s>
  <Your earlier work>
    $work
  </Your earlier work>
  <Your keyword memory>
    $(JSON3.write(a.memory[:keyword]))
  </Your keyword memory>
  <|assistant|>
  Info extraction: 
  """
  response = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=0.4, timeout=180,
                                  stopword=["/n/n", "END", "End", "Obs", "<|", "</"])
  response = split(response, "<|")[1]
  response = split(response, "</")[1]
  response = split(response, "</|assistant|>")[1]
  response = "Info extraction:" * response

  println("")
  @show selfaware_1 = response

  headerToDetect = ["Info extraction:", "Info mapping:", "Info matching:", "Actinput"]
  headers = detectCharacters(response, headerToDetect)

  # headers[1:2] is for when LLM generate more than a paire of "Info extraction" and "Info matching", discard the rest
  chunkedtext = chunktext(response, headers[1:3]) 
  println("")
  _infomatch = chunkedtext["Info matching:"]
  _infomatch = GeneralUtils.getStringBetweenCharacters(_infomatch, '{', '}', endCharLocation="end")
  infomatch = GeneralUtils.JSON3read_stringKey(_infomatch)
  # infomatch = copy(JSON3.read(_infomatch))

  println("")
  @show chunkedtext

  println("")
  @show infomatch

  keywordMemoryUpdate!(a.memory[:keyword], infomatch)

  response = "What I know about user:" * JSON3.write(a.memory[:keyword])         # * response
  println("")
  @show selfaware_2 = response


  return response
end

function sentenceToKeywordMemory(a::agentReflex)

  getonlykeys = ["Actinput", "Obs"]
  worknoplan = similar(a.memory[:shortterm])
  worklength = length(a.memory[:shortterm])
  for (i, (k, v)) in enumerate(a.memory[:shortterm])
    if i >= worklength - 1
      count = 0
      for i in getonlykeys
        if occursin(i, k)
          count += 1
        end
      end
      if count != 0
        worknoplan[k] = v 
      end
    end
  end

  println("")
  @show worknoplan

  work = dictToString(worknoplan)

  aboutYourself =
  """
    Your name is $(a.agentName)
    $(a.roles[a.role])

  """

  prompt = 
  """
  <s>
    <|system|>
      <About yourself>
      $aboutYourself
      $(a.roleSpecificInstruction[a.role])
      </About yourself>
      <Your job> 
      Use the following format strictly:
      Info extraction: repeat all important info from the latest observed result thoroughly
      Info mapping: based on extracted info, explicitly state what each info could match which keyword memory's key
      Info matching: using JSON format, what key in my memory matches which info
      </Your job>
    </|system|>
    <Example>
      <Your earlier work>
      The user wants to buy an electric SUV car under 20000 dollars.
      </Your earlier work>
      <Your keyword memory>
      {\"car type\": null, \"engine type\": null, \"price\": null, \"color\": null, \"financing\": null}
      </Your keyword memory>
      <|assistant|>
        Info extraction:
          - The user is buying an electric SUV car.
          - price is under 20000 dollars
        Info mapping: 
          - "SUV" could matches "car type" key
          - "electric" could matches "engine type" key
          - "under 20000 dollars" could match "price" key
        Info matching: {\"car type\": \"SUV\", \"engine type\": \"electric motor\", \"price\": \"under 20000\", \"color\": null, \"financing\": null}
      </|assistant|>    
    </Example>
  </s>
  <Your earlier work>
    $work
  </Your earlier work>
  <Your keyword memory>
    $(JSON3.write(a.memory[:keyword]))
  </Your keyword memory>
  <|assistant|>
  Info extraction: 
  """
  response = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=0.2, timeout=180,
                                  stopword=["/n/n", "END", "End", "Obs", "<|", "</"])
  response = split(response, "<|")[1]
  response = split(response, "</")[1]
  response = split(response, "</|assistant|>")[1]
  response = "Info extraction:" * response

  println("")
  @show selfaware_1 = response

  headerToDetect = ["Info extraction:", "Info mapping:", "Info matching:", "Actinput"]
  headers = detectCharacters(response, headerToDetect)

  # headers[1:2] is for when LLM generate more than a paire of "Info extraction" and "Info matching", discard the rest
  chunkedtext = chunktext(response, headers[1:3]) 
  println("")
  _infomatch = chunkedtext["Info matching:"]
  _infomatch = GeneralUtils.getStringBetweenCharacters(_infomatch, '{', '}', endCharLocation="end")
  infomatch = GeneralUtils.JSON3read_stringKey(_infomatch)
  # infomatch = copy(JSON3.read(_infomatch))

  println("")
  @show chunkedtext

  println("")
  @show infomatch

  keywordMemoryUpdate!(a.memory[:keyword], infomatch)
end

function keywordMemoryToPlanMatching(a::agentReflex)

  aboutYourself =
  """
    Your name is $(a.agentName)
    $(a.roles[a.role])

  """

  prompt = 
  """
  <s>
    <|system|>
      <About yourself>
      $aboutYourself
      $(a.roleSpecificInstruction[a.role])
      </About yourself>
      <Your job> 
      Use the following format strictly:
      Info mapping: explicitly state what each step of the plan could match which keyword memory's key
      Info matching: using JSON format, what key in my memory matches which info
      </Your job>
    </|system|>
    <Example 1>
      <My plan>
        1. ask the user for car type, 
        2. ask the user for color
      </My plan>
      <User info>
      {\"car type\": "SUV", \"color\": null}
      </User info>
      <|assistant|>
        Info mapping: 
          -In step 1 of the plan, I need to ask the user about the car type. From user info, the car type is "SUV". Therefore this step is done.
          -In step 2 of the plan, I need to ask the user about the color . From user info, the color is null. Therefore this step is not done yet and I'll need to do this step. 
      </|assistant|>    
    </Example 1>
  </s>
  <Your earlier work>
    $work
  </Your earlier work>
  <User info>
    $(JSON3.write(a.memory[:keyword]))
  </User info>
  <|assistant|>
  Info mapping: 
  """
  response = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=0.4, timeout=180,
                                  stopword=["/n/n", "END", "End", "Obs", "<|", "</"])
  response = split(response, "<|")[1]
  response = split(response, "</")[1]
  response = split(response, "</|assistant|>")[1]
  response = "Info mapping:" * response

  println("")
  @show keywordMemoryToPlanMatching = response

  error(1111)

  headerToDetect = ["Info extraction:", "Info mapping:", "Info matching:", "Actinput"]
  headers = detectCharacters(response, headerToDetect)

  # headers[1:2] is for when LLM generate more than a paire of "Info extraction" and "Info matching", discard the rest
  chunkedtext = chunktext(response, headers[1:3]) 
  println("")
  _infomatch = chunkedtext["Info matching:"]
  _infomatch = GeneralUtils.getStringBetweenCharacters(_infomatch, '{', '}', endCharLocation="end")
  infomatch = GeneralUtils.JSON3read_stringKey(_infomatch)
  # infomatch = copy(JSON3.read(_infomatch))

  println("")
  @show chunkedtext

  println("")
  @show infomatch

  keywordMemoryUpdate!(a.memory[:keyword], infomatch)
end

# function actor_mistral_openorca(a::agentReflex, selfaware=nothing)
#   getonlykeys = ["Actinput", "Obs"]
#   worknoplan = similar(a.memory[:shortterm])
#   for (k, v) in a.memory[:shortterm]
#     count = 0
#     for i in getonlykeys
#       if occursin(i, k)
#         count += 1
#       end
#     end
#     if count != 0
#       worknoplan[k] = v 
#     end
#   end

#   work = dictToString(worknoplan)
  
  
  
#   """
#     general prompt format:

#     "
#     <|system|>
#     {role}
#     {tools}
#     {thinkingFormat}
#     <|im_end|>
#     {context}
#     <|im_start|>user
#     {usermsg}
#     <|im_end|>
#     <|im_start|>assistant

#     "

#     Note:
#       {context} = 
#         "
#         {earlierConversation}
#         {env state}
#         {shortterm memory}
#         {longterm memory}
#         "
#   """

#   toolslist = []
#   toolnames = ""
#   toollines = ""
#   for (toolname, v) in a.tools
#     toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
#     toollines *= toolline
#     toolnames *= "$toolname, "
#     push!(toolslist, toolname)
#   end

#   thought = "Thought: you should always think about what to do according to the plan (pay attention to correct numeral calculation and commonsense and do one thing at a time.)"
#   startword = "Thought:"
#   if selfaware !== nothing
#     "
#     Thought: based on what you know, you should focus on what you need to improve first then follow your plan to decide what to do next. (P.S. 1) let's think a single step. 2) pay attention to correct numeral calculation and commonsense.)
#     "
#   end
#                         # your should request the missing information first before making a decision
#   aboutYourself =
#   """
#     Your name is $(a.agentName)
#     $(a.roles[a.role])  
#   """

#   winestocksearchresult = nothing
#   if haskey(a.memory, :winestocksearchresult) && a.memory[:winestockResult] !== nothing
#     winestocksearchresult = 
#     """
#     <winestock search result>
#     $(a.memory[:winestocksearchresult])
#     </winestock search result>
#     """
#   else
#     winestocksearchresult = "\n"
#   end

#   prompt = 
#   """
#     <|system|>
#       <About yourself>
#       $aboutYourself
#       </About yourself>
#       <You have access to the following tools>
#       $toollines
#       </You have access to the following tools>
#       <Your plan>
#       $(a.memory[:shortterm]["Plan 1:"])
#       </Your plan>
#       <What I know about the user>
#       $(JSON3.write(a.memory[:keyword]))
#       </What I know about the user>
#       <Your job> 
#       Use the following format:
#       $thought
#       Act: based on your thought what action to choose?, must be one of [{toolnames}].
#       Actinput: your input to the action (pay attention to the tool's input)
#       Obs: observed result of the action
#       </Your job>
#       <Example>
#       <What I know about the user>
#       $(readKeywordMemory(a))
#       </What I know about the user>
#       Thought: based on what you know, I think he also need to know whether there are any charging station near by his house. I should search the internet to get this info.
#       Act: internetsearch
#       Actinput: {\"internetsearch\": \"EV charging station near Bangkok\"}
#       </Example>
#     </s>
#     <|assistant|>
#     $startword
#   """

#   prompt = replace(prompt, "{toolnames}" => toolnames)

#   println("")
#   @show actor_prompt = prompt

#   response = nothing
#   chunkedtext = nothing
#   latestTask = nothing

#   while true  # while Thought or Act is empty, run actor again

#     response = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=0.4, timeout=300,
#                                 stopword=["Thought:", "Obs:", "<|system|>", "</s>", "<|end|>"],
#                                 seed=rand(1000000:2000000))
#     println("")
#     @show actor_raw = response

#     response = splittext(response, ["/n/n", "END", "End","obs", "Obs", "<|im_end|>"])
#     response = split(response, "<|")[1]
#     response = split(response, "</")[1]
#     # response = split(response, "Thought:")[end]
    
#     latestTask = shortMemLatestTask(a.memory[:shortterm]) +1

#     response = startword * response

#     headerToDetect = ["Plan:", "Self-awareness:", "Thought:", 
#                       "Act:", "Actinput:", "Obs:", 
#                       "Answer:", "Conclusion:", "Summary:"]
    
#     # replace headers with headers with correct attempt and task number
#     response = replaceHeaders(response, headerToDetect, latestTask)  
#     response = split(response, "<|")[1]
#     response = split(response, "</")[1]
    
#     # sometime LLM use wrong keyword. use regex to detect "actinput5:" and replace it with "Actinput"
#     regexmatch = match(r"actinput\d+:", response)
#     respone = regexmatch !== nothing ? response = replace(response, match=>"Actinput:") : response
#     response = replace(response, "actinput:"=>"Actinput:")

#     println("")
#     @show actor_response = response

#     headerToDetect = ["Plan $(a.attempt):",
#                       "Self-awareness $latestTask:", 
#                       "Thought $latestTask:", 
#                       "Act $latestTask:", 
#                       "Actinput $latestTask:", 
#                       "Obs $latestTask:",
#                       "Check $latestTask:",]
#     headers = detectCharacters(response, headerToDetect)
#     chunkedtext = chunktext(response, headers) 

#     # assuming length more than 10 character means LLM has valid thinking
#     check_1 = haskey(chunkedtext, "Thought $latestTask:")
#     check_2 = haskey(chunkedtext, "Act $latestTask:")
#     check_3 = haskey(chunkedtext, "Actinput $latestTask:")

#     # check for a valid toolname
#     check_4 = false
#     for i in toolslist
#       if occursin(i, chunkedtext["Act $latestTask:"])
#         check_4 = true
#         break
#       end
#     end

#     # check for empty Thought
#     check_5 = length(chunkedtext["Thought $latestTask:"]) > 5
#     # check for empty Actinput
#     check_6 = nothing
#     try
#       check_6 = length(chunkedtext["Actinput $latestTask:"]) > 5  
#     catch
#       println("")
#       @show response
#       println("")
#       @show chunkedtext
#       a.memory[:chunkedtext] = chunkedtext
#     end
    

#     # check whether the act has valid json
#     check_7 = true
#     if occursin('{', response)
#       try
#         act = GeneralUtils.getStringBetweenCharacters(response, '{', '}', endCharLocation="end")
#         act = JSON3.read(act)
#         check_7 = true
#       catch
#         check_7 = false
#       end
#     end

#     # print all check_1 to check_6
#     println("check_1: $check_1, check_2: $check_2, check_3: $check_3, check_4: $check_4, 
#             check_5: $check_5, check_6: $check_6, check_7: $check_7")
#     if check_1 && check_2 && check_3 && check_4 && check_5 && check_6 && check_7
      
#       #TODO paraphrase selfaware
#       break
#     end
#     @show retrying_actor = response
#   end
  
#   toolname = toolNameBeingCalled(chunkedtext["Act $latestTask:"], a.tools)

#   # change trailing number to continue a.memory[:shortterm]
#   headerToDetect = ["Question:", "Plan:", "Self-awareness:", "Thought:", 
#                       "Act:", "Actinput:", "Obs:", "...", 
#                       "Answer:", "Conclusion:", "Summary:"]
#   response = replaceHeaders(response, headerToDetect, latestTask)
#   println("")
#   @show actor_response_1 = response
#   headerToDetect = ["Plan $(a.attempt):",
#                       "Thought $latestTask:", 
#                       "Act $latestTask:", 
#                       "Actinput $latestTask:", 
#                       "Obs $latestTask:",
#                       "Check $latestTask:",]
#   headers = detectCharacters(response, headerToDetect)
#   chunkedtext = chunktext(response, headers)
#   println("")
#   @show chunkedtext
  
  
#   toolinput = chunkedtext["Actinput $latestTask:"]


#   # because tools has JSON input but sometime LLM output is not JSON, we need to check.
#   if occursin("{", toolinput)
#     act = GeneralUtils.getStringBetweenCharacters(response, '{', '}', endCharLocation="end")
#     act = copy(JSON3.read(act))
#     chunkedtext["Actinput $latestTask:"] = JSON3.write(act[Symbol(toolname)])
#     toolinput = act[Symbol(toolname)]
#   end
  

#   chunkedtext["Act $latestTask:"] = toolname
  
#   return (toolname=toolname, toolinput=toolinput, chunkedtext=chunkedtext, selfaware=selfaware)
# end

function actor_mistral_openorca(a::agentReflex, selfaware=nothing)

  toolslist = []
  toolnames = ""
  toollines = ""
  for (toolname, v) in a.tools
    toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
    toollines *= toolline
    toolnames *= "$toolname, "
    push!(toolslist, toolname)
  end

  thought = "Thought: you should always think about what to do according to the plan (pay attention to correct numeral calculation and commonsense and do one thing at a time.)"
  startword = "Thought:"
  if selfaware !== nothing
    "
    Thought: based on what you know, you should focus on what you need to improve first then follow your plan to decide what to do next. (P.S. 1) let's think a single step. 2) pay attention to correct numeral calculation and commonsense.)
    "
  end
                        # your should request the missing information first before making a decision
  aboutYourself =
  """
    Your name is $(a.agentName)
    $(a.roles[a.role])  
  """

  winestocksearchresult = false
  if haskey(a.memory, :winestock) && a.memory[:winestock] !== nothing
    # winestocksearchresult = 
    # """
    # <winestock search result>
    # $(a.memory[:winestock])
    # </winestock search result>
    # """
    winestocksearchresult = true
  end

  keywordmemory = ""
  for (k, v) in a.memory[:keyword]
    if v === nothing
      keywordmemory *= "- I don't know the user preferred $k yet \n"
    else  
      keywordmemory *= "- The user preferred $k is $v \n"
    end
  end

  """
  - Car type is SUV
  - Brand is Lexus
  - Price is 20k dollar
  - No info on the car color yet 
  - No info on the financing method yet
  - Luxury level is high
  {\"car type\": "SUV",\"brand\":\"Lexus\",\"price\":\"200000\",\"color\": null,\"financing method\": null, \"luxury level\":\"high\"}
  Thought: Based on the recall, you must address what you don't know first if you have one. Otherwise, you must think about what is the immediate next step to do according to the plan
  Recall: repeat each of the user info, if you didn't know (indicate by null) say, which info I didn't know
  """

  prompt_1 = 
  """
    <s>
    <|system|>
      <About yourself>
        Your name is $(a.agentName)
        $(a.roles[a.role])
      </About yourself>
      <You have access to the following tools>
        $toollines
      </You have access to the following tools>
      <Your job> 
        Use the following format:
        Thought: there are some user's info I still don't know, I must address it as the immediate next step. (PS. 1. pay attention to correct numeral calculation and commonsense 2. you must assume nothing 3. do not ask for confirmation)
        Act: based on your thought what action to choose?, must be one of [{toolnames}].
        Actinput: your input to the action using JSON format (pay attention to the tool's input)
        Obs: observed result of the action
      </Your job>
    </|system|>
    </s>
    <My plan>
      $(a.memory[:shortterm]["Plan 1:"])
    </My plan>
    <User info>
    $keywordmemory
    </User info>
    $winestocksearchresult
    <|assistant|>
    Thought:
  """

  prompt_2 = 
  """
    <s>
    <|system|>
      <About yourself>
        Your name is $(a.agentName)
        $(a.roles[a.role])
      </About yourself>
      <You have access to the following tools>
        $toollines
      </You have access to the following tools>
      <Your job> 
        Use the following format:
        Thought: you always check your plan and explicitly state what you are going to do next. (PS. 1. let's think only one thing at a time. 2. pay attention to correct numeral calculation and commonsense.)
        Act: based on your thought what action to choose?, must be one of [{toolnames}].
        Actinput: your input to the action using JSON format (pay attention to the tool's input)
        Obs: observed result of the action
      </Your job>
    </|system|>
    </s>
    $(dictToString(a.memory[:shortterm]))
    <|assistant|>
    Thought:
  """

  prompt = prompt_1
  if nothing ∉ values(a.memory[:keyword])
    prompt = prompt_2
  end

  prompt = replace(prompt, "{toolnames}" => toolnames)

  println("")
  @show actor_prompt = prompt

  response = nothing
  chunkedtext = nothing
  latestTask = nothing

  while true  # while Thought or Act is empty, run actor again

    response = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=0.4, timeout=300,
                                stopword=["Thought:", "Obs:", "<|system|>", "</s>", "<|end|>"],
                                seed=rand(1000000:2000000))
    println("")
    @show actor_raw = response

    response = splittext(response, ["/n/n", "END", "End","obs", "Obs", "<|im_end|>"])
    response = split(response, "<|")[1]
    response = split(response, "</")[1]
    # response = split(response, "Thought:")[end]
    
    latestTask = shortMemLatestTask(a.memory[:shortterm]) +1

    response = startword * response

    headerToDetect = ["Plan:", "Self-awareness:", "Thought:", 
                      "Act:", "Actinput:", "Obs:", 
                      "Answer:", "Conclusion:", "Summary:"]
    
    # replace headers with headers with correct attempt and task number
    response = replaceHeaders(response, headerToDetect, latestTask)  
    response = split(response, "<|")[1]
    response = split(response, "</")[1]
    
    # sometime LLM use wrong keyword. use regex to detect "actinput5:" and replace it with "Actinput"
    regexmatch = match(r"actinput\d+:", response)
    respone = regexmatch !== nothing ? response = replace(response, match=>"Actinput:") : response
    response = replace(response, "actinput:"=>"Actinput:")

    println("")
    @show actor_response = response

    headerToDetect = ["Plan $(a.attempt):",
                      "Self-awareness $latestTask:", 
                      "Thought $latestTask:", 
                      "Act $latestTask:", 
                      "Actinput $latestTask:", 
                      "Obs $latestTask:",
                      "Check $latestTask:",]
    headers = detectCharacters(response, headerToDetect)
    chunkedtext = chunktext(response, headers) 

    # assuming length more than 10 character means LLM has valid thinking
    check_1 = haskey(chunkedtext, "Thought $latestTask:")
    check_2 = haskey(chunkedtext, "Act $latestTask:")
    check_3 = haskey(chunkedtext, "Actinput $latestTask:")

    # check for a valid toolname
    check_4 = false
    for i in toolslist
      if occursin(i, chunkedtext["Act $latestTask:"])
        check_4 = true
        break
      end
    end

    # check for empty Thought
    check_5 = length(chunkedtext["Thought $latestTask:"]) > 5
    # check for empty Actinput
    check_6 = nothing
    try
      check_6 = length(chunkedtext["Actinput $latestTask:"]) > 5  
    catch
      println("")
      @show response
      println("")
      @show chunkedtext
      a.memory[:chunkedtext] = chunkedtext
    end
    

    # check whether the act has valid json
    check_7 = true
    try
      act = GeneralUtils.getStringBetweenCharacters(response, '{', '}', endCharLocation="end")
      println("")
      @show act
      act = JSON3.read(act)
      check_7 = true
    catch
      check_7 = false
    end

    if check_1 && check_2 && check_3 && check_4 && check_5 && check_6 && check_7
      break
    end
    # print all check_1 to check_6
    println("")
    println("check_1: $check_1, check_2: $check_2, check_3: $check_3, check_4: $check_4, 
            check_5: $check_5, check_6: $check_6, check_7: $check_7")
    @show retrying_actor = response
  end
  
  toolname = toolNameBeingCalled(chunkedtext["Act $latestTask:"], a.tools)

  # change trailing number to continue a.memory[:shortterm]
  headerToDetect = ["Question:", "Plan:", "Self-awareness:", "Thought:", 
                      "Act:", "Actinput:", "Obs:", "...", 
                      "Answer:", "Conclusion:", "Summary:"]
  response = replaceHeaders(response, headerToDetect, latestTask)
  println("")
  @show actor_response_1 = response
  headerToDetect = ["Plan $(a.attempt):",
                      "Thought $latestTask:", 
                      "Act $latestTask:", 
                      "Actinput $latestTask:", 
                      "Obs $latestTask:",
                      "Check $latestTask:",]
  headers = detectCharacters(response, headerToDetect)
  chunkedtext = chunktext(response, headers)
  chunkedtext["Act $latestTask:"] = toolname

  println("")
  @show chunkedtext
    
  toolinput = chunkedtext["Actinput $latestTask:"]

  # # because tools has JSON input but sometime LLM output is not JSON, we need to check.
  # if occursin("{", toolinput)
  #   act = GeneralUtils.getStringBetweenCharacters(response, '{', '}', endCharLocation="end")
  #   act = copy(JSON3.read(act)) 

  #   println("")
  #   @show act

  #   chunkedtext["Actinput $latestTask:"] = JSON3.write(act[Symbol(toolname)])
  #   a.memory[:c] = chunkedtext
  #   toolinput = act[Symbol(toolname)]
  # end
  
  
  
  return (toolname=toolname, toolinput=toolinput, chunkedtext=chunkedtext, selfaware=selfaware)
end


""" Chat with llm.

  Arguments\n
  -----
    a::agent
      an agent
    
  Return\n
  -----
    None

  Example\n
  -----
  ```jldoctest
  julia> using JSON3, UUIDs, Dates, FileIO, MQTTClient, ChatAgent
  julia> const mqttBroker = "mqtt.yiem.cc"
  julia> mqttclient, connection = MakeConnection(mqttBroker, 1883)
  julia> tools=Dict( # update input format
          "askbox"=>Dict(
            :description => "<askbox tool description>Useful for when you need to ask the user for more context. Do not ask the user their own question.</askbox tool description>",
            :input => "<input>Input is a text in JSON format.</input><input example>{\"Q1\": \"How are you doing?\", \"Q2\": \"How may I help you?\"}</input example>",
            :output => "" ,
            :func => nothing,
          ),
        )
  julia> msgMeta = Dict(
    :msgPurpose=> "updateStatus",
    :from=> "agent",       
    :to=> "llmAI",
    :requestresponse=> "request",
    :sendto=> "",      # destination topic
    :replyTo=> "agent/api/v0.1.0/txt/response",     # requester ask responseer to send reply to this topic
    :repondToMsgId=> "",   # responseer is responseing to this msg id
    :taskstatus=> "",       # "complete", "fail", "waiting" or other status
    :timestamp=> Dates.now(),
    :msgId=> "$(uuid4())",
  )
  julia> a = ChatAgent.agentReflex(
          "Jene",
          mqttclient,
          msgMeta,
          agentConfigTopic, # I need a function to send msg to config topic to get load balancer
          role=:sommelier,
          tools=tools
        )
  julia> newAgent = ChatAgent.agentReact(agent)
  julia> response = ChatAgent.conversation(newAgent, "Hi! how are you?")
  ```

  Signature\n
  -----
"""
function conversation(a::agentReflex, usermsg::String; attemptlimit::Int=3)
  a.attemptlimit = attemptlimit
  workstate = nothing
  response = nothing

  _ = addNewMessage(a, "user", usermsg)
  isuseplan = isUsePlans(a)
  # newinfo = extractinfo(a, usermsg)
  # a.env = newinfo !== nothing ? updateEnvState(a, newinfo) : a.env
  @show isuseplan

  if isuseplan # use plan before responding
    if haskey(a.memory[:shortterm], "User:") == false #TODO should change role if user want to buy wine.
      a.memory[:shortterm]["User:"] = usermsg
    end
    workstate, response = work(a)
  end

  # if LLM using askbox, use returning msg form askbox as conversation response
  if workstate == "askbox" || workstate == "formulatedUserResponse"
    #TODO paraphrase msg so that it is human friendlier word.
  else
    response = chat_mistral_openorca(a)
    response = split(response, "\n\n")[1]
    response = split(response, "\n\n")[1]
  end
  
  response = removeTrailingCharacters(response)
  _ = addNewMessage(a, "assistant", response)

  return response
end






"""
  Continuously run llm functions except when llm is getting Answer: or askbox.
  There are many work() depend on thinking mode.
""" 
function work(a::agentReflex)
  workstate = nothing
  response = nothing
  
  # user answering LLM -> Obs
  if length(a.memory[:shortterm]) > 1
    latestTask = shortMemLatestTask(a.memory[:shortterm])
    if haskey(a.memory[:shortterm], "Act $latestTask:")
      if occursin("askbox", a.memory[:shortterm]["Act $latestTask:"])
        a.memory[:shortterm]["Obs $latestTask:"] = "(user response) " * a.messages[end][:content]
      end
    end
  end
  
  while true  # Work loop
    objective = nothing

    # make new plan
    if !haskey(a.memory[:shortterm], "Plan 1:")
      plan = planner_mistral_openorca(a)
      a.memory[:shortterm]["Plan $(a.attempt):"] = plan
      a.memory[:log]["Plan $(a.attempt):"] = plan
      a.task = 1  # reset because new plan is created
      println("")
      @show plan
      println("")
      @show a.attempt
    end

    if a.attempt <= a.attemptlimit
      toolname = nothing
      toolinput = nothing

      # enter actor loop
      actorstate, msgToUser = actor(a)

      if actorstate == "askbox"
        response = msgToUser
        workstate = actorstate
        break
      elseif actorstate == "formulateFinalResponse"
        println("all tasks done")

        response = formulateUserResponse(a)

        # println("")
        # formulatedresponse = response
        # @show formulatedresponse
        
        # a.memory[:shortterm]["response $(a.attempt):"] = response
        # a.memory[:log]["response $(a.attempt):"] = response

        # # evaluate. if score > 6/10 good enough.
        # guideline = writeEvaluationGuideline(a)

        # println("")
        # @show guideline

        # score = grading(a, guideline, response)
        # @show score

        score = 9 #TODO let grading() make a score
        if score > 5 # good enough answer
          println("")
          formulatedresponse_final = response
          @show formulatedresponse_final
          workstate = "formulatedUserResponse"
          a.memory[:shortterm] = OrderedDict{String, Any}()
          a.memory[:log] = OrderedDict{String, Any}()
          break
        else  # self evaluate and reflect then try again
          analysis = analyze(a)
          println("")
          @show analysis
          
          lessonwithcontext = selfReflext(a, analysis)
          
          println("")
          @show lessonwithcontext

          headerToDetect = ["Lesson:", "Context:", ]
          headers = detectCharacters(lessonwithcontext, headerToDetect)
          chunkedtext = chunktext(lessonwithcontext, headers)

          a.memory[:longterm][chunkedtext["Context:"]] = chunkedtext["Lesson:"]
          a.attempt += 1
          a.task = 0
          a.memory[:shortterm] = OrderedDict{String, Any}()
          a.memory[:log] = OrderedDict{String, Any}()
          println("")
          println("RETRY $(a.attempt +1)")
          println("")
        end
      else
        error("undefied condition, actorstate $actorstate $(@__LINE__)")
        break
      end
    else
      error("attempt limit reach")
      break
    end
  end

  # good enough answer
  return  workstate, response
end


"""
  Actor function. 
  
  Arguments:
    a, one of ChatAgent's agent.
    plan, a task by task plan to response
  
  Return:
    case 1) if actor complete the plan successfully.
      actorState = "all tasks done" inidicates that all task in plan were done.
      msgToUser = nothing.
    case 2) if actor needs to talk to user for more context
    actorState = "askbox"
    msgToUser = "message from assistant to user"

"""
function actor(a::agentReflex)

  actorState = nothing
  msgToUser = nothing

  
    # totaltasks = checkTotalTaskInPlan(a)
  

  while true  # Actor loop
    # check whether the current task is completed, skip evaluation if memory has only "Plan 1:"
    # taskrecap = ""
    # if length(keys(a.memory[:shortterm])) != 1
    #   taskrecap = recap(a)
    # end
    # println("")
    # @show taskrecap
    latestTask = shortMemLatestTask(a.memory[:shortterm]) +1
    println(">>> working")
    # work
    selfaware = nothing
    if length(a.memory[:shortterm]) > 2 # must have User:, Plan:, Thought:, Act:, Actinput: already
      selfaware = selfAwareness(a)
      # result = keywordMemoryToPlanMatching(a)
    end

    actorResult = actor_mistral_openorca(a, selfaware)
    toolname, toolinput, chunkedtext, selfaware = actorResult
    println("")
    @show toolname
    @show toolinput
    println(typeof(toolinput))

    if toolname == "askbox"  # chat with user
      msgToUser = askbox(toolinput)
      actorState = toolname

      # add only a single Q1 to memory because LLM need to ask the user only 1 question at a time
      latestTask = shortMemLatestTask(a.memory[:shortterm]) +1
      chunkedtext["Actinput $latestTask:"] = msgToUser
      addShortMem!(a.memory[:shortterm], chunkedtext)

      break
    elseif toolname == "finalanswer"
      addShortMem!(a.memory[:shortterm], chunkedtext)
      println(">>> already done")
      actorState = "formulateFinalResponse"

      break
    else  # function call
      addShortMem!(a.memory[:shortterm], chunkedtext)
      f = a.tools[toolname][:func]                
      toolresult = f(a, actorResult)
      @show toolresult
      if toolname == "winestock"
        a.memory[:winestock] = toolresult
      end
      a.memory[:shortterm]["Obs $latestTask:"] = toolresult
      a.memory[:log]["Obs $latestTask:"] = toolresult
    end    
  end

  return actorState, msgToUser
end



""" Write evaluation guideline.

  Arguments:
    a, one of ChatAgent's agent.
    usermsg, stimulus e.g. question, task and etc.

  Return:
    An evaluation guideline used to guage AI's work.

  Example:

  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> usermsg = "What's AMD latest product?"
  "
  julia> evaluationGuideLine = writeEvaluationGuideline(agent, usermsg)
  ``` 
"""
function writeEvaluationGuideline(a::agentReflex)
  prompt =
  """
    <|system|>
    $(a.roles[a.role])
      <You have access to the following tools>
      askbox: Useful for when you need to ask a customer for more context. Input should be a conversation to customer.
      wikisearch: Useful for when you need to search an encyclopedia Input is keywords and not a question.
      </You have access to the following tools>
      <Your job>
      1. Write an evaluation guideline for wine recommendation in order to be able to evaluate your response.
      2. An example of what the response should be.
      </Your job>
    </s>
    <|assistant|>

  """

  response = sendReceivePrompt(a, prompt)
  return response
end



""" Determine a score out of 10 according to evaluation guideline.

  Arguments:
    a, one of ChatAgent's agent.
    guidelines, an evaluation guideline.
    shorttermMemory, a short term memory that logs what happened.

  Return:
    A score out of 10 based on guideline.

  Example:

  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> shorttermMemory = OrderedDict{String, Any}(
    "user"        => "What's the latest AMD GPU?",
    "Plan 1:"     => "   To answer this question, I will need to search for the latest AMD GPU using the wikisearch tool.\n",
    "Act 1:"      => " wikisearch\n",
    "Actinput 1:" => " amd gpu latest\n",
    "Obs 1:"      => "No info available for your search query.",
    "Act 2:"      => "  wikisearch\n",
    "Actinput 2:" => "  amd graphics card latest\n",
    "Obs 2:"      => "No info available for your search query.")
  julia> guideline = "\nEvaluation Guideline:\n1. Check if the user's question has been understood correctly.\n2. Evaluate the tasks taken to provide the information requested by the user.\n3. Assess whether the correct tools were used for the task.\n4. Determine if the user's request was successfully fulfilled.\n5. Identify any potential improvements or alternative approaches that could be used in the future.\n\nThe response should include:\n1. A clear understanding of the user's question.\n2. The tasks taken to provide the information requested by the user.\n3. An evaluation of whether the correct tools were used for the task.\n4. A confirmation or explanation if the user's request was successfully fulfilled.\n5. Any potential improvements or alternative approaches that could be used in the future."
  julia> score = grading(agent, guideline, shorttermMemory)
  ```
"""
function grading(a, guideline::T, text::T)  where {T<:AbstractString}
  prompt =
  """
    <|system|>
      <You have access to the following tools>
      askbox: Useful for when you need to ask a customer for more context. Input should be a conversation to customer.
      wikisearch: Useful for when you need to search an encyclopedia Input is keywords and not a question.
      </You have access to the following tools>
      <Guideline>
      $guideline
      </Guideline>
      <Your response>
      $text
      </Your response>

      <Your job>
      Evaluate your response using the evaluation guideline then give yourself a score out of 9 for your response.
      </Your job>
      <Example>
      {"Evaluate": "My response is detailed with good comparison between options.", "Score": 6}
      </Example>
    </s>
    <|assistant|>
    {
  """
  println("")
  prompt_grading = prompt
  @show prompt_grading
  println("")
  score = nothing
  while true
    response = sendReceivePrompt(a, prompt, timeout=180)
    try
      response = "{" * split(response, "}")[1] * "}"
      @show response
      @show jsonresponse = JSON3.read(response)
      score = jsonresponse["Score"]
      break
    catch
      println("retry grading")
    end
  end

  return score
end



""" Analize work.

  Arguments:
    a, one of ChatAgent's agent.

  Return:
    A report of analized work.

  Example:

  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> shorttermMemory = OrderedDict{String, Any}(
    "user:"        => "What's the latest AMD GPU?",
    "Plan 1:"     => "   To answer this question, I will need to search for the latest AMD GPU using the wikisearch tool.\n",
    "Act 1:"      => " wikisearch\n",
    "Actinput 1:" => " amd gpu latest\n",
    "Obs 1:"      => "No info available for your search query.",
    "Act 2:"      => "  wikisearch\n",
    "Actinput 2:" => "  amd graphics card latest\n",
    "Obs 2:"      => "No info available for your search query.")
  julia> report = analyze(agent, shorttermMemory)
  ```
"""
function analyze(a)
  shorttermMemory = dictToString(a.memory[:shortterm])
  prompt =
  """
    <|system|>
      <you have access to the following tools>
      askbox: Useful for when you need to ask a customer for more context. Input should be a conversation to customer.
      wikisearch: Useful for when you need to search an encyclopedia Input is keywords and not a question.
      </you have access to the following tools>
      <your earlier work>
      $shorttermMemory
      </your earlier work>
      <your job>
      You job is to do each of the following in detail to analize your work.
      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. What could you do to improve the response?
      </your job>
    </s>
    <|assistant|>
    
  """

  response = sendReceivePrompt(a, prompt, max_tokens=1024, timeout=180)

  return response
end


""" Write a lesson drawn from evaluation.

  Arguments:
    a, one of ChatAgent's agent.
    report, a report resulted from analyzing shorttermMemory

  Return:
    A lesson.

  Example:

  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> report = 
    "What happened: I tried to search for AMD's latest product using the wikisearch tool, 
    but no information was available in the search results.
    Cause and effect relationships:
      1. Searching \"AMD latest product\" -> No info available.
      2. Searching \"most recent product release\" -> No info available.
      3. Searching \"latest product\" -> No info available.
    Analysis of each relationship:
      1. The search for \"AMD latest product\" did not provide any information because the wikisearch tool could not find relevant results for that query.
      2. The search for \"most recent product release\" also did not yield any results, indicating that there might be no recent product releases available or that the information is not accessible through the wikisearch tool.
      3. The search for \"latest product\" similarly resulted in no information being found, suggesting that either the latest product is not listed on the encyclopedia or it is not easily identifiable using the wikisearch tool.
    Improvements: To improve the response, I could try searching for AMD's products on a different 
      source or search engine to find the most recent product release. Additionally, I could ask 
      the user for more context or clarify their question to better understand what they are 
      looking for."
  julia> lesson = selfReflext(agent, report)
  ```
"""
function selfReflext(a, analysis::T) where {T<:AbstractString}
  prompt =
  """
    <|system|>
      <You have access to the following tools>
      askbox: Useful for when you need to ask a customer for more context. Input should be a conversation to customer.
      wikisearch: Useful for when you need to search an encyclopedia Input is keywords and not a question.
      </You have access to the following tools>
      <your report>
      $analysis
      </your report>
      <your job are>
      1. Lesson: what lesson could you learn from your report?.
      2. Context: what is the context this lesson could apply to?
      </your job are>
    </s>
    <|assistant|>
    
  """

  response = sendReceivePrompt(a, prompt, max_tokens=1024)
  return response
end


""" Formulate a response from work for user's stimulus.

  Arguments:
    a, one of ChatAgent's agent.

  Return:
    A response for user's stimulus.

  Example:
  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> shorttermMemory = OrderedDict{String, Any}(
    "user:"        => "What's the latest AMD GPU?",
    "Plan 1:"     => "   To answer this question, I will need to search for the latest AMD GPU using the wikisearch tool.\n",
    "Act 1:"      => " wikisearch\n",
    "Actinput 1:" => " amd gpu latest\n",
    "Obs 1:"      => "No info available for your search query.",
    "Act 2:"      => "  wikisearch\n",
    "Actinput 2:" => "  amd graphics card latest\n",
    "Obs 2:"      => "No info available for your search query.")

  julia> report = formulateUserResponse(agent, shorttermMemory)
  ```
"""
function formulateUserResponse(a)
  conversation = messagesToString_nomark(a.messages, addressAIas="I")
  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 talk with the user>
      $conversation
      </your talk with the user>

      <your earlier work>
      $work
      </your earlier work>
    
      <your job>      
      Based on your talk with the user and your work, present a response that compares and justifies each option in great detail to the user.
      </your job>   
    </s>
    <|assistant|>
    Recommendation:
  """
  response = sendReceivePrompt(a, prompt, max_tokens=1024, timeout=300)
  return response
end

# function formulateUserResponse(a)
#   conversation = messagesToString_nomark(a.messages, addressAIas="I")
#   work = dictToString(a.memory[:shortterm])

#   prompt =
#   """
#     <|system|>
#     Symbol:
#     Plan: a plan
#     Thought: your thought
#     Act: the action you took
#     Actinput: the input to the action
#     Obs: the result of the action

#     Your talk with the user:
#     $conversation

#     Your work:
#     $work
    
#     From your talk with the user and your work, formulate a response for the user .
#     </s>
#     <|assistant|>
#     response:
#   """
#   response = sendReceivePrompt(a, prompt)
#   return response
# end


""" Extract important info from text into key-value pair text.

  Arguments:
    a, one of ChatAgent's agent.
    text, a text you want to extract info

  Return:
    key-value pair text.

  Example:
  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> text = "We are holding a wedding party at the beach."
  julia> extract(agent, text)
  "location=beach, event=wedding party"
  ```
"""
function extractinfo(a, text::T)  where {T<:AbstractString}
  # determine whether there are any important info in an input text
  prompt =
  """
    <|system|>
      <user's message>
      $text
      </user's message>
      <your job>
      Determine whether there are important info in the user's message. Answer: {Yes/No/Not sure}
      </your job>
    </s>
    Answer: 
  """
  response = sendReceivePrompt(a, prompt, temperature=0.0)
  if occursin("Yes", response)
    prompt =
    """
      <|system|>
        <user's message>
        $text
        </user's message>
        <your job>
        Extract important info from the user's message into keys and values using this format: key=value,.
        p.s.1 you can extract many key-value pairs.
        </your job>
      </s>

    """

    response = sendReceivePrompt(a, prompt, temperature=0.0)
    return response
  else
    return nothing
  end
end 


""" Update important info from key-value pair text into another key-value pair text.

  Arguments:
    a, one of ChatAgent's agent
    text, a key-value pair text

  Return:
    updated key-value pair text

  Example:
  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> currentinfo = "location=beach, event=wedding party"
  julia> newinfo = "wine_type=full body, dry and medium tannin\nprice_range=50 dollars"
  julia> updateEnvState(agent, currentinfo, newinfo)
  "  location=beach, event=wedding party, wine_type=full body, dry and medium tannin, price_range=50 dollars"
  ```
"""
function updateEnvState(a, newinfo)
  prompt =
  """
    <|im_start|>system
    Current state:
    $(a.env)

    New info:
    $newinfo
    
    Your job is to update or add information from new info into the current state which use key-value format.
    <|im_end|>
    Updated Current State:\n
  """

  response = sendReceivePrompt(a, prompt, temperature=0.0)
  return response
end



""" Determine whether LLM should go to next task.

  Arguments:
    a, one of ChatAgent's agent.

  Return:
    "Yes" or "no" decision to go next task.

  Example:
  ```jldoctest
  julia> using ChatAgent
  julia> agent = ChatAgent.agentReflex("Jene")
  julia> shorttermMemory = OrderedDict{String, Any}(
    "user:"        => "What's the latest AMD GPU?",
    "Plan 1:"     => "   To answer this question, I will need to search for the latest AMD GPU using the wikisearch tool.\n",
    "Act 1:"      => " wikisearch\n",
    "Actinput 1:" => " amd gpu latest\n",
    "Obs 1:"      => "No info available for your search query.",
    "Act 2:"      => "  wikisearch\n",
    "Actinput 2:" => "  amd graphics card latest\n",
    "Obs 2:"      => "No info available for your search query.")

  julia> decision = checkTaskCompletion(agent)
  "Yes"
  ```
"""
function checkTaskCompletion(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>
      Check whether each task of your plan has been completed.
      </Your job>

      <Example 1>
        Task 1 of the plan: Ask user about their preferred topping of a pizza.
        Obs: I love Malvasia.
        assistant: After checking all my work's observed results, I can't find any relevant info that the user tell me what is their preferred topping in pizza. Thus, task 1 isn't done yet.
        Task 2 of the plan: Ask user if they have any preferred type of car. 
        Obs: I like a semi truck. 
        assistant: After checking all my work's observed results, I found that the user like a semi truck. Thus, task 2 is done.
        Task 3 of the plan: How much you are looking to spend for a new house?
        Obs: 50K THB. 
        assistant: After checking all my work's observed results, I found that the user have a budget of 50,000 Baht. Thus, task 3 is done.
      </Example 1>

      Let's think step by step.
    </s>
    <|assistant|> After 
  """
  response = nothing
  _response = nothing
  _response = sendReceivePrompt(a, prompt, max_tokens=1024)
  @show checkTaskCompletion_raw = _response
  _response = split(_response, "</")[1]
  _response = split(_response, "\n\n")[1]
  # response = "I " * split(_response, "{")[1]  # sometime response have more than 1 {answer: done}

  decision = nothing
  # if occursin("done", response)
  #   decision = true
  # else
  #   decision = false
  # end
  
  return decision, response
end


function recap(a)
  # stimulus = a.memory[:shortterm]["user:"]
  getonlykeys = ["Actinput", "Obs"]
  worknoplan = similar(a.memory[:shortterm])
  for (k, v) in a.memory[:shortterm]
    count = 0
    for i in getonlykeys
      if occursin(i, k)
        count += 1
      end
    end
    if count != 0
      worknoplan[k] = v 
    end
  end

  work = dictToString(worknoplan)

  toolnames = ""
  toollines = ""
  for (toolname, v) in a.tools
    toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
    toollines *= toolline
    toolnames *= "$toolname, "
  end

  # 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: 
  # """



  prompt =
  """
    <|system|>
    $(a.roles[a.role])
      <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>
      <You have access to the following tools>
      $toollines
      </You have access to the following tools>
      <Your earlier work>
      $work
      </Your earlier work>
      <Your job>
      Extract info: extract each info in details from your earlier work according to the Actinput context.
      </Your job>

      Let's think step by step.
    </s>
    <|assistant|>
    Extracted info: 
  """
  aware = "Self-awareness: map the info from the recap to the plan's tasks then state your mapping."
  response = sendReceivePrompt(a, prompt, max_tokens=1024, temperature=0.0)
  response = split(response, "</")[1]
  response = split(response, "<|")[1]
  response = split(response, "\n\n")[1]
  
  return response
end

""" Direct conversation is not an agent, messages does not pass through logic loop
  but goes directly to LLM.
"""
function directconversation(a::agentReflex, usermsg::String)
  response = nothing

  _ = addNewMessage(a, "user", usermsg)

  response = chat_mistral_openorca(a)
  response = removeTrailingCharacters(response)
  _ = addNewMessage(a, "assistant", response)
  return response
end


""" Convert keyword memory into a string.

  Arguments\n
    a : one of ChatAgent's agent.
    keywordmemory : a dictionary of keyword memory.

  Return\n
    result : a string of LLM readout from keyword memory

  Example:
  ```jldoctest
  julia> using ChatAgent
  julia> a = ChatAgent.agentReflex("Jene")
  julia> keywordmemory = OrderedDict{String, Any}(
    "food type"       => nothing,
    "tannin level"    => "low to medium",
    "intensity level" => "medium-bodied",
    "acidity level"   => nothing,
    "price range"     => "fifteen dollars",
    "wine type"       => "Red",
    "sweetness level" => "dry",
    )

  julia> readout = readKeywordMemory(a, keywordmemory=keywordmemory)
  " - The user did not provide food type yet
    - The user prefers a low to medium tannin level
    - The user prefers a medium-bodied intensity level
    - The user did not provide acidity level yet
    - The user prefers price range is fifteen dollars
    - The user prefers a Red wine type
    - The user prefers a dry sweetness level"
  ```
"""
function readKeywordMemory(a; keywordmemory::Union{AbstractDict, Nothing}=nothing)

  keywordmemory = keywordmemory !== nothing ? keywordmemory : a.memory[:keyword]
  result = ""

  if !isempty(keywordmemory)
    new_keywordmemory = deepcopy(keywordmemory)
    @show keywordmemory


    # prepare reversed dict for pop! coz I need to preserve key order
    reversed_keywordmemory = Dict()
    while length(new_keywordmemory) > 0
      k, v = pop!(new_keywordmemory)
      reversed_keywordmemory[k] = v
    end

    while length(reversed_keywordmemory) > 0
      tempdict = OrderedDict()
      for i in 1:4
        if length(reversed_keywordmemory) == 0
          break
        else
          k, v = pop!(reversed_keywordmemory)
          tempdict[k] = v
        end
      end

      # ask LLM to read tempdict
      jsonstr = JSON3.write(tempdict)
      prompt = 
      """
      <s>
        <|system|>
          <About yourself>
          Your name is $(a.agentName)
          $(a.roles[a.role])
          </About yourself>
          <Your job> 
          Readout all the key and its value pairs in memory, one by one. Do not say anything else.
          </Your job>
        </|system|>
        <Example 1>
          <Memory>
            {\"car type\": "SUV",\"brand\":\"Lexus\",\"price\":\"20k dollar\",\"color\": null,\"financing method\": null, \"luxury level\":\"high\"}
          </Memory>
          <|assistant|>
            - Car type is SUV
            - Brand is Lexus
            - Price is 20k dollar
            - No info on the car color yet 
            - No info on the financing method yet
            - Luxury level is high
          </|assistant|>
        </Example 1>
      </s>
      <Memory>
      User preference: $jsonstr
      </Memory>
      <|assistant|>
      """

      response = sendReceivePrompt(a, prompt, max_tokens=512, temperature=0.0)
      response = split(response, "</|assistant|>")[1]

      # store LLM readout string to result
      result = result * response
    end
  end

  return result
end
























































end # module