[대화형 AI] TaskWeaver에 대해 알아보자.

개요

MS 대화형 AI는 서비스 제품인 1. MS Copilot과 오픈소스로 공개된 2. TaskWeaver, 3. AutoGen 으로 나뉩니다. 그 중 TaskWeaver에 대해 알아봅시다.

 

2. TaskWeaver

오픈소스 소개

• Task Weaver는 데이터 분석 작업을 목적으로 개발된 AI입니다.

• 사용자의 프롬포트에서 요청을 해석하고, 실행 가능한 코드 스니펫을 생성해 사용자에게 결과와 함께 제공합니다.

• 데이터 분석 결과와 함께 관련 코드까지 제공한다는 장점이 있습니다.

• UI 인터페이스도 제공할 수 있습니다.

 

오픈소스 특징

• Task Weaver는 2개의 대화형 AI에 각각 역할을 설정해 AI간 대화(interaction)를 통해 프롬포트 결과를 출력합니다.
  • Planner: 사용자가 입력한 프롬포트를 읽고 분석 단계를 설정합니다. Code Interpreter에게 단계별 분석 내용을 전달합니다. 모든 단계가 정상적으로 종료되면 사용자에게 결과를 반환합니다.
  • Code Interpreter: Planner에게 전달받은 단계를 실행할 코드를 작성합니다.

 

• 다른 대화형 AI 서비스들과 달리 역할을 맡은 AI간 대화를 통해 결과를 도출한다는 차이가 있습니다.

 

• Task Weaver는 사용할 언어모델(LLM)을 선택할 수 있습니다.
  • LLM 목록은 마이크소로프트의 OpenAI, Azure OpenAi뿐만 아니라 Google의 Gemini, 개인 학습한 LLM 모델까지 다양하게 사용할 수 있습니다.
  • 사용하는 LLM은 API 형태로 호출/응답할 수 있어야 합니다.

 

• 코드 스니펫은 파이썬 기반으로 작성됩니다.
  • 코드 스니펫 작성 후 테스트와 검증까지 진행해 실제로 명령한 프롬포트대로 동작하는 코드를 받을 수 있다는 장점이 있습니다.
    • 로컬에서 테스트가 진행되어 실제 로컬 파일을 읽고 결과를 도출할 수 있는지 확인 가능

• 분석 목적별로 추가 플러그인을 설정할 수 있습니다.
  • paper_summary: pdf파일을 읽고 내용 요약
  • sql_pull_data: DB의 데이터를 SQL로 가져옴
  • vision_web_explorer:
  • web_search: 웹에서 내용을 검색해서 관련 웹페이지 URL, 제목과 요약을 제공
  • document_retriever: 주어진 문서들 중 유사도가 높은 상위 K개에서 관련 내용 검색

 

• Task Weaver 작동 원리 

  

 

1. 사용자가 Task Weaver에 프롬포트를 입력한다.
2. Planner는 사용자의 프롬포트를 읽고 분석 단계를 생성한다.
3. 첫번째 단계를 CodeInterpreter에게 전달한다.
4. CodeInterpreter는 해당 내용을 코드로 구현한다. 구현된 코드는 로컬 환경에서 제대로 동작하는지 테스트, 검증한다.
5. 제대로 워킹하는 경우 Planner에게 해당 코드와 결과를 전달하며, Planner는 다음 계획을 다시 CodeInterpreter에게 전달하는 방식으로 마지막 단계까지 반복한다.5-1. 제대로 워킹하지 않으면 Code interpreter 내에서 코드를 수정한다.
6. 마지막 단계까지 잘 working 하면 Planner는 결과를 사용자에게 반환한다.

 

서비스 예시1 - 로컬 파일 이상치 탐색

• 입력 프롬포트 1: Human ▶ 특정 로컬 파일을 열어서 anomaly detection 해줘. Count 칼럼을 anomaly etection에 사용해. (Please load the data file /home/heenj/projects/TaskWeaver/project/sample_data/demo_data.csv then analyze Count column for anomaly detection.)

• 결과 1
  • 요청 사항대로 해당 csv 파일을 로드 후 이상치 탐색을 잘 진행하였다.

╭───< Planner >
 ├─► [init_plan]
 │   1. load the data file
 │   2. confirm the column to be detected anomalies <interactively depends on 1>
 │   3. detect anomalies on the loaded data <sequentially depends on 2>
 │   4. report the detected anomalies to the user <interactively depends on 3>
 ├─► [plan]
 │   1. instruct CodeInterpreter to load the data file
 │   2. ask the user to confirm the column to be detected anomalies
 │   3. instruct CodeInterpreter to detect anomalies on the loaded data
 │   4. report the detected anomalies to the user
 ├─► [current_plan_step] 1. instruct CodeInterpreter to load the data file
 ├──● Please load the data file /home/heenj/projects/vaiv/TaskWeaver/project/sample_data/demo_data.csv
 ╰──● sending message to CodeInterpreter
 ╭───< CodeInterpreter >
 ├─► [thought] ProgramApe will generate a code snippet to load the data file /home/heenj/projects/vaiv/TaskWeaver/project/sample_data/demo_data.csv.
 ├─► [python]
 │   import pandas as pd
 │
 │   data = pd.read_csv("/home/heenj/projects/vaiv/TaskWeaver/project/sample_data/demo_data.csv")
 │   loaded_data = data
 ├─► [verification] NONE
 ├─► [code_error] No code verification is performed.
 ├─► [execution_status] SUCCESS
 ├─► [execution_result]
 │   The execution of the generated python code above has succeeded
 │
 │   The result of above Python code after execution is:
 │                  TimeBucket  Count
 │   0    2023-02-01T00:00:00Z  67814
 │   1    2023-02-01T04:00:00Z  84569
 │   ..                    ...    ...
 │   718  2023-05-31T16:00:00Z  89912
 │   719  2023-05-31T20:00:00Z  87249
 │
 │   [720 rows x 2 columns]
 ├─► [artifact_paths] []
 ├──● The following python code has been executed:
 │   ```python
 │   import pandas as pd
 │
 │   data = pd.read_csv("/home/heenj/projects/vaiv/TaskWeaver/project/sample_data/demo_data.csv")
 │   loaded_data = data
 │   ```
 │
 │
 │   The execution of the generated python code above has succeeded
 │
 │   The result of above Python code after execution is:
 │                  TimeBucket  Count
 │   0    2023-02-01T00:00:00Z  67814
 │   1    2023-02-01T04:00:00Z  84569
 │   ..                    ...    ...
 │   718  2023-05-31T16:00:00Z  89912
 │   719  2023-05-31T20:00:00Z  87249
 │
 │   [720 rows x 2 columns]
 ╰──● sending message to Planner
 ╭───< Planner >
 ├─► [init_plan]
 │   1. load the data file
 │   2. confirm the column to be detected anomalies <interactively depends on 1>
 │   3. detect anomalies on the loaded data <sequentially depends on 2>
 │   4. report the detected anomalies to the user <interactively depends on 3>
 ├─► [plan]
 │   1. instruct CodeInterpreter to load the data file
 │   2. ask the user to confirm the column to be detected anomalies
 │   3. instruct CodeInterpreter to detect anomalies on the loaded data
 │   4. report the detected anomalies to the user
 ├─► [current_plan_step] 2. ask the user to confirm the column to be detected anomalies
 ├──● Please confirm the column to be detected anomalies
 ╰──● sending message to User
 TaskWeaver ▶  Please confirm the column to be detected anomalies
    Human   ▶  Count
 ╭───< Planner >
 ├─► [init_plan]
 │   1. load the data file
 │   2. confirm the column to be detected anomalies <interactively depends on 1>
 │   3. detect anomalies on the loaded data <sequentially depends on 2>
 │   4. report the detected anomalies to the user <interactively depends on 3>
 ├─► [plan]
 │   1. instruct CodeInterpreter to load the data file
 │   2. ask the user to confirm the column to be detected anomalies
 │   3. instruct CodeInterpreter to detect anomalies on the loaded data
 │   4. report the detected anomalies to the user
 ├─► [current_plan_step] 3. instruct CodeInterpreter to detect anomalies on the loaded data
 ├──● Please detect anomalies on the loaded data with column Count
 ╰──● sending message to CodeInterpreter
 ╭───< CodeInterpreter >
 ├─► [thought] ProgramApe will generate a code snippet to detect с on the loaded data with column Count.
 ├─► [python]
 │   import numpy as np
 │   from scipy.stats import zscore
 │
 │   loaded_data["Is_Anomaly"] = np.where(zscore(loaded_data["Count"]) > 3, True, False)
 │   anomaly_data = loaded_data[loaded_data["Is_Anomaly"] == True]
 │   anomaly_data
 ├─► [verification] NONE
 ├─► [code_error] No code verification is performed.
 ├─► [execution_status] SUCCESS
 ├─► [execution_result]
 │   The execution of the generated python code above has succeeded
 │
 │   The result of above Python code after execution is:
 │                  TimeBucket   Count  Is_Anomaly
 │   46   2023-02-08T16:00:00Z  128843        True
 │   308  2023-03-24T08:00:00Z  135922        True
 │   ..                    ...     ...         ...
 │   565  2023-05-06T04:00:00Z  134175        True
 │   680  2023-05-25T08:00:00Z  135727        True
 │
 │   [5 rows x 3 columns]
 ├─► [artifact_paths] []
 ├──● The following python code has been executed:
 │   ```python
 │   import numpy as np
 │   from scipy.stats import zscore
 │
 │   loaded_data["Is_Anomaly"] = np.where(zscore(loaded_data["Count"]) > 3, True, False)
 │   anomaly_data = loaded_data[loaded_data["Is_Anomaly"] == True]
 │   anomaly_data
 │   ```
 │
 │
 │   The execution of the generated python code above has succeeded
 │
 │   The result of above Python code after execution is:
 │                  TimeBucket   Count  Is_Anomaly
 │   46   2023-02-08T16:00:00Z  128843        True
 │   308  2023-03-24T08:00:00Z  135922        True
 │   ..                    ...     ...         ...
 │   565  2023-05-06T04:00:00Z  134175        True
 │   680  2023-05-25T08:00:00Z  135727        True
 │
 │   [5 rows x 3 columns]
 ╰──● sending message to Planner
 ╭───< Planner >
 ├─► [init_plan]
 │   1. load the data file
 │   2. confirm the column to be detected anomalies <interactively depends on 1>
 │   3. detect anomalies on the loaded data <sequentially depends on 2>
 │   4. report the detected anomalies to the user <interactively depends on 3>
 ├─► [plan]
 │   1. instruct CodeInterpreter to load the data file
 │   2. ask the user to confirm the column to be detected anomalies
 │   3. instruct CodeInterpreter to detect anomalies on the loaded data
 │   4. report the detected anomalies to the user
 ├─► [current_plan_step] 4. report the detected anomalies to the user
 ├──● The detected anomalies are:
 │
 │                  TimeBucket   Count  Is_Anomaly
 │   46   2023-02-08T16:00:00Z  128843        True
 │   308  2023-03-24T08:00:00Z  135922        True
 │   ..                    ...     ...         ...
 │   565  2023-05-06T04:00:00Z  134175        True
 │   680  2023-05-25T08:00:00Z  135727        True
 │
 │   [5 rows x 3 columns]
 ╰──● sending message to User
 TaskWeaver ▶  The detected anomalies are:

               TimeBucket   Count  Is_Anomaly
46   2023-02-08T16:00:00Z  128843        True
308  2023-03-24T08:00:00Z  135922        True
..                    ...     ...         ...
565  2023-05-06T04:00:00Z  134175        True
680  2023-05-25T08:00:00Z  135727        True

[5 rows x 3 columns]

 

 

• 입력 프롬포트 2: Human ▶ 이상치 데이터를 강조해서 시계열 데이터 plot을 그려줘. (Can you show me those data time-series plot highlighting anomaly data specially?)

• 결과 2
  • 입력 프롬포트 1의 내용을 기반으로 plot을 잘 그려줍니다.

Human   ▶  Can you show me those data time-series plot?
 ╭───< Planner >
 ├─► [init_plan]
 │   1. load the data file
 │   2. confirm the column to be detected anomalies <interactively depends on 1>
 │   3. detect anomalies on the loaded data <sequentially depends on 2>
 │   4. report the detected anomalies to the user <interactively depends on 3>
 │   5. plot the time-series of the data <interactively depends on 4>
 ├─► [plan]
 │   1. instruct CodeInterpreter to load the data file
 │   2. ask the user to confirm the column to be detected anomalies
 │   3. instruct CodeInterpreter to detect anomalies on the loaded data
 │   4. report the detected anomalies to the user
 │   5. instruct CodeInterpreter to plot the time-series of the data
 ├─► [current_plan_step] 5. instruct CodeInterpreter to plot the time-series of the data
 ├──● Please plot the time-series of the data
 ╰──● sending message to CodeInterpreter
 ╭───< CodeInterpreter >
 ├─► [thought] ProgramApe will generate a code snippet to plot the time-series of the data.
 ├─► [python]
 │   import matplotlib.pyplot as plt
 │
 │   plt.plot(loaded_data["TimeBucket"], loaded_data["Count"])
 │   plt.xlabel("Time")
 │   plt.ylabel("Count")
 │   plt.title("Time-Series Plot of the Data")
 │   plt.show()
 ├─► [verification] NONE
 ├─► [code_error] No code verification is performed.
 ├─► [execution_status] SUCCESS
 ├─► [execution_result]
 │   The execution of the generated python code above has succeeded
 │
 │   The execution is successful but no output is generated.
 │   The following artifacts were generated:
 │   - type: image ; uri: file:///home/heenj/projects/vaiv/TaskWeaver/project/workspace/sessions/20240206-042030-a4a427a6/cwd/post-20240206-043138-0aaf4e88-display-1_image.png ;
 │   description: <Figure size 2400x1200 with 1 Axes>
 │
 ├─► [artifact_paths] ['/home/heenj/projects/vaiv/TaskWeaver/project/workspace/sessions/20240206-042030-a4a427a6/cwd/post-20240206-043138-0aaf4e88-display-1_image.png']
 ├──● The following python code has been executed:
 │   ```python
 │   import matplotlib.pyplot as plt
 │
 │   plt.plot(loaded_data["TimeBucket"], loaded_data["Count"])
 │   plt.xlabel("Time")
 │   plt.ylabel("Count")
 │   plt.title("Time-Series Plot of the Data")
 │   plt.show()
 │   ```
 │
 │
 │   The execution of the generated python code above has succeeded
 │
 │   The execution is successful but no output is generated.
 │   The following artifacts were generated:
 │   - type: image ; uri: file:///home/heenj/projects/vaiv/TaskWeaver/project/workspace/sessions/20240206-042030-a4a427a6/cwd/post-20240206-043138-0aaf4e88-display-1_image.png ;
 │   description: <Figure size 2400x1200 with 1 Axes>
 │
 ╰──● sending message to Planner
 ╭───< Planner >
 ├─► [init_plan]
 │   1. load the data file
 │   2. confirm the column to be detected anomalies <interactively depends on 1>
 │   3. detect anomalies on the loaded data <sequentially depends on 2>
 │   4. report the detected anomalies to the user <interactively depends on 3>
 │   5. plot the time-series of the data <interactively depends on 4>
 ├─► [plan]
 │   1. instruct CodeInterpreter to load the data file
 │   2. ask the user to confirm the column to be detected anomalies
 │   3. instruct CodeInterpreter to detect anomalies on the loaded data
 │   4. report the detected anomalies to the user
 │   5. instruct CodeInterpreter to plot the time-series of the data
 ├─► [current_plan_step] 5. instruct CodeInterpreter to plot the time-series of the data
 ├──● The time-series plot of the data is generated and saved as [file:///home/heenj/projects/vaiv/TaskWeaver/project/workspace/sessions/20240206-042030-a4a427a6/cwd/post-20240206-
 │   043138-0aaf4e88-display-1_image.png](file:///home/heenj/projects/vaiv/TaskWeaver/project/workspace/sessions/20240206-042030-a4a427a6/cwd/post-20240206-043138-0aaf4e88-display-
 │   1_image.png)
 ╰──● sending message to User
 TaskWeaver ▶  The time-series plot of the data is generated and saved as [file:///home/heenj/projects/vaiv/TaskWeaver/project/workspace/sessions/20240206-042030-a4a427a6/cwd/post-20240206-043138-0aaf4e88-display-1_image.png](file:///home/heenj/projects/vaiv/TaskWeaver/project/workspace/sessions/20240206-042030-a4a427a6/cwd/post-20240206-043138-0aaf4e88-display-1_image.png)

'

• CLI 환경에서 실행하는 경우 image를 특정 경로에 저장해줍니다.

 

서비스 예시2 - 야후 파이낸스 주가 데이터

• 입력프롬포트 1: Human ▶ 2024년 2월 삼성전자 주가를 yahoo finance에서 가져와줘. (Can you get Samsung electronics price in Feburary, 2024 using yahoo finance?)

 

• 결과 1
  • yfinance 라이브러리를 이용해 삼성전자 주가를 가져오는걸 볼 수 있습니다.
  • 로컬에 yfinance 라이브러리가 설치돼 있지 않아서 처음 테스트에서 에러가 발생했습니다.
  • 에러를 디버깅해 !pip install 을 사용해서 임시로 설치하고 결과를 출력하였습니다.

╭───< Planner >
 ├─► [init_plan]
 │   1. confirm the product name <interactively depends on 1>
 │   2. confirm the month and year <interactively depends on 1>
 │   3. get the price data from yahoo finance <sequentially depends on 2>
 │   4. report the price data to the user <interactively depends on 3>
 ├─► [plan]
 │   1. ask the user to confirm the product name
 │   2. ask the user to confirm the month and year
 │   3. instruct CodeInterpreter to get the price data from yahoo finance
 │   4. report the price data to the user
 ├─► [current_plan_step] 1. ask the user to confirm the product name
 ├──● Please confirm the product name
 ╰──● sending message to User
 TaskWeaver ▶  Please confirm the product name
    Human   ▶  Samsung electronics
 ╭───< Planner >
 ├─► [init_plan]
 │   1. confirm the product name <interactively depends on 1>
 │   2. confirm the month and year <interactively depends on 1>
 │   3. get the price data from yahoo finance <sequentially depends on 2>
 │   4. report the price data to the user <interactively depends on 3>
 ├─► [plan]
 │   1. ask the user to confirm the product name
 │   2. ask the user to confirm the month and year
 │   3. instruct CodeInterpreter to get the price data from yahoo finance
 │   4. report the price data to the user
 ├─► [current_plan_step] 2. ask the user to confirm the month and year
 ├──● Please confirm the month and year
 ╰──● sending message to User
 TaskWeaver ▶  Please confirm the month and year
    Human   ▶  02/2024
 ╭───< Planner >
 ├─► [init_plan]
 │   1. confirm the product name <interactively depends on 1>
 │   2. confirm the month and year <interactively depends on 1>
 │   3. get the price data from yahoo finance <sequentially depends on 2>
 │   4. report the price data to the user <interactively depends on 3>
 ├─► [plan]
 │   1. ask the user to confirm the product name
 │   2. ask the user to confirm the month and year
 │   3. instruct CodeInterpreter to get the price data from yahoo finance
 │   4. report the price data to the user
 ├─► [current_plan_step] 3. instruct CodeInterpreter to get the price data from yahoo finance
 ├──● Please get the price data of Samsung electronics in 02/2024 from yahoo finance
 ╰──● sending message to CodeInterpreter
 ╭───< CodeInterpreter >
 ├─► [thought] ProgramApe will generate a code snippet to get the price data of Samsung electronics in 02/2024 from yahoo finance.
 ├─► [python]
 │   import yfinance as yf
 │
 │   ticker = "005930.KS"
 │   start_date = "2024-02-01"
 │   end_date = "2024-02-29"
 │
 │   data = yf.download(ticker, start_date, end_date)
 │   price_data = data["Close"]
 ├─► [verification] NONE
 ├─► [code_error] No code verification is performed.
 ├─► [execution_status] FAILURE
 ├─► [execution_result]
 │   The execution of the generated python code above has failed
 │
 │   During execution, the following messages were logged:
 │   Traceback (most recent call last):
 │
 │     Cell In[5], line 1
 │       import yfinance as yf
 │
 │   ModuleNotFoundError: No module named 'yfinance'
 │
 │
 ├─► [artifact_paths] []
 ├──● The following python code has been executed:
 │   ```python
 │   import yfinance as yf
 │
 │   ticker = "005930.KS"
 │   start_date = "2024-02-01"
 │   end_date = "2024-02-29"
 │
 │   data = yf.download(ticker, start_date, end_date)
 │   price_data = data["Close"]
 │   ```
 │
 │
 │   The execution of the generated python code above has failed
 │
 │   During execution, the following messages were logged:
 │   Traceback (most recent call last):
 │
 │     Cell In[5], line 1
 │       import yfinance as yf
 │
 │   ModuleNotFoundError: No module named 'yfinance'
 │
 │
 ├─► [revise_message]
 │   The execution of the previous generated code has failed. If you think you can fix the problem by rewriting the code, please generate code and run it again.
 │   Otherwise, please explain the problem to me.
 ╰──● sending message to CodeInterpreter
 ╭───< CodeInterpreter >
 ├─► [thought] ProgramApe understands that the previous code failed to execute due to the import of the yfinance module, which is not pre-installed.
 ├─► [thought] ProgramApe will attempt to install the yfinance module and then generate a new code snippet to get the price data of Samsung electronics in 02/2024 from yahoo
 │   finance.
 ├─► [python]
 │   !pip install yfinance
 │
 │   import yfinance as yf
 │
 │   ticker = "005930.KS"
 │   start_date = "2024-02-01"
 │   end_date = "2024-02-29"
 │
 │   data = yf.download(ticker, start_date, end_date)
 │   price_data = data["Close"]
 ├─► [verification] NONE
 ├─► [code_error] No code verification is performed.
 ├─► [execution_status] SUCCESS
 ├─► [execution_result]
 │   The execution of the generated python code above has succeeded
 │
 │   The result of above Python code after execution is:
 │   Date
 │   2024-02-01    73600.0
 │   2024-02-02    75200.0
 │   2024-02-05    74300.0
 │   2024-02-06    74600.0
 │   Name: Close, dtype: float64
 ├─► [artifact_paths] []
 ├──● The following python code has been executed:
 │   ```python
 │   !pip install yfinance
 │
 │   import yfinance as yf
 │
 │   ticker = "005930.KS"
 │   start_date = "2024-02-01"
 │   end_date = "2024-02-29"
 │
 │   data = yf.download(ticker, start_date, end_date)
 │   price_data = data["Close"]
 │   ```
 │
 │
 │   The execution of the generated python code above has succeeded
 │
 │   The result of above Python code after execution is:
 │   Date
 │   2024-02-01    73600.0
 │   2024-02-02    75200.0
 │   2024-02-05    74300.0
 │   2024-02-06    74600.0
 │   Name: Close, dtype: float64
 ╰──● sending message to Planner
 ╭───< Planner >
 ├─► [init_plan]
 │   1. confirm the product name <interactively depends on 1>
 │   2. confirm the month and year <interactively depends on 1>
 │   3. get the price data from yahoo finance <sequentially depends on 2>
 │   4. report the price data to the user <interactively depends on 3>
 ├─► [plan]
 │   1. ask the user to confirm the product name
 │   2. ask the user to confirm the month and year
 │   3. instruct CodeInterpreter to get the price data from yahoo finance
 │   4. report the price data to the user
 ├─► [current_plan_step] 4. report the price data to the user
 ├──● The price data of Samsung electronics in 02/2024 is:
 │
 │   Date
 │   2024-02-01    73600.0
 │   2024-02-02    75200.0
 │   2024-02-05    74300.0
 │   2024-02-06    74600.0
 │   Name: Close, dtype: float64
 ╰──● sending message to User
 TaskWeaver ▶  The price data of Samsung electronics in 02/2024 is:

Date
2024-02-01    73600.0
2024-02-02    75200.0
2024-02-05    74300.0
2024-02-06    74600.0
Name: Close, dtype: float64

 

• 아래 사진은 2023년 삼성전자 주가를 가져온 후 향후 7일간 주가를 예측해서 plot을 그려달라고한 결과입니다.

 

 

서비스 예시 3: paper_summary

• 파이썬 pdftotext 라이브러리를 통해 단순히 pdf를 글자로 변환하고 이를 프롬포트로 넣는걸 확인했습니다.
  • 특히 표가 들어간 경우 형식 보존이 안되어 결과 품질이 좋지 못했습니다.

 

 

서비스 이용 이슈

• 어느 LLM을 사용하느냐에 따라 사용자 경험에 차이가 발생할 수 있습니다..
  • Gemini 경우 한글 이해도가 많이 떨어져 결과가 좋지 않습니다.
  • LLM별로 발생할 수 있는 에러 이슈들을 해결하기 위해 내부 코드들을 수정할 필요가 있습니다.

• 요청사항이 많아질수록 시간이 많이 소요되며 에러가 발생하기 쉽습니다.
  • 프롬포트를 통해 계획을 세우는 Planner와 코드를 작성하는 Code Interpreter 역할간 대화(interaction)가 많아지기 때문입니다.
  • 특히 대화가 길어지면 특정 LLM들은 최대 토큰 수 정책에 따른 에러도 발생하기 쉽습니다.

• 프롬프트로 처음에 설정한 내용을 Planner가 다시 물어보는 경우가 있습니다.
  • 조금 귀찮지만 요청한 내용을 정확하게 정의할 수 있어 나쁘진 않아 보입니다.

 

비용

• 코드는 오픈소스로 공개돼 있습니다.

• 다만 사용할 LLM의 API Key가 필요합니다.

 

 

참조

https://www.microsoft.com/en-us/windows/copilot-ai-features?r=1#faq

https://copilot.microsoft.com/

[최초공개] M365 코파일럿 첫 체험기! 2주 사용 후 느낀, 장단점 솔직 리뷰