LangAGI-Lab / camel

Counseling with CAMEL

Setup

import argparse
import json
import multiprocessing
import re
import traceback
from abc import ABC, abstractmethod
from pathlib import Path

import requests
from langchain.prompts import PromptTemplate
from langchain_openai import OpenAI

Define Agents

class Agent():
    def __init__(self, vLLM_server, model_id):
        self.llm = OpenAI(
            temperature=0.0,
            openai_api_key='EMPTY',
            openai_api_base=vLLM_server,
            max_tokens=512,
            model=model_id
        )

    def generate(self):
        pass

class CBTAgent(Agent):
    def __init__(self, prompt, vLLM_server, model_id):
        super().__init__(vLLM_server, model_id)
        self.prompt_template = PromptTemplate(
            input_variables=[
                "client_information",
                "reason_counseling",
                'history',
            ],
            template=prompt
        )

    def generate(self, client_information, reason, history):
        history_text = '\n'.join(
            [
                f"{message['role'].capitalize()}: {message['message']}"
                for message in history
            ]
        )
        prompt = self.prompt_template.format(
            client_information=client_information,
            reason_counseling=reason,
            history= history_text
        )
        response = self.llm.invoke(prompt)

        try:
            cbt_technique = response.split("Counseling")[0].replace("\n", "")
        except:
            cbt_technique = None
        try:
            cbt_plan = response.split("Counseling planning:\n")[1].split("\nCBT")[0]
        except:
            cbt_plan = None

        return cbt_technique, cbt_plan

class CounsleorAgent(Agent):
    def __init__(self,  prompt, vLLM_server, model_id, cbt_plan):
       super().__init__(vLLM_server, model_id)
       self.cbt_plan = cbt_plan
       self.prompt_template = PromptTemplate(
            input_variables=[
                "client_information",
                "reason_counseling",
                "cbt_plan",
                "history"
            ],
            template=prompt
        )

    def generate(self, client_information, reason, history):
        history_text = '\n'.join(
            [
                f"{message['role'].capitalize()}: {message['message']}"
                for message in history
            ]
        )
        prompt = self.prompt_template.format(
            client_information=client_information,
            reason_counseling=reason,
            cbt_plan=self.cbt_plan,
            history=history_text,
        )
        # print(prompt)
        response = self.llm.invoke(prompt)
        # print(f"Response: {response}")

        if "'message':" in response:
            response = response.split("'message':")[1].split(", {")[0].replace("\"","").replace("]", "").replace("}", "")
        return response.split("Counselor:")[-1].replace("\n", "").replace("\\", "").replace("\"","").strip()

Define prompt templates

RESPONSE_PROMPT="""<|start_header_id|>system<|end_header_id|>

You are playing the role of a counselor in a psychological counseling session. Your task is to use the provided client information and counseling planning to generate the next counselor utterance in the dialogue. The goal is to create a natural and engaging response that builds on the previous conversation and aligns with the counseling plan.<|eot_id|><|start_header_id|>user<|end_header_id|>

Client Information:
{client_information}

Reason for seeking counseling:
{reason_counseling}

Counseling planning:
{cbt_plan}

Counseling Dialogue:
{history}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""

CBT_PLAN_PROMPT="""<|start_header_id|>system<|end_header_id|>

You are a counselor specializing in CBT techniques. Your task is to use the provided client information, and dialogue to generate an appropriate CBT technique and a detailed counseling plan.<|eot_id|><|start_header_id|>user<|end_header_id|>

Types of CBT Techniques:
Efficiency Evaluation, Pie Chart Technique, Alternative Perspective, Decatastrophizing, Pros and Cons Analysis, Evidence-Based Questioning, Reality Testing, Continuum Technique, Changing Rules to Wishes, Behavior Experiment, Problem-Solving Skills Training, Systematic Exposure

Client Information:
{client_information}

Reason for seeking counseling:
{reason_counseling}

Counseling Dialogue:
{history}

Choose an appropriate CBT technique and create a counseling plan based on that technique.<|eot_id|><|start_header_id|>assistant<|end_header_id|>"""

Start!

def collect_info(name, age, gender, occupation, education, matrital_status, family_details, reason):
    CLINET_INFO = f"""Name: {name}
Age: {age}
Gender: {gender}
Occupation: {occupation}
Education: {education}
Marital Status: {matrital_status}
Family Details: {family_details}"""

    REASON_FOR_COUNSELING = reason
    HISTORY_INIT = f"Counselor: Hi {name}, it's nice to meet you. How can I assist you today?\nClient: "

    return CLINET_INFO, REASON_FOR_COUNSELING, HISTORY_INIT

def start_demo(intake_form, reason, history_init):
    model_id = "DLI-Lab/camel"
    vLLM_server = ```YOUR vLLM SERVER```
    max_turns = 20

    print("Welcome to the Multi-Turn ClientAgent Demo!\n")
    print(f"[Intake Form]")
    print(intake_form)
    print("Type 'exit' to quit the demo.\n")

    print("====== Counseling Session ======\n")
    first_response = history_init.split('Counselor: ')[-1].split('\nClient')[0]
    print(f"Counselor: {first_response}")

    num_turn = 0
    while num_turn < max_turns:
        if num_turn == 0:
            user_input = input("You (Client): ")
            # print(f"You (Client): {user_input}")
            history_init = history_init + user_input
            history = [
                {"role": "Counselor", "message": history_init.split("Counselor: ")[-1].split("\nClient")[0]},
                {"role": "Client", "message": history_init.split("Client: ")[-1]}
            ]
            # print("CBT Planning")
            CBT_Planner = CBTAgent(CBT_PLAN_PROMPT, vLLM_server, model_id)
            cbt_technique, cbt_plan = CBT_Planner.generate(intake_form, reason, history)
            # print(f"CBT Technique: {cbt_technique}")
            # print(f"CBT Plan: {cbt_plan}")

            num_turn+=1
        else:
            counselor = CounsleorAgent(RESPONSE_PROMPT, vLLM_server, model_id, cbt_plan)
            counselor_response = counselor.generate(intake_form, reason, history)
            print(f"Counselor: {counselor_response}")

            history.append({"role": "Counselor", "message": counselor_response})

            user_input = input("You (Client): ")

            if user_input.lower() == 'exit':
                print("\n====== Exiting the demo. Goodbye! ======\n")
                break

            print(f"You (Client): {user_input}")
            history.append({"role": "Client", "message": user_input})

            num_turn+=1

    print("Demo completed.")
    return cbt_plan, history


## Example
# name = "Laura"
# age = "45"
# gender = "female"
# occupation =  "Office Job"
# education = "College Graduate"
# matrital_status = "Single"
# family_details = "Lives alone"

name = input("Let's begin the pre-counseling session. What is your name? ")
age = input("How old are you? ")
gender = input("What is your gender? (e.g., Male, Female)")
occupation = input("What is your occupation? ")
education = input("What is your highest level of education? (e.g., College Graduate)")
marital_status = input("What is your marital status? (e.g., Single, Married)")
family_details = input("Can you briefly describe your family situation? (e.g., Lives alone)")
reason = input("What brings you here for counseling? Please explain briefly. ")


CLINET_INFO, REASON_FOR_COUNSELING, HISTORY_INIT = collect_info(name, age, gender, occupation, education, matrital_status, family_details, reason)
cbt_plan, history = start_demo(CLINET_INFO, REASON_FOR_COUNSELING, HISTORY_INIT)

print(f"CBT Plan: {cbt_plan}\n\n")

for message in history:
    print(f"{message['role']}: {message['message']}")