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

update

Browse Source
This commit is contained in:
tonaerospace
2025-03-21 10:03:08 +07:00
parent 8fc3afe348
commit bfadd53033
5 changed files with 614 additions and 650 deletions
Download Patch File Download Diff File Expand all files Collapse all files

548
src/interface.jl
View File

@@ -1,6 +1,6 @@
module interface module interface
export addNewMessage, conversation, decisionMaker, evaluator, reflector, generatechat, export addNewMessage, conversation, decisionMaker, reflector, generatechat,
generalconversation, detectWineryName, generateSituationReport generalconversation, detectWineryName, generateSituationReport
using JSON3, DataStructures, Dates, UUIDs, HTTP, Random, PrettyPrinting, Serialization, using JSON3, DataStructures, Dates, UUIDs, HTTP, Random, PrettyPrinting, Serialization,
@@ -56,6 +56,8 @@ end
- `state::T2` - `state::T2`
a game state a game state
# Keyword Arguments
# Return # Return
- `thoughtDict::Dict` - `thoughtDict::Dict`
@@ -90,8 +92,6 @@ julia> output_thoughtDict = Dict(
# TODO # TODO
- [] update docstring - [] update docstring
- [x] implement the function
- [] implement RAG to pull similar experience
- [] use customerinfo - [] use customerinfo
- [] user storeinfo - [] user storeinfo
@@ -294,15 +294,13 @@ function decisionMaker(a::T; recent::Integer=5)::Dict{Symbol,Any} where {T<:agen
Dict(:name => "user", :text => usermsg) Dict(:name => "user", :text => usermsg)
] ]
#[WORKING] change qwen format put in model format # change qwen format put in model format
prompt = GeneralUtils.formatLLMtext(_prompt; formatname="llama3instruct") prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *= """
<|start_header_id|>assistant<|end_header_id|>
"""
response = a.func[:text2textInstructLLM](prompt) response = a.func[:text2textInstructLLM](prompt)
response = GeneralUtils.remove_french_accents(response) response = GeneralUtils.remove_french_accents(response)
response = replace(response, '*'=>"") response = replace(response, "**"=>"")
response = replace(response, "***"=>"")
response = replace(response, "<|eot_id|>"=>"") response = replace(response, "<|eot_id|>"=>"")
# check if response contain more than one functions from ["CHATBOX", "CHECKINVENTORY", "ENDCONVERSATION"] # check if response contain more than one functions from ["CHATBOX", "CHECKINVENTORY", "ENDCONVERSATION"]
@@ -395,284 +393,280 @@ function decisionMaker(a::T; recent::Integer=5)::Dict{Symbol,Any} where {T<:agen
end end
""" Assigns a scalar value to each new child node to be used for selec- # """ Assigns a scalar value to each new child node to be used for selec-
tion and backpropagation. This value effectively quantifies the agent's progress in task completion, # tion and backpropagation. This value effectively quantifies the agent's progress in task completion,
serving as a heuristic to steer the search algorithm towards the most promising regions of the tree. # serving as a heuristic to steer the search algorithm towards the most promising regions of the tree.
# Arguments # # Arguments
- `a::T1` # - `a::T1`
one of Yiem's agent # one of Yiem's agent
- `state::T2` # - `state::T2`
a game state # a game state
# Return # # Return
- `evaluation::Tuple{String, Integer}` # - `evaluation::Tuple{String, Integer}`
evaluation and score # evaluation and score
# Example # # Example
```jldoctest # ```jldoctest
julia> # julia>
``` # ```
# Signature # # Signature
""" # """
function evaluator(config::T1, state::T2 # function evaluator(config::T1, state::T2
)::Tuple{String,Integer} where {T1<:AbstractDict,T2<:AbstractDict} # )::Tuple{String,Integer} where {T1<:AbstractDict,T2<:AbstractDict}
systemmsg = # systemmsg =
""" # """
Analyze the trajectories of a solution to a question answering task. The trajectories are # Analyze the trajectories of a solution to a question answering task. The trajectories are
labeled by environmental observations about the situation, thoughts that can reason about # labeled by environmental observations about the situation, thoughts that can reason about
the current situation and actions that can be three types: # the current situation and actions that can be three types:
1) CHECKINVENTORY[query], which you can use to find wine in your inventory. # 1) CHECKINVENTORY[query], which you can use to find wine in your inventory.
2) CHATBOX[text], which you can use to interact with the user. # 2) CHATBOX[text], which you can use to interact with the user.
Given a question and a trajectory, evaluate its correctness and provide your reasoning and # Given a question and a trajectory, evaluate its correctness and provide your reasoning and
analysis in detail. Focus on the latest thought, action, and observation. Incomplete trajectories # analysis in detail. Focus on the latest thought, action, and observation. Incomplete trajectories
can be correct if the thoughts and actions so far are correct, even if the answer is not found # can be correct if the thoughts and actions so far are correct, even if the answer is not found
yet. Do not generate additional thoughts or actions. Then ending with the correctness score s # yet. Do not generate additional thoughts or actions. Then ending with the correctness score s
where s is an integer from 0 to 10. # where s is an integer from 0 to 10.
You should only respond in JSON format as describe below: # You should only respond in JSON format as describe below:
{"evaluation": "your evaluation", "score": "your evaluation score"} # {"evaluation": "your evaluation", "score": "your evaluation score"}
Here are some examples: # Here are some examples:
user: # user:
{ # {
"question": "I'm looking for a sedan with an automatic driving feature.", # "question": "I'm looking for a sedan with an automatic driving feature.",
"thought_1": "I have many types of sedans in my inventory, each with diverse features.", # "thought_1": "I have many types of sedans in my inventory, each with diverse features.",
"thought_2": "But there is only 1 model that has the feature customer wanted.", # "thought_2": "But there is only 1 model that has the feature customer wanted.",
"thought_3": "I should check our inventory first to see if we have it.", # "thought_3": "I should check our inventory first to see if we have it.",
"action_1": {"name": "inventory", "input": "Yiem model A"}, # "action_1": {"name": "inventory", "input": "Yiem model A"},
"observation_1": "Yiem model A is in stock." # "observation_1": "Yiem model A is in stock."
} # }
assistant # assistant
{ # {
"evaluation": "This trajectory is correct as it is reasonable to check an inventory for info provided in the question. # "evaluation": "This trajectory is correct as it is reasonable to check an inventory for info provided in the question.
It is also better to have simple searches corresponding to a single entity, making this the best action.", # It is also better to have simple searches corresponding to a single entity, making this the best action.",
"score": 10 # "score": 10
} # }
user: # user:
{ # {
"question": "Do you have an all-in-one pen with 4 colors and a pencil for sale?", # "question": "Do you have an all-in-one pen with 4 colors and a pencil for sale?",
"thought_1": "Let me check our inventory first to see if I have it.", # "thought_1": "Let me check our inventory first to see if I have it.",
"action_1": {"name": "inventory", "input": "pen with 4 color and a pencil."}, # "action_1": {"name": "inventory", "input": "pen with 4 color and a pencil."},
"observation_1": "I found {1: "Pilot Dr. grip 4-in-1 pen", 2: "Rotting pencil"}", # "observation_1": "I found {1: "Pilot Dr. grip 4-in-1 pen", 2: "Rotting pencil"}",
"thought_2": "Ok, I have what the user is asking. Let's tell the user.", # "thought_2": "Ok, I have what the user is asking. Let's tell the user.",
"action_2": {"name": "CHATBOX", "input": "Yes, we do have a Pilot Dr. grip 4-in-1 pen and a Rotting pencil"}, # "action_2": {"name": "CHATBOX", "input": "Yes, we do have a Pilot Dr. grip 4-in-1 pen and a Rotting pencil"},
"observation_1": "This is not what I wanted." # "observation_1": "This is not what I wanted."
} # }
assistant: # assistant:
{ # {
"evaluation": "This trajectory is incorrect as my search term should be related to a 4-colors pen with a pencil in it, # "evaluation": "This trajectory is incorrect as my search term should be related to a 4-colors pen with a pencil in it,
not a pen and a pencil seperately. A better search term should have been a 4-colors pen with a pencil, all-in-one.", # not a pen and a pencil seperately. A better search term should have been a 4-colors pen with a pencil, all-in-one.",
"score": 0 # "score": 0
} # }
Let's begin! # Let's begin!
""" # """
usermsg = """ # usermsg = """
$(JSON3.write(state[:thoughtHistory])) # $(JSON3.write(state[:thoughtHistory]))
""" # """
chathistory = # chathistory =
[ # [
Dict(:name => "system", :text => systemmsg), # Dict(:name => "system", :text => systemmsg),
Dict(:name => "user", :text => usermsg) # Dict(:name => "user", :text => usermsg)
] # ]
# put in model format # # put in model format
prompt = formatLLMtext(chathistory, "llama3instruct") # prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *= """
<|start_header_id|>assistant<|end_header_id|>
{
"""
pprint(prompt) # pprint(prompt)
externalService = config[:externalservice][:text2textinstruct] # externalService = config[:externalservice][:text2textinstruct]
# apply LLM specific instruct format # # apply LLM specific instruct format
externalService = config[:externalservice][:text2textinstruct] # externalService = config[:externalservice][:text2textinstruct]
msgMeta = GeneralUtils.generate_msgMeta( # msgMeta = GeneralUtils.generate_msgMeta(
externalService[:mqtttopic], # externalService[:mqtttopic],
senderName="evaluator", # senderName="evaluator",
senderId=string(uuid4()), # senderId=string(uuid4()),
receiverName="text2textinstruct", # receiverName="text2textinstruct",
mqttBroker=config[:mqttServerInfo][:broker], # mqttBroker=config[:mqttServerInfo][:broker],
mqttBrokerPort=config[:mqttServerInfo][:port], # mqttBrokerPort=config[:mqttServerInfo][:port],
) # )
outgoingMsg = Dict( # outgoingMsg = Dict(
:msgMeta => msgMeta, # :msgMeta => msgMeta,
:payload => Dict( # :payload => Dict(
:text => prompt, # :text => prompt,
:kwargs => Dict( # :kwargs => Dict(
:max_tokens => 512, # :max_tokens => 512,
:stop => ["<|eot_id|>"], # :stop => ["<|eot_id|>"],
) # )
) # )
) # )
for attempt in 1:5 # for attempt in 1:5
try # try
response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg) # response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg)
_responseJsonStr = response[:response][:text] # _responseJsonStr = response[:response][:text]
expectedJsonExample = """ # expectedJsonExample = """
Here is an expected JSON format: # Here is an expected JSON format:
{"evaluation": "...", "score": "..."} # {"evaluation": "...", "score": "..."}
""" # """
responseJsonStr = jsoncorrection(config, _responseJsonStr, expectedJsonExample) # responseJsonStr = jsoncorrection(config, _responseJsonStr, expectedJsonExample)
evaluationDict = copy(JSON3.read(responseJsonStr)) # evaluationDict = copy(JSON3.read(responseJsonStr))
# check if dict has all required value # # check if dict has all required value
dummya::AbstractString = evaluationDict[:evaluation] # dummya::AbstractString = evaluationDict[:evaluation]
dummyb::Integer = evaluationDict[:score] # dummyb::Integer = evaluationDict[:score]
return (evaluationDict[:evaluation], evaluationDict[:score]) # return (evaluationDict[:evaluation], evaluationDict[:score])
catch e # catch e
io = IOBuffer() # io = IOBuffer()
showerror(io, e) # showerror(io, e)
errorMsg = String(take!(io)) # errorMsg = String(take!(io))
st = sprint((io, v) -> show(io, "text/plain", v), stacktrace(catch_backtrace())) # st = sprint((io, v) -> show(io, "text/plain", v), stacktrace(catch_backtrace()))
println("\nAttempt $attempt. Error occurred: $errorMsg\n$st ", Dates.now(), " ", @__FILE__, " ", @__LINE__) # println("\nAttempt $attempt. Error occurred: $errorMsg\n$st ", Dates.now(), " ", @__FILE__, " ", @__LINE__)
end # end
end # end
error("evaluator failed to generate an evaluation") # error("evaluator failed to generate an evaluation")
end # end
""" # """
# Arguments # # Arguments
# Return # # Return
# Example # # Example
```jldoctest # ```jldoctest
julia> # julia>
``` # ```
# TODO # # TODO
- [] update docstring # - [] update docstring
- [x] implement the function # - [x] implement the function
- [x] add try block. check result that it is expected before returning # - [x] add try block. check result that it is expected before returning
# Signature # # Signature
""" # """
function reflector(config::T1, state::T2)::String where {T1<:AbstractDict,T2<:AbstractDict} # function reflector(config::T1, state::T2)::String where {T1<:AbstractDict,T2<:AbstractDict}
# https://github.com/andyz245/LanguageAgentTreeSearch/blob/main/hotpot/hotpot.py # # https://github.com/andyz245/LanguageAgentTreeSearch/blob/main/hotpot/hotpot.py
_prompt = # _prompt =
""" # """
You are a helpful sommelier working for a wine store. # You are a helpful sommelier working for a wine store.
Your goal is to recommend the best wine from your inventory that match the user preferences. # Your goal is to recommend the best wine from your inventory that match the user preferences.
You will be given a question and a trajectory of the previous help you've done for a user. # You will be given a question and a trajectory of the previous help you've done for a user.
You were unsuccessful in helping the user either because you guessed the wrong answer with Finish[answer], or you didn't know the user enough. # You were unsuccessful in helping the user either because you guessed the wrong answer with Finish[answer], or you didn't know the user enough.
In a few sentences, Diagnose a possible reason for failure and devise a new, concise, high level plan that aims to mitigate the same failure. # In a few sentences, Diagnose a possible reason for failure and devise a new, concise, high level plan that aims to mitigate the same failure.
Use complete sentences. # Use complete sentences.
You should only respond in JSON format as describe below: # You should only respond in JSON format as describe below:
{"reflection": "your relection"} # {"reflection": "your relection"}
Here are some examples: # Here are some examples:
Previous Trial: # Previous Trial:
{ # {
"question": "Hello, I would like a get a bottle of wine", # "question": "Hello, I would like a get a bottle of wine",
"thought_1": "A customer wants to buy a bottle of wine. Before making a recommendation, I need to know more about their preferences.", # "thought_1": "A customer wants to buy a bottle of wine. Before making a recommendation, I need to know more about their preferences.",
"action_1": {"name": "CHATBOX", "input": "What is the occasion for which you're buying this wine?"}, # "action_1": {"name": "CHATBOX", "input": "What is the occasion for which you're buying this wine?"},
"observation_1": "We are holding a wedding party", # "observation_1": "We are holding a wedding party",
"thought_2": "A wedding party, that's a great occasion! The customer might be looking for a celebratory drink. Let me ask some more questions to narrow down the options.", # "thought_2": "A wedding party, that's a great occasion! The customer might be looking for a celebratory drink. Let me ask some more questions to narrow down the options.",
"action_2": {"name": "CHATBOX", "input": "What type of food will you be serving at the wedding?"}, # "action_2": {"name": "CHATBOX", "input": "What type of food will you be serving at the wedding?"},
"observation_2": "It will be Thai dishes.", # "observation_2": "It will be Thai dishes.",
"thought_3": "With Thai food, I should recommend a wine that complements its spicy and savory flavors. And since it's a celebratory occasion, the customer might prefer a full-bodied wine.", # "thought_3": "With Thai food, I should recommend a wine that complements its spicy and savory flavors. And since it's a celebratory occasion, the customer might prefer a full-bodied wine.",
"action_3": {"name": "CHATBOX", "input": "What is your budget for this bottle of wine?"}, # "action_3": {"name": "CHATBOX", "input": "What is your budget for this bottle of wine?"},
"observation_3": "I would spend up to 50 bucks.", # "observation_3": "I would spend up to 50 bucks.",
"thought_4": "Now that I have some more information, it's time to narrow down the options.", # "thought_4": "Now that I have some more information, it's time to narrow down the options.",
"action_4": {"name": "winestock", "input": "red wine with full body, pairs well with spicy food, budget \$50"}, # "action_4": {"name": "winestock", "input": "red wine with full body, pairs well with spicy food, budget \$50"},
"observation_4": "I found the following wines in our stock: \n{\n 1: El Enemigo Cabernet Franc 2019\n2: Tantara Chardonnay 2017\n\n}\n", # "observation_4": "I found the following wines in our stock: \n{\n 1: El Enemigo Cabernet Franc 2019\n2: Tantara Chardonnay 2017\n\n}\n",
"thought_5": "Now that I have a list of potential wines, I need to know more about the customer's taste preferences.", # "thought_5": "Now that I have a list of potential wines, I need to know more about the customer's taste preferences.",
"action_5": {"name": "CHATBOX", "input": "What type of wine characteristics are you looking for? (e.g. tannin level, sweetness, intensity, acidity)"}, # "action_5": {"name": "CHATBOX", "input": "What type of wine characteristics are you looking for? (e.g. tannin level, sweetness, intensity, acidity)"},
"observation_5": "I like full-bodied red wine with low tannin.", # "observation_5": "I like full-bodied red wine with low tannin.",
"thought_6": "Now that I have more information about the customer's preferences, it's time to make a recommendation.", # "thought_6": "Now that I have more information about the customer's preferences, it's time to make a recommendation.",
"action_6": {"name": "recommendbox", "input": "El Enemigo Cabernet Franc 2019"}, # "action_6": {"name": "recommendbox", "input": "El Enemigo Cabernet Franc 2019"},
"observation_6": "I don't like the one you recommend. I want dry wine." # "observation_6": "I don't like the one you recommend. I want dry wine."
} # }
{ # {
"reflection": "I asked the user about the occasion, food type, and budget, and then searched for wine in the inventory right away. However, I should have asked the user for the specific wine type and their preferences in order to gather more information before making a recommendation." # "reflection": "I asked the user about the occasion, food type, and budget, and then searched for wine in the inventory right away. However, I should have asked the user for the specific wine type and their preferences in order to gather more information before making a recommendation."
} # }
Let's begin! # Let's begin!
Previous trial: # Previous trial:
$(JSON3.write(state[:thoughtHistory])) # $(JSON3.write(state[:thoughtHistory]))
{"reflection" # {"reflection"
""" # """
# apply LLM specific instruct format # # apply LLM specific instruct format
externalService = config[:externalservice][:text2textinstruct] # externalService = config[:externalservice][:text2textinstruct]
llminfo = externalService[:llminfo] # llminfo = externalService[:llminfo]
prompt = # prompt =
if llminfo[:name] == "llama3instruct" # if llminfo[:name] == "llama3instruct"
formatLLMtext_llama3instruct("system", _prompt) # formatLLMtext_llama3instruct("system", _prompt)
else # else
error("llm model name is not defied yet $(@__LINE__)") # error("llm model name is not defied yet $(@__LINE__)")
end # end
msgMeta = GeneralUtils.generate_msgMeta( # msgMeta = GeneralUtils.generate_msgMeta(
a.config[:externalservice][:text2textinstruct][:mqtttopic], # a.config[:externalservice][:text2textinstruct][:mqtttopic],
senderName="reflector", # senderName="reflector",
senderId=string(uuid4()), # senderId=string(uuid4()),
receiverName="text2textinstruct", # receiverName="text2textinstruct",
mqttBroker=config[:mqttServerInfo][:broker], # mqttBroker=config[:mqttServerInfo][:broker],
mqttBrokerPort=config[:mqttServerInfo][:port], # mqttBrokerPort=config[:mqttServerInfo][:port],
) # )
outgoingMsg = Dict( # outgoingMsg = Dict(
:msgMeta => msgMeta, # :msgMeta => msgMeta,
:payload => Dict( # :payload => Dict(
:text => prompt, # :text => prompt,
:kwargs => Dict( # :kwargs => Dict(
:max_tokens => 512, # :max_tokens => 512,
:stop => ["<|eot_id|>"], # :stop => ["<|eot_id|>"],
) # )
) # )
) # )
for attempt in 1:5 # for attempt in 1:5
try # try
response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg) # response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg)
_responseJsonStr = response[:response][:text] # _responseJsonStr = response[:response][:text]
expectedJsonExample = """ # expectedJsonExample = """
Here is an expected JSON format: # Here is an expected JSON format:
{"reflection": "..."} # {"reflection": "..."}
""" # """
responseJsonStr = jsoncorrection(config, _responseJsonStr, expectedJsonExample) # responseJsonStr = jsoncorrection(config, _responseJsonStr, expectedJsonExample)
reflectionDict = copy(JSON3.read(responseJsonStr)) # reflectionDict = copy(JSON3.read(responseJsonStr))
# check if dict has all required value # # check if dict has all required value
dummya::AbstractString = reflectionDict[:reflection] # dummya::AbstractString = reflectionDict[:reflection]
return reflectionDict[:reflection] # return reflectionDict[:reflection]
catch e # catch e
io = IOBuffer() # io = IOBuffer()
showerror(io, e) # showerror(io, e)
errorMsg = String(take!(io)) # errorMsg = String(take!(io))
st = sprint((io, v) -> show(io, "text/plain", v), stacktrace(catch_backtrace())) # st = sprint((io, v) -> show(io, "text/plain", v), stacktrace(catch_backtrace()))
println("\nAttempt $attempt. Error occurred: $errorMsg\n$st ", Dates.now(), " ", @__FILE__, " ", @__LINE__) # println("\nAttempt $attempt. Error occurred: $errorMsg\n$st ", Dates.now(), " ", @__FILE__, " ", @__LINE__)
end # end
end # end
error("reflector failed to generate a thought") # error("reflector failed to generate a thought")
end # end
""" Chat with llm. """ Chat with llm.
@@ -864,46 +858,6 @@ function think(a::T)::NamedTuple{(:actionname, :result),Tuple{String,String}} wh
) )
result = chatresponse result = chatresponse
# # store thoughtDict after the conversation finish
# if a.memory[:events][end][:thought][:action_name] == "ENDCONVERSATION"
# # generateSituationReport in the agent didn't include the last conversation
# # so the function will be called here
# a.memory[:recap] = generateSituationReport(a, a.func[:text2textInstructLLM]; skiprecent=0)
# for (i, event) in enumerate(a.memory[:events])
# if event[:subject] == "assistant"
# # create timeline of the last 3 conversation except the last one.
# # The former will be used as caching key and the latter will be the caching target
# # in vector database
# all_recapkeys = keys(a.memory[:recap]) # recap as caching
# all_recapkeys_vec = [r for r in all_recapkeys] # convert to a vector
# # select from 1 to 2nd-to-lase event (i.e. excluding the latest which is assistant's response)
# _recapkeys_vec = all_recapkeys_vec[1:i-1]
# # select only previous 3 recaps
# recapkeys_vec =
# if length(_recapkeys_vec) <= 3 # 1st message is a user's hello msg
# _recapkeys_vec # choose all
# else
# _recapkeys_vec[end-2:end]
# end
# #[PENDING] if there is specific data such as number, donot store in database
# tempmem = DataStructures.OrderedDict()
# for k in recapkeys_vec
# tempmem[k] = a.memory[:recap][k]
# end
# recap = GeneralUtils.dictToString_noKey(tempmem)
# thoughtDict = a.memory[:events][i][:thought] # latest assistant thoughtDict
# a.func[:insertSommelierDecision](recap, thoughtDict)
# else
# # skip
# end
# end
# println("Caching conversation done")
# end
elseif actionname == "CHECKINVENTORY" elseif actionname == "CHECKINVENTORY"
if rawresponse !== nothing if rawresponse !== nothing
vd = GeneralUtils.dfToVectorDict(rawresponse) vd = GeneralUtils.dfToVectorDict(rawresponse)
@@ -1040,10 +994,7 @@ function generatechat(a::sommelier, thoughtDict)
] ]
# put in model format # put in model format
prompt = GeneralUtils.formatLLMtext(_prompt; formatname="llama3instruct") prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *= """
<|start_header_id|>assistant<|end_header_id|>
"""
try try
response = a.func[:text2textInstructLLM](prompt) response = a.func[:text2textInstructLLM](prompt)
@@ -1170,10 +1121,7 @@ function generatechat(a::companion)
] ]
# put in model format # put in model format
prompt = GeneralUtils.formatLLMtext(_prompt; formatname="llama3instruct") prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *= """
<|start_header_id|>assistant<|end_header_id|>
"""
response = a.text2textInstructLLM(prompt) response = a.text2textInstructLLM(prompt)
@@ -1342,10 +1290,7 @@ function generatequestion(a, text2textInstructLLM::Function; recent=nothing)::St
] ]
# put in model format # put in model format
prompt = GeneralUtils.formatLLMtext(_prompt; formatname="llama3instruct") prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *= """
<|start_header_id|>assistant<|end_header_id|>
"""
try try
response = text2textInstructLLM(prompt) response = text2textInstructLLM(prompt)
@@ -1448,6 +1393,9 @@ function generateSituationReport(a, text2textInstructLLM::Function; skiprecent::
Let's begin! Let's begin!
""" """
header = ["Event_$i:" for i in eachindex(a.memory[:events])]
dictkey = lowercase.(["Event_$i" for i in eachindex(a.memory[:events])])
if length(a.memory[:events]) <= skiprecent if length(a.memory[:events]) <= skiprecent
return nothing return nothing
end end
@@ -1473,14 +1421,9 @@ function generateSituationReport(a, text2textInstructLLM::Function; skiprecent::
] ]
# put in model format # put in model format
prompt = GeneralUtils.formatLLMtext(_prompt; formatname="llama3instruct") prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *= """
<|start_header_id|>assistant<|end_header_id|>
"""
response = text2textInstructLLM(prompt) response = text2textInstructLLM(prompt)
header = ["Event_$i:" for i in eachindex(a.memory[:events])]
dictkey = lowercase.(["Event_$i" for i in eachindex(a.memory[:events])])
responsedict = GeneralUtils.textToDict(response, header; responsedict = GeneralUtils.textToDict(response, header;
dictKey=dictkey, symbolkey=true) dictKey=dictkey, symbolkey=true)
@@ -1531,10 +1474,7 @@ function detectWineryName(a, text)
] ]
# put in model format # put in model format
prompt = GeneralUtils.formatLLMtext(_prompt; formatname="llama3instruct") prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *= """
<|start_header_id|>assistant<|end_header_id|>
"""
try try
response = a.func[:text2textInstructLLM](prompt) response = a.func[:text2textInstructLLM](prompt)

141
src/llmfunction.jl
View File

@@ -326,7 +326,7 @@ julia>
# TODO # TODO
- [] update docstring - [] update docstring
- [WORKING] implement the function - implement the function
# Signature # Signature
""" """
@@ -382,7 +382,6 @@ function extractWineAttributes_1(a::T1, input::T2)::String where {T1<:agent, T2<
Let's begin! Let's begin!
""" """
attributes =
header = ["Comprehension:", "Wine_name:", "Winery:", "Vintage:", "Region:", "Country:", "Wine_type:", "Grape_varietal:", "Tasting_notes:", "Wine_price:", "Occasion:", "Food_to_be_paired_with_wine:"] header = ["Comprehension:", "Wine_name:", "Winery:", "Vintage:", "Region:", "Country:", "Wine_type:", "Grape_varietal:", "Tasting_notes:", "Wine_price:", "Occasion:", "Food_to_be_paired_with_wine:"]
dictkey = ["comprehension", "wine_name", "winery", "vintage", "region", "country", "wine_type", "grape_varietal", "tasting_notes", "wine_price", "occasion", "food_to_be_paired_with_wine"] dictkey = ["comprehension", "wine_name", "winery", "vintage", "region", "country", "wine_type", "grape_varietal", "tasting_notes", "wine_price", "occasion", "food_to_be_paired_with_wine"]
errornote = "" errornote = ""
@@ -407,7 +406,7 @@ function extractWineAttributes_1(a::T1, input::T2)::String where {T1<:agent, T2<
# check wheter all attributes are in the response # check wheter all attributes are in the response
checkFlag = false checkFlag = false
for word in attributes for word in header
if !occursin(word, response) if !occursin(word, response)
errornote = "$word attribute is missing in previous attempts" errornote = "$word attribute is missing in previous attempts"
println("Attempt $attempt $errornote ", Dates.now(), " ", @__FILE__, " ", @__LINE__) println("Attempt $attempt $errornote ", Dates.now(), " ", @__FILE__, " ", @__LINE__)
@@ -417,11 +416,19 @@ function extractWineAttributes_1(a::T1, input::T2)::String where {T1<:agent, T2<
end end
checkFlag == true ? continue : nothing checkFlag == true ? continue : nothing
responsedict = copy(JSON3.read(response)) # check whether response has all header
detected_kw = GeneralUtils.detect_keyword(header, response)
if sum(values(detected_kw)) < length(header)
errornote = "\nYiemAgent extractWineAttributes_1() response does not have all header"
continue
elseif sum(values(detected_kw)) > length(header)
errornote = "\nYiemAgent extractWineAttributes_1() response has duplicated header"
continue
end
responsedict = GeneralUtils.textToDict(response, header;
dictKey=dictkey, symbolkey=true)
# convert delete!(responsedict, :comprehension)
delete!(responsedict, :reasoning)
delete!(responsedict, :tasting_notes) delete!(responsedict, :tasting_notes)
delete!(responsedict, :occasion) delete!(responsedict, :occasion)
delete!(responsedict, :food_to_be_paired_with_wine) delete!(responsedict, :food_to_be_paired_with_wine)
@@ -431,9 +438,9 @@ function extractWineAttributes_1(a::T1, input::T2)::String where {T1<:agent, T2<
# check if winery, wine_name, region, country, wine_type, grape_varietal's value are in the query because sometime AI halucinates # check if winery, wine_name, region, country, wine_type, grape_varietal's value are in the query because sometime AI halucinates
checkFlag = false checkFlag = false
for i in attributes for i in dictkey
j = Symbol(i) j = Symbol(i)
if j ∉ [:reasoning, :tasting_notes, :occasion, :food_to_be_paired_with_wine] if j ∉ [:comprehension, :tasting_notes, :occasion, :food_to_be_paired_with_wine]
# in case j is wine_price it needs to be checked differently because its value is ranged # in case j is wine_price it needs to be checked differently because its value is ranged
if j == :wine_price if j == :wine_price
if responsedict[:wine_price] != "NA" if responsedict[:wine_price] != "NA"
@@ -516,7 +523,7 @@ function extractWineAttributes_2(a::T1, input::T2)::String where {T1<:agent, T2<
conversiontable = conversiontable =
""" """
Conversion Table: <Conversion Table>
Intensity level: Intensity level:
1 to 2: May correspond to "light-bodied" or a similar description. 1 to 2: May correspond to "light-bodied" or a similar description.
2 to 3: May correspond to "med light bodied", "medium light" or a similar description. 2 to 3: May correspond to "med light bodied", "medium light" or a similar description.
@@ -541,6 +548,7 @@ function extractWineAttributes_2(a::T1, input::T2)::String where {T1<:agent, T2<
3 to 4: May correspond to "medium acidity" or a similar description. 3 to 4: May correspond to "medium acidity" or a similar description.
4 to 5: May correspond to "semi high acidity" or a similar description. 4 to 5: May correspond to "semi high acidity" or a similar description.
4 to 5: May correspond to "high acidity" or a similar description. 4 to 5: May correspond to "high acidity" or a similar description.
</Conversion Table>
""" """
systemmsg = systemmsg =
@@ -554,67 +562,64 @@ function extractWineAttributes_2(a::T1, input::T2)::String where {T1<:agent, T2<
The preference form requires the following information: The preference form requires the following information:
sweetness, acidity, tannin, intensity sweetness, acidity, tannin, intensity
You must follow the following guidelines: <You must follow the following guidelines>
1) If specific information required in the preference form is not available in the query or there isn't any, mark with 'NA' to indicate this. 1) If specific information required in the preference form is not available in the query or there isn't any, mark with 'NA' to indicate this.
Additionally, words like 'any' or 'unlimited' mean no information is available. Additionally, words like 'any' or 'unlimited' mean no information is available.
2) Use the conversion table to convert the descriptive word level of sweetness, intensity, tannin, and acidity into a corresponding integer. 2) Use the conversion table to convert the descriptive word level of sweetness, intensity, tannin, and acidity into a corresponding integer.
3) Do not generate other comments. 3) Do not generate other comments.
</You must follow the following guidelines>
You should then respond to the user with the following points: <You should then respond to the user with>
- sweetness_keyword: The exact keywords in the user's query describing the sweetness level of the wine. Sweetness_keyword: The exact keywords in the user's query describing the sweetness level of the wine.
- sweetness: ( S ), where ( S ) represents integers indicating the range of sweetness levels. Example: 1-2 Sweetness: ( S ), where ( S ) represents integers indicating the range of sweetness levels. Example: 1-2
- acidity_keyword: The exact keywords in the user's query describing the acidity level of the wine. Acidity_keyword: The exact keywords in the user's query describing the acidity level of the wine.
- acidity: ( A ), where ( A ) represents integers indicating the range of acidity level. Example: 3-5 Acidity: ( A ), where ( A ) represents integers indicating the range of acidity level. Example: 3-5
- tannin_keyword: The exact keywords in the user's query describing the tannin level of the wine. Tannin_keyword: The exact keywords in the user's query describing the tannin level of the wine.
- tannin: ( T ), where ( T ) represents integers indicating the range of tannin level. Example: 1-3 Tannin: ( T ), where ( T ) represents integers indicating the range of tannin level. Example: 1-3
- intensity_keyword: The exact keywords in the user's query describing the intensity level of the wine. Intensity_keyword: The exact keywords in the user's query describing the intensity level of the wine.
- intensity: ( I ), where ( I ) represents integers indicating the range of intensity level. Example: 2-4 Intensity: ( I ), where ( I ) represents integers indicating the range of intensity level. Example: 2-4
</You should then respond to the user with>
You should only respond in the form (JSON) as described below: <You should only respond in format as described below>
{ Sweetness_keyword: ...
"sweetness_keyword": ..., Sweetness: ...
"sweetness": ..., Acidity_keyword: ...
"acidity_keyword": ..., Acidity: ...
"acidity": ..., Tannin_keyword: ...
"tannin_keyword": ..., Tannin: ...
"tannin": ..., Intensity_keyword: ...
"intensity_keyword": ..., Intensity: ...
"intensity": ... </You should only respond in format as described below>
}
Here are some examples: <Here are some examples>
User's query: I want a wine with a medium-bodied, low acidity, medium tannin. User's query: I want a wine with a medium-bodied, low acidity, medium tannin.
{ Sweetness_keyword: NA
"sweetness_keyword": "NA", Sweetness: NA
"sweetness": "NA", Acidity_keyword: low acidity
"acidity_keyword": "low acidity", Acidity: 1-2
"acidity": "1-2", Tannin_keyword: medium tannin
"tannin_keyword": "medium tannin", Tannin: 3-4
"tannin": "3-4", Intensity_keyword: medium-bodied
"intensity_keyword": "medium-bodied", Intensity: 3-4
"intensity": "3-4"
}
User's query: German red wine, under 100, pairs with spicy food
{
"sweetness_keyword": "NA",
"sweetness": "NA",
"acidity_keyword": "NA",
"acidity": "NA",
"tannin_keyword": "NA",
"tannin": "NA",
"intensity_keyword": "NA",
"intensity": "NA"
}
User's query: German red wine, under 100, pairs with spicy food
Sweetness_keyword: NA
Sweetness: NA
Acidity_keyword: NA
Acidity: NA
Tannin_keyword: NA
Tannin: NA
Intensity_keyword: NA
Intensity: NA
</Here are some examples>
Let's begin! Let's begin!
""" """
header = ["Sweetness_keyword:", "Sweetness:", "Acidity_keyword:", "Acidity:", "Tannin_keyword:", "Tannin:", "Intensity_keyword:", "Intensity:"]
dictkey = ["sweetness_keyword", "sweetness", "acidity_keyword", "acidity", "tannin_keyword", "tannin", "intensity_keyword", "intensity"]
errornote = "" errornote = ""
for attempt in 1:5 for attempt in 1:10
usermsg = usermsg =
""" """
$conversiontable $conversiontable
@@ -629,14 +634,22 @@ function extractWineAttributes_2(a::T1, input::T2)::String where {T1<:agent, T2<
] ]
# put in model format # put in model format
prompt = GeneralUtils.formatLLMtext(_prompt; formatname="llama3instruct") prompt = GeneralUtils.formatLLMtext(_prompt; formatname="qwen")
prompt *=
"""
<|start_header_id|>assistant<|end_header_id|>
"""
response = a.func[:text2textInstructLLM](prompt) response = a.func[:text2textInstructLLM](prompt)
responsedict = copy(JSON3.read(response))
# check whether response has all header
detected_kw = GeneralUtils.detect_keyword(header, response)
if sum(values(detected_kw)) < length(header)
errornote = "\nYiemAgent extractWineAttributes_2() response does not have all header"
continue
elseif sum(values(detected_kw)) > length(header)
errornote = "\nYiemAgent extractWineAttributes_2() response has duplicated header"
continue
end
responsedict = GeneralUtils.textToDict(response, header;
dictKey=dictkey, symbolkey=true)
# check whether each describing keyword is in the input to prevent halucination # check whether each describing keyword is in the input to prevent halucination
for i in ["sweetness", "acidity", "tannin", "intensity"] for i in ["sweetness", "acidity", "tannin", "intensity"]

9
test/etc.jl
View File

@@ -1,9 +0,0 @@
using GeneralUtils
response = "trajectory_evaluation:\nThe trajectory is correct so far. The thought accurately reflects the user's question, and the action taken is a valid attempt to retrieve data from the database that matches the specified criteria.\n\nanswer_evaluation:\nThe observation provides information about two red wines from Bordeaux rive droite in France, which partially answers the question. However, it does not provide a complete answer as it only lists the wine names and characteristics, but does not explicitly state whether there are any other wines that match the criteria.\n\naccepted_as_answer: No\n\nscore: 6\nThe trajectory is mostly correct, but the observation does not fully address the question.\n\nsuggestion: Consider adding more filters or parameters to the database query to retrieve a complete list of wines that match the specified criteria."
responsedict = GeneralUtils.textToDict(response,
["trajectory_evaluation", "answer_evaluation", "accepted_as_answer", "score", "suggestion"],
rightmarker=":", symbolkey=true)

272
test/runtests.jl
View File

@@ -1,272 +0,0 @@
using Revise
using JSON, JSON3, Dates, UUIDs, PrettyPrinting, LibPQ, Base64, DataFrames
using YiemAgent, GeneralUtils
using Base.Threads
# ---------------------------------------------- 100 --------------------------------------------- #
# load config
config = JSON3.read("/appfolder/app/dev/YiemAgent/test/config.json")
# config = copy(JSON3.read("../mountvolume/config.json"))
function executeSQL(sql::T) where {T<:AbstractString}
DBconnection = LibPQ.Connection("host=192.168.88.12 port=10201 dbname=wineDB user=yiemtechnologies password=yiemtechnologies@Postgres_0.0")
result = LibPQ.execute(DBconnection, sql)
close(DBconnection)
return result
end
function executeSQLVectorDB(sql)
DBconnection = LibPQ.Connection("host=192.168.88.12 port=10203 dbname=SQLVectorDB user=yiemtechnologies password=yiemtechnologies@Postgres_0.0")
result = LibPQ.execute(DBconnection, sql)
close(DBconnection)
return result
end
function text2textInstructLLM(prompt::String)
msgMeta = GeneralUtils.generate_msgMeta(
config[:externalservice][:text2textinstruct][:mqtttopic];
msgPurpose="inference",
senderName="yiemagent",
senderId=string(uuid4()),
receiverName="text2textinstruct_small",
mqttBrokerAddress=config[:mqttServerInfo][:broker],
mqttBrokerPort=config[:mqttServerInfo][:port],
)
outgoingMsg = Dict(
:msgMeta => msgMeta,
:payload => Dict(
:text => prompt,
:kwargs => Dict(
:num_ctx => 16384,
:temperature => 0.2,
)
)
)
_response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg; timeout=6000)
response = _response[:response][:text]
return response
end
# get text embedding from a LLM service
function getEmbedding(text::T) where {T<:AbstractString}
msgMeta = GeneralUtils.generate_msgMeta(
config[:externalservice][:text2textinstruct][:mqtttopic];
msgPurpose="embedding",
senderName="yiemagent",
senderId=string(uuid4()),
receiverName="text2textinstruct_small",
mqttBrokerAddress=config[:mqttServerInfo][:broker],
mqttBrokerPort=config[:mqttServerInfo][:port],
)
outgoingMsg = Dict(
:msgMeta => msgMeta,
:payload => Dict(
:text => [text] # must be a vector of string
)
)
response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg; timeout=6000)
embedding = response[:response][:embeddings]
return embedding
end
function findSimilarTextFromVectorDB(text::T1, tablename::T2, embeddingColumnName::T3,
vectorDB::Function; limit::Integer=1
)::DataFrame where {T1<:AbstractString, T2<:AbstractString, T3<:AbstractString}
# get embedding from LLM service
embedding = getEmbedding(text)[1]
# check whether there is close enough vector already store in vectorDB. if no, add, else skip
sql = """
SELECT *, $embeddingColumnName <-> '$embedding' as distance
FROM $tablename
ORDER BY distance LIMIT $limit;
"""
response = vectorDB(sql)
df = DataFrame(response)
return df
end
function similarSQLVectorDB(query; maxdistance::Integer=100)
tablename = "sqlllm_decision_repository"
# get embedding of the query
df = findSimilarTextFromVectorDB(query, tablename,
"function_input_embedding", executeSQLVectorDB)
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
if row != 0 && distance < maxdistance
# if there is usable SQL, return it.
output_b64 = df[1, :function_output_base64] # pick the closest match
output_str = String(base64decode(output_b64))
rowid = df[1, :id]
println("\n~~~ found similar sql. row id $rowid, distance $distance ", @__FILE__, " ", @__LINE__)
return (dict=output_str, distance=distance)
else
println("\n~~~ similar sql not found, max distance $maxdistance ", @__FILE__, " ", @__LINE__)
return (dict=nothing, distance=nothing)
end
end
function insertSQLVectorDB(query::T1, SQL::T2; maxdistance::Integer=1) where {T1<:AbstractString, T2<:AbstractString}
tablename = "sqlllm_decision_repository"
# get embedding of the query
# query = state[:thoughtHistory][:question]
df = findSimilarTextFromVectorDB(query, tablename,
"function_input_embedding", executeSQLVectorDB)
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
if row == 0 || distance > maxdistance # no close enough SQL stored in the database
query_embedding = getEmbedding(query)[1]
query = replace(query, "'" => "")
sql_base64 = base64encode(SQL)
sql_ = replace(SQL, "'" => "")
sql = """
INSERT INTO $tablename (function_input, function_output, function_output_base64, function_input_embedding) VALUES ('$query', '$sql_', '$sql_base64', '$query_embedding');
"""
println("\n~~~ added new decision to vectorDB ", @__FILE__, " ", @__LINE__)
println(sql)
_ = executeSQLVectorDB(sql)
end
end
function similarSommelierDecision(recentevents::T1; maxdistance::Integer=5
)::Union{AbstractDict, Nothing} where {T1<:AbstractString}
tablename = "sommelier_decision_repository"
# find similar
println("\n~~~ search vectorDB for this: $recentevents ", @__FILE__, " ", @__LINE__)
df = findSimilarTextFromVectorDB(recentevents, tablename,
"function_input_embedding", executeSQLVectorDB)
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
if row != 0 && distance < maxdistance
# if there is usable decision, return it.
rowid = df[1, :id]
println("\n~~~ found similar decision. row id $rowid, distance $distance ", @__FILE__, " ", @__LINE__)
output_b64 = df[1, :function_output_base64] # pick the closest match
_output_str = String(base64decode(output_b64))
output = copy(JSON3.read(_output_str))
return output
else
println("\n~~~ similar decision not found, max distance $maxdistance ", @__FILE__, " ", @__LINE__)
return nothing
end
end
function insertSommelierDecision(recentevents::T1, decision::T2; maxdistance::Integer=5
) where {T1<:AbstractString, T2<:AbstractDict}
tablename = "sommelier_decision_repository"
# find similar
df = findSimilarTextFromVectorDB(recentevents, tablename,
"function_input_embedding", executeSQLVectorDB)
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
if row == 0 || distance > maxdistance # no close enough SQL stored in the database
recentevents_embedding = a.func[:getEmbedding](recentevents)[1]
recentevents = replace(recentevents, "'" => "")
decision_json = JSON3.write(decision)
decision_base64 = base64encode(decision_json)
decision = replace(decision_json, "'" => "")
sql = """
INSERT INTO $tablename (function_input, function_output, function_output_base64, function_input_embedding) VALUES ('$recentevents', '$decision', '$decision_base64', '$recentevents_embedding');
"""
println("\n~~~ added new decision to vectorDB ", @__FILE__, " ", @__LINE__)
println(sql)
_ = executeSQLVectorDB(sql)
else
println("~~~ similar decision previously cached, distance $distance ", @__FILE__, " ", @__LINE__)
end
end
sessionId = "12345"
externalFunction = (
getEmbedding=getEmbedding,
text2textInstructLLM=text2textInstructLLM,
executeSQL=executeSQL,
similarSQLVectorDB=similarSQLVectorDB,
insertSQLVectorDB=insertSQLVectorDB,
similarSommelierDecision=similarSommelierDecision,
insertSommelierDecision=insertSommelierDecision,
)
a = YiemAgent.sommelier(
externalFunction;
name="Ton",
id=sessionId, # agent instance id
retailername="Yiem",
)
while true
println("your respond: ")
user_answer = readline()
response = YiemAgent.conversation(a, Dict(:text=> user_answer))
println("\n$response")
end
# response = YiemAgent.conversation(a, Dict(:text=> "I want to get a French red wine under 100."))

292
test/test_1.jl Normal file
View File

@@ -0,0 +1,292 @@
using Revise
using JSON, JSON3, Dates, UUIDs, PrettyPrinting, LibPQ, Base64, DataFrames
using YiemAgent, GeneralUtils
using Base.Threads
# ---------------------------------------------- 100 --------------------------------------------- #
# load config
config = JSON3.read("/appfolder/app/dev/YiemAgent/test/config.json")
# config = copy(JSON3.read("../mountvolume/config.json"))
function executeSQL(sql::T) where {T<:AbstractString}
host = config[:externalservice][:wineDB][:host]
port = config[:externalservice][:wineDB][:port]
dbname = config[:externalservice][:wineDB][:dbname]
user = config[:externalservice][:wineDB][:user]
password = config[:externalservice][:wineDB][:password]
DBconnection = LibPQ.Connection("host=$host port=$port dbname=$dbname user=$user password=$password")
result = LibPQ.execute(DBconnection, sql)
close(DBconnection)
return result
end
function executeSQLVectorDB(sql)
host = config[:externalservice][:SQLVectorDB][:host]
port = config[:externalservice][:SQLVectorDB][:port]
dbname = config[:externalservice][:SQLVectorDB][:dbname]
user = config[:externalservice][:SQLVectorDB][:user]
password = config[:externalservice][:SQLVectorDB][:password]
DBconnection = LibPQ.Connection("host=$host port=$port dbname=$dbname user=$user password=$password")
result = LibPQ.execute(DBconnection, sql)
close(DBconnection)
return result
end
function text2textInstructLLM(prompt::String; maxattempt=3)
msgMeta = GeneralUtils.generate_msgMeta(
config[:externalservice][:loadbalancer][:mqtttopic];
msgPurpose="inference",
senderName="yiemagent",
senderId=sessionId,
receiverName="text2textinstruct_small",
mqttBrokerAddress=config[:mqttServerInfo][:broker],
mqttBrokerPort=config[:mqttServerInfo][:port],
)
outgoingMsg = Dict(
:msgMeta => msgMeta,
:payload => Dict(
:text => prompt,
:kwargs => Dict(
:num_ctx => 16384,
:temperature => 0.2,
)
)
)
response = nothing
for attempts in 1:maxattempt
_response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg; timeout=300, maxattempt=maxattempt)
payload = _response[:response]
if _response[:success] && payload[:text] !== nothing
response = _response[:response][:text]
break
else
println("\n<text2textInstructLLM()> attempt $attempts/$maxattempt failed ", @__FILE__, ":", @__LINE__, " $(Dates.now())")
pprintln(outgoingMsg)
println("</text2textInstructLLM()> attempt $attempts/$maxattempt failed ", @__FILE__, ":", @__LINE__, " $(Dates.now())\n")
sleep(3)
end
end
return response
end
# get text embedding from a LLM service
function getEmbedding(text::T) where {T<:AbstractString}
msgMeta = GeneralUtils.generate_msgMeta(
config[:externalservice][:loadbalancer][:mqtttopic];
msgPurpose="embedding",
senderName="yiemagent",
senderId=sessionId,
receiverName="text2textinstruct_small",
mqttBrokerAddress=config[:mqttServerInfo][:broker],
mqttBrokerPort=config[:mqttServerInfo][:port],
)
outgoingMsg = Dict(
:msgMeta => msgMeta,
:payload => Dict(
:text => [text] # must be a vector of string
)
)
response = GeneralUtils.sendReceiveMqttMsg(outgoingMsg; timeout=120)
embedding = response[:response][:embeddings]
return embedding
end
function findSimilarTextFromVectorDB(text::T1, tablename::T2, embeddingColumnName::T3,
vectorDB::Function; limit::Integer=1
)::DataFrame where {T1<:AbstractString, T2<:AbstractString, T3<:AbstractString}
# get embedding from LLM service
embedding = getEmbedding(text)[1]
# check whether there is close enough vector already store in vectorDB. if no, add, else skip
sql = """
SELECT *, $embeddingColumnName <-> '$embedding' as distance
FROM $tablename
ORDER BY distance LIMIT $limit;
"""
response = vectorDB(sql)
df = DataFrame(response)
return df
end
function similarSQLVectorDB(query; maxdistance::Integer=100)
tablename = "sqlllm_decision_repository"
# get embedding of the query
df = findSimilarTextFromVectorDB(query, tablename,
"function_input_embedding", executeSQLVectorDB)
# println(df[1, [:id, :function_output]])
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
# distance = 100 # CHANGE this is for testing only
if row != 0 && distance < maxdistance
# if there is usable SQL, return it.
output_b64 = df[1, :function_output_base64] # pick the closest match
output_str = String(base64decode(output_b64))
rowid = df[1, :id]
println("\n~~~ found similar sql. row id $rowid, distance $distance ", @__FILE__, ":", @__LINE__, " $(Dates.now())")
return (dict=output_str, distance=distance)
else
println("\n~~~ similar sql not found, max distance $maxdistance ", @__FILE__, ":", @__LINE__, " $(Dates.now())")
return (dict=nothing, distance=nothing)
end
end
function insertSQLVectorDB(query::T1, SQL::T2; maxdistance::Integer=3) where {T1<:AbstractString, T2<:AbstractString}
tablename = "sqlllm_decision_repository"
# get embedding of the query
# query = state[:thoughtHistory][:question]
df = findSimilarTextFromVectorDB(query, tablename,
"function_input_embedding", executeSQLVectorDB)
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
if row == 0 || distance > maxdistance # no close enough SQL stored in the database
query_embedding = getEmbedding(query)[1]
query = replace(query, "'" => "")
sql_base64 = base64encode(SQL)
sql_ = replace(SQL, "'" => "")
sql = """
INSERT INTO $tablename (function_input, function_output, function_output_base64, function_input_embedding) VALUES ('$query', '$sql_', '$sql_base64', '$query_embedding');
"""
# println("\n~~~ added new decision to vectorDB ", @__FILE__, ":", @__LINE__, " $(Dates.now())")
# println(sql)
_ = executeSQLVectorDB(sql)
end
end
function similarSommelierDecision(recentevents::T1; maxdistance::Integer=5
)::Union{AbstractDict, Nothing} where {T1<:AbstractString}
tablename = "sommelier_decision_repository"
# find similar
println("\n~~~ search vectorDB for this: $recentevents ", @__FILE__, " ", @__LINE__)
df = findSimilarTextFromVectorDB(recentevents, tablename,
"function_input_embedding", executeSQLVectorDB)
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
if row != 0 && distance < maxdistance
# if there is usable decision, return it.
rowid = df[1, :id]
println("\n~~~ found similar decision. row id $rowid, distance $distance ", @__FILE__, " ", @__LINE__)
output_b64 = df[1, :function_output_base64] # pick the closest match
_output_str = String(base64decode(output_b64))
output = copy(JSON3.read(_output_str))
return output
else
println("\n~~~ similar decision not found, max distance $maxdistance ", @__FILE__, " ", @__LINE__)
return nothing
end
end
function insertSommelierDecision(recentevents::T1, decision::T2; maxdistance::Integer=5
) where {T1<:AbstractString, T2<:AbstractDict}
tablename = "sommelier_decision_repository"
# find similar
df = findSimilarTextFromVectorDB(recentevents, tablename,
"function_input_embedding", executeSQLVectorDB)
row, col = size(df)
distance = row == 0 ? Inf : df[1, :distance]
if row == 0 || distance > maxdistance # no close enough SQL stored in the database
recentevents_embedding = a.func[:getEmbedding](recentevents)[1]
recentevents = replace(recentevents, "'" => "")
decision_json = JSON3.write(decision)
decision_base64 = base64encode(decision_json)
decision = replace(decision_json, "'" => "")
sql = """
INSERT INTO $tablename (function_input, function_output, function_output_base64, function_input_embedding) VALUES ('$recentevents', '$decision', '$decision_base64', '$recentevents_embedding');
"""
println("\n~~~ added new decision to vectorDB ", @__FILE__, " ", @__LINE__)
println(sql)
_ = executeSQLVectorDB(sql)
else
println("~~~ similar decision previously cached, distance $distance ", @__FILE__, " ", @__LINE__)
end
end
sessionId = "12345"
externalFunction = (
getEmbedding=getEmbedding,
text2textInstructLLM=text2textInstructLLM,
executeSQL=executeSQL,
similarSQLVectorDB=similarSQLVectorDB,
insertSQLVectorDB=insertSQLVectorDB,
similarSommelierDecision=similarSommelierDecision,
insertSommelierDecision=insertSommelierDecision,
)
a = YiemAgent.sommelier(
externalFunction;
name="Ton",
id=sessionId, # agent instance id
retailername="Yiem",
)
while true
println("your respond: ")
user_answer = readline()
response = YiemAgent.conversation(a, Dict(:text=> user_answer))
println("\n$response")
end
# response = YiemAgent.conversation(a, Dict(:text=> "I want to get a French red wine under 100."))