ton/ChatAgent_v2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update

Browse Source
This commit is contained in:
tonaerospace
2023-12-19 16:30:32 +00:00
parent d90ff8d3fc
commit 59e3b3027f
3 changed files with 222 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files

252
src/interface.jl
View File

@@ -120,19 +120,21 @@ function chat_mistral_openorca(a::agentReflex)
"
"""
conversation = messagesToString(a.messages)
prompt =
"""
<|im_start|>system
<|system|>
$(a.roles[a.role])
Your earlier talk with the user:
$(a.earlierConversation)
<|im_end|>
$(messagesToString(a.messages))
<|/s|>
$conversation
<|im_start|>assistant
"""
response = sendReceivePrompt(a, prompt)
response = split(response, "\n\n")[1]
response = split(response, "<|im_end|>")[1]
return response
@@ -181,7 +183,7 @@ function planner_mistral_openorca(a::agentReflex)
assistant_plan_prompt =
"""
<|im_start|>system
<|system|>
$(a.roles[a.role])
The required info you need for wine recommendation:
- type of food: ask the user
@@ -190,41 +192,115 @@ function planner_mistral_openorca(a::agentReflex)
- wine price range: ask the user
- ambient temperature at the serving location: ask the user
- wines we have in stock
You provide a personalized recommendation of wine based on the user's info above by describing the benefits of each wine in detail.
You have access to the following tools:
$toollines
Your work:
Your earlier work:
$shorttermMemory
Use the following format:
Objective: the objective you intend to do
Aware: according to your work and your conversation with the user, ask yourself what info you have and what info you doesn't have?
Plan: first you should always think about the objective, the info you have, the info you doesn't have thoroughly then extract and devise a complete, step by step plan (pay attention to correct numeral calculation and commonsense).
Your task is to do the following:
Plan: first you should always think about your conversation with the user and your earlier work thoroughly then extract and devise a complete, step by step plan to achieve your objective (pay attention to correct numeral calculation and commonsense).
P.S.1 each step of the plan should be a single action.
P.S.2 ask the user if you don't have info.
<|im_end|>
<|/s|>
$conversation
<|im_start|>assistant
Objective:
<|assistant|>
Plan:
"""
#WORKING remove () in steps as LLM sometimes use for (addtional info) but interfere with updatePlan
result = sendReceivePrompt(a, assistant_plan_prompt, max_tokens=512, temperature=0.1)
plan = sendReceivePrompt(a, assistant_plan_prompt, max_tokens=512, temperature=0.1)
plan = split(plan, "<|user|>")[1]
plan = split(plan, "<|assistant|>")[1]
plan = split(plan, "\n\n")[1]
x = split(result, "<|im_end|>")[1]
@show x
x = split(x, "Step")[1]
x = split(x, "Plan:")
objective = x[1]
plan = x[2]
return objective, plan
return plan
end
# function planner_mistral_openorca(a::agentReflex)
# """
# general prompt format:
# "
# <|im_start|>system
# {role}
# {tools}
# {thinkingFormat}
# <|im_end|>
# {context}
# <|im_start|>user
# {usermsg}
# <|im_end|>
# <|im_start|>assistant
# "
# Note:
# {context} =
# "
# {earlierConversation}
# {env state}
# {shortterm memory}
# {longterm memory}
# "
# """
# conversation = messagesToString(a.messages)
# toollines = ""
# for (toolname, v) in a.tools
# if toolname ∉ [""]
# toolline = "$toolname: $(v[:description]) $(v[:input]) $(v[:output])\n"
# toollines *= toolline
# end
# end
# # skip objective and plan because LLM is going to generate new plan
# shorttermMemory = dictToString(a.memory[:shortterm], skiplist=["Objective:", "Plan 0:"])
# assistant_plan_prompt =
# """
# <|system|>
# $(a.roles[a.role])
# The required info you need for wine recommendation:
# - type of food: ask the user
# - occasion: ask the user
# - user's personal taste of wine: ask the user
# - wine price range: ask the user
# - ambient temperature at the serving location: ask the user
# - wines we have in stock
# You provide a personalized recommendation of wine based on the user's info above by describing the benefits of each wine in detail.
# You have access to the following tools:
# $toollines
# Your earlier work:
# $shorttermMemory
# Use the following format:
# Objective: what do you think the user needs?
# Plan: first you should always think about your conversation with the user and your earlier work thoroughly then extract and devise a complete, step by step plan (pay attention to correct numeral calculation and commonsense).
# P.S.1 each step of the plan should be a single action.
# P.S.2 ask the user if you don't have info.
# <|/s|>
# $conversation
# <|assistant|>
# Objective:
# """
# result = sendReceivePrompt(a, assistant_plan_prompt, max_tokens=512, temperature=0.1)
# x = split(result, "<|user|>")[1]
# x = split(result, "<|assistant|>")[1]
# x = split(x, "Step")[1]
# x = split(x, "Plan:")
# objective = x[1]
# plan = x[2]
# return objective, plan
# end
""" Update the current plan.
"""
function updatePlan(a::agentReflex)
@@ -242,7 +318,7 @@ function updatePlan(a::agentReflex)
prompt =
"""
<|im_start|>system
<|system|>
$(a.roles[a.role])
The required info you need for wine recommendation:
- wine price range: ask the user
@@ -266,13 +342,14 @@ function updatePlan(a::agentReflex)
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).
<|im_end|>
</s|>
Updated plan:
"""
result = sendReceivePrompt(a, prompt, max_tokens=512, temperature=0.1)
@show updatedPlan = result
a.memory[:shortterm]["Plan 0:"] = result
end
function actor_mistral_openorca(a::agentReflex)
@@ -323,13 +400,15 @@ function actor_mistral_openorca(a::agentReflex)
prompt =
"""
<|im_start|>system
<|system|>
$(a.roles[a.role])
You have access to the following tools:
$toollines
$(a.thinkingFormat[:actor])
<|im_end|>
Your earlier work:
$shorttermMemory
$(a.thinkingFormat[:actor])
<|/s|>
<|assistant|>
Thought $(a.step):
"""
prompt = replace(prompt, "{toolnames}" => toolnames)
@@ -342,7 +421,7 @@ function actor_mistral_openorca(a::agentReflex)
chunkedtext = nothing
tempcounter = 0.0
while true # while Thought is empty, run actor again
while true # while Thought or Act is empty, run actor again
tempcounter += 0.1
@show tempcounter
response = sendReceivePrompt(a, prompt, temperature=tempcounter)
@@ -375,8 +454,10 @@ function actor_mistral_openorca(a::agentReflex)
chunkedtext = chunktext(response, headers)
# assuming length more than 10 character means LLM has valid thinking
if length(chunkedtext["Thought $(a.step):"]) > 10
break
if haskey(chunkedtext, "Thought $(a.step):") && haskey(chunkedtext, "Act $(a.step):")
if length(chunkedtext["Thought $(a.step):"]) > 10 && length(chunkedtext["Act $(a.step):"]) > 10
break
end
end
end
@@ -483,12 +564,14 @@ end
function work(a::agentReflex)
workstate = nothing
response = nothing
#BUG there is no Obs 2:
# user answering LLM -> Obs
if haskey(a.memory[:shortterm], "Act $(a.step):")
if occursin("chatbox", a.memory[:shortterm]["Act $(a.step):"])
a.memory[:shortterm]["Obs $(a.step):"] = a.messages[end][:content]
if length(a.memory[:shortterm]) != 0
latest_step = dictLatestStep(a.memory[:shortterm])
if haskey(a.memory[:shortterm], "Act $latest_step:")
if occursin("chatbox", a.memory[:shortterm]["Act $latest_step:"])
a.memory[:shortterm]["Obs $latest_step:"] = a.messages[end][:content]
end
end
end
@@ -499,41 +582,15 @@ function work(a::agentReflex)
if a.attempt <= a.attemptlimit
toolname = nothing
toolinput = nothing
# if length(a.memory[:shortterm]) != 0
# updatePlan(a)
# @show updatedPlan = a.memory[:shortterm]["Plan 0:"]
# else
# objective, plan = planner_mistral_openorca(a)
# a.memory[:shortterm]["Objective:"] = objective
# a.memory[:shortterm]["Plan $(a.attempt):"] = plan
# a.memory[:log]["Plan $(a.attempt):"] = plan
# println("")
# @show objective
# @show plan
# end
objective, plan = planner_mistral_openorca(a)
a.memory[:shortterm]["Objective:"] = objective
plan = planner_mistral_openorca(a)
a.memory[:shortterm]["Plan $(a.attempt):"] = plan
a.memory[:log]["Plan $(a.attempt):"] = plan
a.step = 0 # reset because new plan is created
println("")
@show objective
@show plan
# sometimes LLM add not-need word I don't want
# plan = splittext(response, ["Step 1", "<|im_end|>", "Response", "Execution",
# "Result", "Recommendation", "My response"])
# plan = replace(plan, "Plan:"=>"")
# println("")
# @show plan
println("")
@show a.attempt
@@ -636,24 +693,25 @@ function actor(a::agentReflex)
while true # Actor loop
# decide whether to repeat step or do the next step
decision, reason = goNogo(a)
println("")
@show decision
@show reason
# a.memory[:shortterm]["Check $(a.step):"] = reason
decision = "Yes" # yes because a.step start at 0
if a.step != 0
decision, reason = goNogo(a)
end
if decision == "Yes" # in case there is a cancel, go straight to evaluation
a.step += 1
elseif decision == "No"
# repeat the latest step
elseif decision == "No" # repeat the latest step
a.memory[:shortterm] = removeHeaders(a.memory[:shortterm], a.step, ["Plan"])
a.memory[:log] = removeHeaders(a.memory[:log], a.step, ["Plan"])
println("repeating step $(a.step)")
else
error("undefined condition decision = $decision $(@__LINE__)")
end
@show a.step
#WORKING checkStepCompletion
iscomplete = checkStepCompletion(a::agentReflex)
if iscomplete
if a.step < totalsteps # the last step of the plan is responding, let work() do this part
toolname, toolinput = actor_mistral_openorca(a)
@@ -933,7 +991,7 @@ function formulateUserresponse(a)
prompt =
"""
<|im_start|>system
<|system|>
Symbol:
Plan: a plan
Thought: your thought
@@ -948,8 +1006,9 @@ function formulateUserresponse(a)
$work
From your talk with the user and your work, formulate a response for the user.
<|im_end|>
response:
<|/s|>
<|assistant|>
response:
"""
response = sendReceivePrompt(a, prompt)
return response
@@ -1080,7 +1139,7 @@ function goNogo(a)
prompt =
"""
<|im_start|>system
<|system|>
Symbol meaning:
Plan: a plan
Thought: your thought
@@ -1088,18 +1147,23 @@ function goNogo(a)
Actinput: the input to the action
Obs: the result of the action
Your work:
Your earlier work:
$work
Your job is to check whether step $(a.step) of your work is completed according to the plan and choose only one of the following choices.
choice 1: If you get what you intend to do and you are ready to do the next step of the plan say, "{Yes}". And what is the rationale behind the decision to do the next step?
choice 2: If you didn't get what you intend to do and you need to repeat the latest step say, "{No}". And what is the rationale behind the decision to repeat the latest step?
<|im_end|>
<|im_start|>assistant
Your job is to check whether step $(a.step) of your work is completed according to the plan.
So for instance the following:
step 2 of the plan: Ask user about the occasion type. But you can't find any relevant info of occasion type in your work.
assistant: Step 2 isn't done yet. {No}
step 5 of the plan: Ask user if they have any preference for the style of wine. And you found relevant info in your work such as the user like full-bodied wine.
assistant: Step 5 is done. {Yes}
</s|>
<|assistant|>
"""
response = sendReceivePrompt(a, prompt)
@show goNogo_response = response
decision = nothing
reason = nothing
if occursin("Yes", response)
@@ -1132,7 +1196,7 @@ function checkStepCompletion(a::agentReflex)
prompt =
"""
<|im_start|>system
<|system|>
Symbol meaning:
Plan: a plan
Thought: your thought
@@ -1144,8 +1208,8 @@ function checkStepCompletion(a::agentReflex)
$plan
What is step $(a.step) of the plan?
<|im_end|>
<|im_start|>assistant
<|/s|>
<|assistant|>
"""
@@ -1153,17 +1217,7 @@ function checkStepCompletion(a::agentReflex)
response = split(response, "<|im_end|>")[1]
# if occursin("N/A", response)
# response = replace(response, "(N/A)"=>"")
# elseif occursin(response, "(not specified, assume casual)")
# response = replace(response, "(not specified, assume casual)"=>"")
# else
# end
# mistral 7B already know info example: 2. Determine the occasion (wedding party).
if occursin("(", response) && occursin(")", response)
result = true
end
return result
end