Code examples for meeting summaries and transcription. A part of article on my blog -> kolegaliterat.pl

crew_meeting_summary.py

import os
os.environ["OPENAI_API_KEY"] = OPEN_AI_API_KEY

from crewai import Agent, Task, Crew, Process
from crewai_tools import FileReadTool, TXTSearchTool, DirectoryReadTool
from langchain_openai import OpenAI, ChatOpenAI

#Tools
transcription_directory = DirectoryReadTool(FOLDER_PATH)
summaries_directory = DirectoryReadTool(FOLDER_PATH)
file_read_tool = FileReadTool()
txt_search_tool = TXTSearchTool()

#Agents personas
project_manager = Agent(
    role = 'Gamedev Project Manager',
    goal = 'To establish project priorities and identify potential risks within game development, devise strategies to mitigate those risks, and develop actionable work plans.',
    backstory = """The agent functions as a project manager at an independent gaming company, aiding teams in concentrating on critical tasks. Known for a communication style that is both direct and centered on problem-solving.""",
    LLM = OpenAI(model_name = "gpt-4-turbo"),
    tools = [transcription_directory, file_read_tool, txt_search_tool],
    verbose = True,
    memory = True,
    cache = True,
    allow_delegation = True,
    max_iter = 10,
)

scrum_master = Agent(
    role = 'Scrum Master in Gamedev Team',
    goal = 'To outline objectives for the upcoming sprint.',
    backstory = """The agent holds the position of scrum master within an independent gaming company and is currently engaged in determining the tasks for the next sprint, utilizing insights from the project manager's strategy and planning.""",
    LLM = OpenAI(model_name = "gpt-4-turbo"),
    tools = [summaries_directory, file_read_tool, txt_search_tool],
    verbose = True,
    memory = True,
    cache = True,
    max_iter = 10,
)

#Tasks
meeting_summary = Task(
    description = ("Concentrate on capturing the essential details discussed during the meeting"),
    expected_output = "Produce a concise bullet-point list that encapsulates the core information for each topic addressed.",
    agent = project_manager,
    output_file = OUTPUT_FOLDER,

)

decision_summary = Task(
    description = ("Summarize the decisions made in a meeting."
                   "Assign a numerical value between 0 and 5 next to each decision to indicate its level of impact, with higher numbers signifying greater impact. "
                   "Include any mentioned solutions related to these decisions."
    ),
    expected_output = "A bullet list of decision. With impact score in ()",
    agent = project_manager,
    output_file = OUTPUT_FOLDER,
    
)

actions = Task(
    description = ("Formatting Instructions: Utilize Markdown syntax to format the list. Begin each action item with - [ ] to denote an unchecked task."
                   "Accompany each action with a priority label, ranging from P0 to P5, where P0 indicates the highest level of urgency."
                   "Phrase each item as a task that must be completed."
    ),
    expected_output = "Generate a list of actionable items for the forthcoming sprint.",
    agent = scrum_master,
    output_file = OUTPUT_FOLDER
)

#Crew gathering
crew = Crew(
    agents = [project_manager, scrum_master],
    tasks = [meeting_summary, decision_summary, actions],
    memory=True,
    cache=True,
    process=Process.hierarchical,
    max_rpm = 100,
    manager_llm=ChatOpenAI(model="gpt-4-turbo")
)

result = crew.kickoff()
print(result)

transcriber.py

from openai import OpenAI
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from progress.bar import ChargingBar

##App files
from src.helper import DataFiles

class Transcriber():
    def __init__(self, api_key: str, file_paths: str):
        self.open_ai_client = OpenAI(api_key=api_key)
        self.file_paths = file_paths
        self.transcription_folder = DataFiles("transcription")

    def transcribe_meeting_call(self):
        progress = ChargingBar("Cleaning transcriptions!", max=len(self.file_paths))

        for part, file in enumerate(self.file_paths, start=0):
            transcription = self.__create_transcription(file)

            self.transcription_folder.save_data(f"part_{part}.txt", transcription)

            progress.next()
        
        progress.finish()

    def __create_transcription(self, file_path: str):
        with open(file_path, "rb") as audio_file:
        
            transcription = self.open_ai_client.audio.transcriptions.create(
                model="whisper-1",
                file=audio_file,
                response_format="text"
        )
        
        return transcription

class PurityClerk():
    def __init__(self, api_key, file_paths):
        self.llm_cleaner = ChatOpenAI(api_key=api_key, model="gpt-4-turbo")
        self.output_parser = StrOutputParser()
        self.summaries_folder = DataFiles("crew/meetings")
        self.file_paths = file_paths
       
    def prepare_summary(self):
        progress = ChargingBar("Cleaning transcriptions!", max=len(self.file_paths))

        for path in self.file_paths:
            with open(path, 'r') as file:
                data = file.read()
            
            summary_filename = path[-10:]
            
            summary = self.__clean_up_transcription(data)

            self.summaries_folder.save_data(summary_filename, summary)

            progress.next()
        
        progress.finish()

    def __clean_up_transcription(self, content):
        ###Langchain can be easilly removed and replaced by a request to the OpenAI API. 
        
        prompt = ChatPromptTemplate.from_messages([
            ("system", "##ROLE \
                        You're well known person who specilizes in creating great summaries based on meeting transcprition. \
                        You're always focused on looking for crucial information from meeting. \
                        ###TASK \
                        Clean up transcription of a meeting. Focus on important information, remove repetitions and small-talk. \
                        Summary needs to have max 300 words and be in english"),
            ("user", "{transcription}")
        ])

        chain = prompt | self.llm_cleaner | self.output_parser

        summary = chain.invoke({"transcription": content})

        return summary