mirror of https://github.com/microsoft/autogen.git
476 lines
33 KiB
Plaintext
476 lines
33 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"attachments": {},
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"# Group Chat with Customized Speaker Selection Method\n",
|
|
"\n",
|
|
"AutoGen offers conversable agents powered by LLM, tool or human, which can be used to perform tasks collectively via automated chat. This framework allows tool use and human participation through multi-agent conversation.\n",
|
|
"Please find documentation about this feature [here](https://microsoft.github.io/autogen/docs/Use-Cases/agent_chat).\n",
|
|
"\n",
|
|
"In this notebook, we demonstrate how to pass a cumstomized agent selection method to GroupChat. The customized function looks like this:\n",
|
|
"\n",
|
|
"```python\n",
|
|
"def custom_speaker_selection_func(last_speaker, groupchat):\n",
|
|
" \"\"\"Define a customized speaker selection function.\n",
|
|
" A recommended way is to define a transition for each speaker in the groupchat.\n",
|
|
"\n",
|
|
" Parameters:\n",
|
|
" - last_speaker: Agent\n",
|
|
" The last speaker in the group chat.\n",
|
|
" - groupchat: GroupChat\n",
|
|
" The GroupChat object\n",
|
|
" Return:\n",
|
|
" Return one of the following:\n",
|
|
" 1. an `Agent` class, it must be one of the agents in the group chat.\n",
|
|
" 2. a string from ['auto', 'manual', 'random', 'round_robin'] to select a default method to use.\n",
|
|
" 3. None, which indicates the chat should be terminated.\n",
|
|
" \"\"\"\n",
|
|
" pass\n",
|
|
"\n",
|
|
"groupchat = autogen.GroupChat(\n",
|
|
" speaker_selection_method=custom_speaker_selection_func,\n",
|
|
" ...,\n",
|
|
")\n",
|
|
"```\n",
|
|
"The last speaker and the groupchat object are passed to the function. Commonly used variables from groupchat are `groupchat.messages` an `groupchat.agents`, which is the message history and the agents in the group chat respectively. You can access other attributes of the groupchat, such as `groupchat.allowed_speaker_transitions_dict` for pre-defined allowed_speaker_transitions_dict. \n",
|
|
"\n",
|
|
"````{=mdx}\n",
|
|
":::info Requirements\n",
|
|
"Install `pyautogen`:\n",
|
|
"```bash\n",
|
|
"pip install pyautogen\n",
|
|
"```\n",
|
|
"\n",
|
|
"For more information, please refer to the [installation guide](/docs/installation/).\n",
|
|
":::\n",
|
|
"````"
|
|
]
|
|
},
|
|
{
|
|
"attachments": {},
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Set your API Endpoint\n",
|
|
"\n",
|
|
"The [`config_list_from_json`](https://microsoft.github.io/autogen/docs/reference/oai/openai_utils#config_list_from_json) function loads a list of configurations from an environment variable or a json file."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"import autogen\n",
|
|
"\n",
|
|
"config_list = autogen.config_list_from_json(\n",
|
|
" \"OAI_CONFIG_LIST\",\n",
|
|
" filter_dict={\n",
|
|
" \"model\": [\"gpt-4\", \"gpt-4-1106-preview\"],\n",
|
|
" },\n",
|
|
")"
|
|
]
|
|
},
|
|
{
|
|
"attachments": {},
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"````{=mdx}\n",
|
|
":::tip\n",
|
|
"Learn more about configuring LLMs for agents [here](/docs/topics/llm_configuration).\n",
|
|
":::\n",
|
|
"````\n",
|
|
"\n",
|
|
"## Construct Agents\n",
|
|
"\n",
|
|
"- Planner: Give a plan and revise.\n",
|
|
"- Admin: Human in the loop to approve or terminate the process.\n",
|
|
"- Engineer: Retrieve papers from the internet by writing code.\n",
|
|
"- Executor: Execute the code.\n",
|
|
"- Scientist: Read the papers and write a summary.\n",
|
|
"\n",
|
|
"The pipeline is the following:\n",
|
|
"\n",
|
|
"1. The planner interact with Admin (user) to revise a plan. Only when the Admin types \"Approve\", we can move to the next step.\n",
|
|
"2. The engineer will write code to retrieve papers from the internet. The code will be executed by executor.\n",
|
|
"3. When the code is executed successfully, the scientist will read the papers and write a summary.\n",
|
|
"4. The summary will be reviewed by the Admin and give comments. When the Admin types \"TERMINATE\", the process will be terminated.\n"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 2,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"gpt4_config = {\n",
|
|
" \"cache_seed\": 42, # change the cache_seed for different trials\n",
|
|
" \"temperature\": 0,\n",
|
|
" \"config_list\": config_list,\n",
|
|
" \"timeout\": 120,\n",
|
|
"}\n",
|
|
"\n",
|
|
"planner = autogen.AssistantAgent(\n",
|
|
" name=\"Planner\",\n",
|
|
" system_message=\"\"\"Planner. Suggest a plan. Revise the plan based on feedback from admin and critic, until admin approval.\n",
|
|
"The plan may involve an engineer who can write code and a scientist who doesn't write code.\n",
|
|
"Explain the plan first. Be clear which step is performed by an engineer, and which step is performed by a scientist.\n",
|
|
"\"\"\",\n",
|
|
" llm_config=gpt4_config,\n",
|
|
")\n",
|
|
"\n",
|
|
"user_proxy = autogen.UserProxyAgent(\n",
|
|
" name=\"Admin\",\n",
|
|
" system_message=\"A human admin. Interact with the planner to discuss the plan. Plan execution needs to be approved by this admin.\",\n",
|
|
" code_execution_config=False,\n",
|
|
")\n",
|
|
"\n",
|
|
"engineer = autogen.AssistantAgent(\n",
|
|
" name=\"Engineer\",\n",
|
|
" llm_config=gpt4_config,\n",
|
|
" system_message=\"\"\"Engineer. You follow an approved plan. You write python/shell code to solve tasks. Wrap the code in a code block that specifies the script type. The user can't modify your code. So do not suggest incomplete code which requires others to modify. Don't use a code block if it's not intended to be executed by the executor.\n",
|
|
"Don't include multiple code blocks in one response. Do not ask others to copy and paste the result. Check the execution result returned by the executor.\n",
|
|
"If the result indicates there is an error, fix the error and output the code again. Suggest the full code instead of partial code or code changes. If the error can't be fixed or if the task is not solved even after the code is executed successfully, analyze the problem, revisit your assumption, collect additional info you need, and think of a different approach to try.\n",
|
|
"\"\"\",\n",
|
|
")\n",
|
|
"scientist = autogen.AssistantAgent(\n",
|
|
" name=\"Scientist\",\n",
|
|
" llm_config=gpt4_config,\n",
|
|
" system_message=\"\"\"Scientist. You follow an approved plan. You are able to categorize papers after seeing their abstracts printed. You don't write code.\"\"\",\n",
|
|
")\n",
|
|
"\n",
|
|
"executor = autogen.UserProxyAgent(\n",
|
|
" name=\"Executor\",\n",
|
|
" system_message=\"Executor. Execute the code written by the engineer and report the result.\",\n",
|
|
" human_input_mode=\"NEVER\",\n",
|
|
" code_execution_config={\n",
|
|
" \"last_n_messages\": 3,\n",
|
|
" \"work_dir\": \"paper\",\n",
|
|
" \"use_docker\": False,\n",
|
|
" }, # Please set use_docker=True if docker is available to run the generated code. Using docker is safer than running the generated code directly.\n",
|
|
")\n",
|
|
"\n",
|
|
"from typing import Dict, List\n",
|
|
"\n",
|
|
"from autogen import Agent\n",
|
|
"\n",
|
|
"\n",
|
|
"def custom_speaker_selection_func(last_speaker: Agent, groupchat: autogen.GroupChat):\n",
|
|
" \"\"\"Define a customized speaker selection function.\n",
|
|
" A recommended way is to define a transition for each speaker in the groupchat.\n",
|
|
"\n",
|
|
" Returns:\n",
|
|
" Return an `Agent` class or a string from ['auto', 'manual', 'random', 'round_robin'] to select a default method to use.\n",
|
|
" \"\"\"\n",
|
|
" messages = groupchat.messages\n",
|
|
"\n",
|
|
" if len(messages) <= 1:\n",
|
|
" return planner\n",
|
|
"\n",
|
|
" if last_speaker is user_proxy:\n",
|
|
" if \"Approve\" in messages[-1][\"content\"]:\n",
|
|
" # If the last message is approved, let the engineer to speak\n",
|
|
" return engineer\n",
|
|
" elif messages[-2][\"name\"] == \"Planner\":\n",
|
|
" # If it is the planning stage, let the planner to continue\n",
|
|
" return planner\n",
|
|
" elif messages[-2][\"name\"] == \"Scientist\":\n",
|
|
" # If the last message is from the scientist, let the scientist to continue\n",
|
|
" return scientist\n",
|
|
"\n",
|
|
" elif last_speaker is planner:\n",
|
|
" # Always let the user to speak after the planner\n",
|
|
" return user_proxy\n",
|
|
"\n",
|
|
" elif last_speaker is engineer:\n",
|
|
" if \"```python\" in messages[-1][\"content\"]:\n",
|
|
" # If the last message is a python code block, let the executor to speak\n",
|
|
" return executor\n",
|
|
" else:\n",
|
|
" # Otherwise, let the engineer to continue\n",
|
|
" return engineer\n",
|
|
"\n",
|
|
" elif last_speaker is executor:\n",
|
|
" if \"exitcode: 1\" in messages[-1][\"content\"]:\n",
|
|
" # If the last message indicates an error, let the engineer to improve the code\n",
|
|
" return engineer\n",
|
|
" else:\n",
|
|
" # Otherwise, let the scientist to speak\n",
|
|
" return scientist\n",
|
|
"\n",
|
|
" elif last_speaker is scientist:\n",
|
|
" # Always let the user to speak after the scientist\n",
|
|
" return user_proxy\n",
|
|
"\n",
|
|
" else:\n",
|
|
" return \"random\"\n",
|
|
"\n",
|
|
"\n",
|
|
"groupchat = autogen.GroupChat(\n",
|
|
" agents=[user_proxy, engineer, scientist, planner, executor],\n",
|
|
" messages=[],\n",
|
|
" max_round=20,\n",
|
|
" speaker_selection_method=custom_speaker_selection_func,\n",
|
|
")\n",
|
|
"manager = autogen.GroupChatManager(groupchat=groupchat, llm_config=gpt4_config)"
|
|
]
|
|
},
|
|
{
|
|
"attachments": {},
|
|
"cell_type": "markdown",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Start Chat"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 8,
|
|
"metadata": {},
|
|
"outputs": [
|
|
{
|
|
"name": "stdout",
|
|
"output_type": "stream",
|
|
"text": [
|
|
"\u001b[33mAdmin\u001b[0m (to chat_manager):\n",
|
|
"\n",
|
|
"Find a latest paper about gpt-4 on arxiv and find its potential applications in software.\n",
|
|
"\n",
|
|
"--------------------------------------------------------------------------------\n",
|
|
"\u001b[33mPlanner\u001b[0m (to chat_manager):\n",
|
|
"\n",
|
|
"**Initial Plan:**\n",
|
|
"\n",
|
|
"1. **Scientist's Task: Literature Review**\n",
|
|
" - The scientist will conduct a comprehensive literature review to find the latest paper about GPT-4 on arXiv. This involves using search queries related to GPT-4 and filtering results by the most recent publications.\n",
|
|
"\n",
|
|
"2. **Scientist's Task: Analysis of the Paper**\n",
|
|
" - Once the latest paper is identified, the scientist will read through the paper to understand its contents, focusing on the methodology, results, and discussions about potential applications in software.\n",
|
|
"\n",
|
|
"3. **Scientist's Task: Identifying Potential Applications**\n",
|
|
" - The scientist will then brainstorm and list potential applications of GPT-4 in software, based on the findings from the paper. This may include applications in natural language processing, code generation, chatbots, and more.\n",
|
|
"\n",
|
|
"4. **Engineer's Task: Technical Feasibility Assessment**\n",
|
|
" - The engineer will review the list of potential applications provided by the scientist and assess the technical feasibility of each application. This involves considering the current state of software technology, the capabilities of GPT-4, and the practicality of integrating GPT-4 into existing systems.\n",
|
|
"\n",
|
|
"5. **Engineer's Task: Prototype Development Plan**\n",
|
|
" - For applications deemed technically feasible, the engineer will draft a plan for developing a prototype that demonstrates the use of GPT-4 in a software application. This plan will outline the required resources, estimated timeline, and the steps for implementation.\n",
|
|
"\n",
|
|
"6. **Joint Task: Finalizing the Plan**\n",
|
|
" - The scientist and engineer will collaborate to finalize the plan, ensuring that it is scientifically sound and technically viable. They will prepare a document detailing the plan for potential applications and the prototype development.\n",
|
|
"\n",
|
|
"7. **Presentation to Admin**\n",
|
|
" - The finalized plan will be presented to the admin for approval. The admin will review the plan and provide feedback.\n",
|
|
"\n",
|
|
"8. **Revisions Based on Feedback**\n",
|
|
" - Based on the admin's feedback, the scientist and engineer will make necessary revisions to the plan. This iterative process will continue until the admin approves the plan.\n",
|
|
"\n",
|
|
"**Awaiting Admin's Feedback:** Please review the initial plan and provide feedback on any adjustments or additional details you would like to see.\n",
|
|
"\n",
|
|
"--------------------------------------------------------------------------------\n",
|
|
"\u001b[33mAdmin\u001b[0m (to chat_manager):\n",
|
|
"\n",
|
|
"Approve\n",
|
|
"\n",
|
|
"--------------------------------------------------------------------------------\n",
|
|
"\u001b[33mEngineer\u001b[0m (to chat_manager):\n",
|
|
"\n",
|
|
"Since the plan has been approved, I will now proceed with the first step, which is to find the latest paper about GPT-4 on arXiv. To do this, I will write a Python script that uses the arXiv API to search for papers related to GPT-4 and filter them by the most recent publications.\n",
|
|
"\n",
|
|
"Here is the Python script that accomplishes this task:\n",
|
|
"\n",
|
|
"```python\n",
|
|
"import requests\n",
|
|
"from datetime import datetime\n",
|
|
"\n",
|
|
"# Define the URL for the arXiv API\n",
|
|
"ARXIV_API_URL = \"http://export.arxiv.org/api/query\"\n",
|
|
"\n",
|
|
"# Define the search parameters\n",
|
|
"search_query = \"all:gpt-4\"\n",
|
|
"start = 0\n",
|
|
"max_results = 1\n",
|
|
"sort_by = \"submittedDate\"\n",
|
|
"sort_order = \"descending\"\n",
|
|
"\n",
|
|
"# Construct the query\n",
|
|
"query_params = {\n",
|
|
" \"search_query\": search_query,\n",
|
|
" \"start\": start,\n",
|
|
" \"max_results\": max_results,\n",
|
|
" \"sortBy\": sort_by,\n",
|
|
" \"sortOrder\": sort_order\n",
|
|
"}\n",
|
|
"\n",
|
|
"# Send the request to the arXiv API\n",
|
|
"response = requests.get(ARXIV_API_URL, params=query_params)\n",
|
|
"\n",
|
|
"# Check if the request was successful\n",
|
|
"if response.status_code == 200:\n",
|
|
" # Parse the response\n",
|
|
" feed = response.text\n",
|
|
" # Find the entry element, which contains the paper information\n",
|
|
" start_entry = feed.find('<entry>')\n",
|
|
" end_entry = feed.find('</entry>')\n",
|
|
" entry = feed[start_entry:end_entry]\n",
|
|
" \n",
|
|
" # Extract the title\n",
|
|
" start_title = entry.find('<title>') + 7\n",
|
|
" end_title = entry.find('</title>')\n",
|
|
" title = entry[start_title:end_title].strip()\n",
|
|
" \n",
|
|
" # Extract the published date\n",
|
|
" start_published = entry.find('<published>') + 12\n",
|
|
" end_published = entry.find('</published>')\n",
|
|
" published = entry[start_published:end_published].strip()\n",
|
|
" \n",
|
|
" # Extract the summary\n",
|
|
" start_summary = entry.find('<summary>') + 9\n",
|
|
" end_summary = entry.find('</summary>')\n",
|
|
" summary = entry[start_summary:end_summary].strip()\n",
|
|
" \n",
|
|
" # Extract the authors\n",
|
|
" authors = []\n",
|
|
" start_author = entry.find('<author>')\n",
|
|
" end_author = entry.find('</author>')\n",
|
|
" while start_author != -1 and end_author != -1:\n",
|
|
" start_name = entry.find('<name>', start_author) + 6\n",
|
|
" end_name = entry.find('</name>', start_author)\n",
|
|
" author_name = entry[start_name:end_name].strip()\n",
|
|
" authors.append(author_name)\n",
|
|
" start_author = entry.find('<author>', end_author)\n",
|
|
" end_author = entry.find('</author>', start_author)\n",
|
|
" \n",
|
|
" # Print the results\n",
|
|
" print(f\"Title: {title}\")\n",
|
|
" print(f\"Published Date: {published}\")\n",
|
|
" print(f\"Authors: {', '.join(authors)}\")\n",
|
|
" print(f\"Summary: {summary}\")\n",
|
|
"else:\n",
|
|
" print(\"Failed to retrieve data from arXiv API.\")\n",
|
|
"```\n",
|
|
"\n",
|
|
"This script will output the title, published date, authors, and summary of the most recent paper related to GPT-4 on arXiv. Please note that the actual content of the paper and its potential applications in software will need to be analyzed manually after retrieving the paper information.\n",
|
|
"\n",
|
|
"--------------------------------------------------------------------------------\n",
|
|
"\u001b[31m\n",
|
|
">>>>>>>> EXECUTING CODE BLOCK 0 (inferred language is python)...\u001b[0m\n",
|
|
"\u001b[33mExecutor\u001b[0m (to chat_manager):\n",
|
|
"\n",
|
|
"exitcode: 0 (execution succeeded)\n",
|
|
"Code output: \n",
|
|
"Title: A Data-Centric Approach To Generate Faithful and High Quality Patient\n",
|
|
" Summaries with Large Language Models\n",
|
|
"Published Date: 024-02-23T16:32:28Z\n",
|
|
"Authors: Stefan Hegselmann, Shannon Zejiang Shen, Florian Gierse, Monica Agrawal, David Sontag, Xiaoyi Jiang\n",
|
|
"Summary: Patients often face difficulties in understanding their hospitalizations,\n",
|
|
"while healthcare workers have limited resources to provide explanations. In\n",
|
|
"this work, we investigate the potential of large language models to generate\n",
|
|
"patient summaries based on doctors' notes and study the effect of training data\n",
|
|
"on the faithfulness and quality of the generated summaries. To this end, we\n",
|
|
"develop a rigorous labeling protocol for hallucinations, and have two medical\n",
|
|
"experts annotate 100 real-world summaries and 100 generated summaries. We show\n",
|
|
"that fine-tuning on hallucination-free data effectively reduces hallucinations\n",
|
|
"from 2.60 to 1.55 per summary for Llama 2, while preserving relevant\n",
|
|
"information. Although the effect is still present, it is much smaller for GPT-4\n",
|
|
"when prompted with five examples (0.70 to 0.40). We also conduct a qualitative\n",
|
|
"evaluation using hallucination-free and improved training data. GPT-4 shows\n",
|
|
"very good results even in the zero-shot setting. We find that common\n",
|
|
"quantitative metrics do not correlate well with faithfulness and quality.\n",
|
|
"Finally, we test GPT-4 for automatic hallucination detection, which yields\n",
|
|
"promising results.\n",
|
|
"\n",
|
|
"\n",
|
|
"--------------------------------------------------------------------------------\n",
|
|
"\u001b[33mScientist\u001b[0m (to chat_manager):\n",
|
|
"\n",
|
|
"Based on the abstract provided, the paper titled \"A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Models\" explores the use of large language models, including GPT-4, to generate patient summaries from doctors' notes. The study focuses on the impact of training data on the faithfulness and quality of the generated summaries and also investigates the potential of GPT-4 for automatic hallucination detection.\n",
|
|
"\n",
|
|
"**Potential Applications in Software:**\n",
|
|
"\n",
|
|
"1. **Healthcare Documentation Automation:**\n",
|
|
" - GPT-4 could be used to develop software that assists healthcare professionals in creating accurate and comprehensive patient summaries by automatically processing doctors' notes and other medical records.\n",
|
|
"\n",
|
|
"2. **Clinical Decision Support Systems:**\n",
|
|
" - Integrating GPT-4 into clinical decision support systems could provide healthcare workers with insights and suggestions based on a patient's medical history, potentially improving diagnosis and treatment planning.\n",
|
|
"\n",
|
|
"3. **Patient Education and Communication:**\n",
|
|
" - Software applications could leverage GPT-4 to translate complex medical information into patient-friendly summaries, enhancing patient understanding of their health conditions and treatments.\n",
|
|
"\n",
|
|
"4. **Medical Training and Simulation:**\n",
|
|
" - GPT-4 could be used to create realistic medical scenarios for training medical students and professionals, simulating patient interactions and generating case studies.\n",
|
|
"\n",
|
|
"5. **Data Quality Assurance:**\n",
|
|
" - The paper suggests that GPT-4 can be used for automatic hallucination detection, which refers to the identification of inaccuracies or fabrications in generated text. This could be applied to software that ensures the quality and reliability of medical documentation.\n",
|
|
"\n",
|
|
"6. **Research and Development:**\n",
|
|
" - GPT-4 could assist researchers in summarizing and synthesizing large volumes of medical literature, aiding in the discovery of new insights and the development of novel treatments.\n",
|
|
"\n",
|
|
"7. **Personalized Health Monitoring:**\n",
|
|
" - Software applications could use GPT-4 to provide personalized health monitoring and advice by analyzing user input, such as symptoms or lifestyle factors, and generating tailored health recommendations.\n",
|
|
"\n",
|
|
"These potential applications highlight the versatility of GPT-4 in the realm of healthcare software, offering opportunities to enhance patient care, improve healthcare workflows, and support medical education and research.\n",
|
|
"\n",
|
|
"--------------------------------------------------------------------------------\n",
|
|
"\u001b[33mAdmin\u001b[0m (to chat_manager):\n",
|
|
"\n",
|
|
"TERMINATE\n",
|
|
"\n",
|
|
"--------------------------------------------------------------------------------\n"
|
|
]
|
|
},
|
|
{
|
|
"data": {
|
|
"text/plain": [
|
|
"ChatResult(chat_id=None, chat_history=[{'content': 'Find a latest paper about gpt-4 on arxiv and find its potential applications in software.', 'role': 'assistant'}, {'content': \"**Initial Plan:**\\n\\n1. **Scientist's Task: Literature Review**\\n - The scientist will conduct a comprehensive literature review to find the latest paper about GPT-4 on arXiv. This involves using search queries related to GPT-4 and filtering results by the most recent publications.\\n\\n2. **Scientist's Task: Analysis of the Paper**\\n - Once the latest paper is identified, the scientist will read through the paper to understand its contents, focusing on the methodology, results, and discussions about potential applications in software.\\n\\n3. **Scientist's Task: Identifying Potential Applications**\\n - The scientist will then brainstorm and list potential applications of GPT-4 in software, based on the findings from the paper. This may include applications in natural language processing, code generation, chatbots, and more.\\n\\n4. **Engineer's Task: Technical Feasibility Assessment**\\n - The engineer will review the list of potential applications provided by the scientist and assess the technical feasibility of each application. This involves considering the current state of software technology, the capabilities of GPT-4, and the practicality of integrating GPT-4 into existing systems.\\n\\n5. **Engineer's Task: Prototype Development Plan**\\n - For applications deemed technically feasible, the engineer will draft a plan for developing a prototype that demonstrates the use of GPT-4 in a software application. This plan will outline the required resources, estimated timeline, and the steps for implementation.\\n\\n6. **Joint Task: Finalizing the Plan**\\n - The scientist and engineer will collaborate to finalize the plan, ensuring that it is scientifically sound and technically viable. They will prepare a document detailing the plan for potential applications and the prototype development.\\n\\n7. **Presentation to Admin**\\n - The finalized plan will be presented to the admin for approval. The admin will review the plan and provide feedback.\\n\\n8. **Revisions Based on Feedback**\\n - Based on the admin's feedback, the scientist and engineer will make necessary revisions to the plan. This iterative process will continue until the admin approves the plan.\\n\\n**Awaiting Admin's Feedback:** Please review the initial plan and provide feedback on any adjustments or additional details you would like to see.\", 'name': 'Planner', 'role': 'user'}, {'content': 'Approve', 'role': 'assistant'}, {'content': 'Since the plan has been approved, I will now proceed with the first step, which is to find the latest paper about GPT-4 on arXiv. To do this, I will write a Python script that uses the arXiv API to search for papers related to GPT-4 and filter them by the most recent publications.\\n\\nHere is the Python script that accomplishes this task:\\n\\n```python\\nimport requests\\nfrom datetime import datetime\\n\\n# Define the URL for the arXiv API\\nARXIV_API_URL = \"http://export.arxiv.org/api/query\"\\n\\n# Define the search parameters\\nsearch_query = \"all:gpt-4\"\\nstart = 0\\nmax_results = 1\\nsort_by = \"submittedDate\"\\nsort_order = \"descending\"\\n\\n# Construct the query\\nquery_params = {\\n \"search_query\": search_query,\\n \"start\": start,\\n \"max_results\": max_results,\\n \"sortBy\": sort_by,\\n \"sortOrder\": sort_order\\n}\\n\\n# Send the request to the arXiv API\\nresponse = requests.get(ARXIV_API_URL, params=query_params)\\n\\n# Check if the request was successful\\nif response.status_code == 200:\\n # Parse the response\\n feed = response.text\\n # Find the entry element, which contains the paper information\\n start_entry = feed.find(\\'<entry>\\')\\n end_entry = feed.find(\\'</entry>\\')\\n entry = feed[start_entry:end_entry]\\n \\n # Extract the title\\n start_title = entry.find(\\'<title>\\') + 7\\n end_title = entry.find(\\'</title>\\')\\n title = entry[start_title:end_title].strip()\\n \\n # Extract the published date\\n start_published = entry.find(\\'<published>\\') + 12\\n end_published = entry.find(\\'</published>\\')\\n published = entry[start_published:end_published].strip()\\n \\n # Extract the summary\\n start_summary = entry.find(\\'<summary>\\') + 9\\n end_summary = entry.find(\\'</summary>\\')\\n summary = entry[start_summary:end_summary].strip()\\n \\n # Extract the authors\\n authors = []\\n start_author = entry.find(\\'<author>\\')\\n end_author = entry.find(\\'</author>\\')\\n while start_author != -1 and end_author != -1:\\n start_name = entry.find(\\'<name>\\', start_author) + 6\\n end_name = entry.find(\\'</name>\\', start_author)\\n author_name = entry[start_name:end_name].strip()\\n authors.append(author_name)\\n start_author = entry.find(\\'<author>\\', end_author)\\n end_author = entry.find(\\'</author>\\', start_author)\\n \\n # Print the results\\n print(f\"Title: {title}\")\\n print(f\"Published Date: {published}\")\\n print(f\"Authors: {\\', \\'.join(authors)}\")\\n print(f\"Summary: {summary}\")\\nelse:\\n print(\"Failed to retrieve data from arXiv API.\")\\n```\\n\\nThis script will output the title, published date, authors, and summary of the most recent paper related to GPT-4 on arXiv. Please note that the actual content of the paper and its potential applications in software will need to be analyzed manually after retrieving the paper information.', 'name': 'Engineer', 'role': 'user'}, {'content': \"exitcode: 0 (execution succeeded)\\nCode output: \\nTitle: A Data-Centric Approach To Generate Faithful and High Quality Patient\\n Summaries with Large Language Models\\nPublished Date: 024-02-23T16:32:28Z\\nAuthors: Stefan Hegselmann, Shannon Zejiang Shen, Florian Gierse, Monica Agrawal, David Sontag, Xiaoyi Jiang\\nSummary: Patients often face difficulties in understanding their hospitalizations,\\nwhile healthcare workers have limited resources to provide explanations. In\\nthis work, we investigate the potential of large language models to generate\\npatient summaries based on doctors' notes and study the effect of training data\\non the faithfulness and quality of the generated summaries. To this end, we\\ndevelop a rigorous labeling protocol for hallucinations, and have two medical\\nexperts annotate 100 real-world summaries and 100 generated summaries. We show\\nthat fine-tuning on hallucination-free data effectively reduces hallucinations\\nfrom 2.60 to 1.55 per summary for Llama 2, while preserving relevant\\ninformation. Although the effect is still present, it is much smaller for GPT-4\\nwhen prompted with five examples (0.70 to 0.40). We also conduct a qualitative\\nevaluation using hallucination-free and improved training data. GPT-4 shows\\nvery good results even in the zero-shot setting. We find that common\\nquantitative metrics do not correlate well with faithfulness and quality.\\nFinally, we test GPT-4 for automatic hallucination detection, which yields\\npromising results.\\n\", 'name': 'Executor', 'role': 'user'}, {'content': 'Based on the abstract provided, the paper titled \"A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Models\" explores the use of large language models, including GPT-4, to generate patient summaries from doctors\\' notes. The study focuses on the impact of training data on the faithfulness and quality of the generated summaries and also investigates the potential of GPT-4 for automatic hallucination detection.\\n\\n**Potential Applications in Software:**\\n\\n1. **Healthcare Documentation Automation:**\\n - GPT-4 could be used to develop software that assists healthcare professionals in creating accurate and comprehensive patient summaries by automatically processing doctors\\' notes and other medical records.\\n\\n2. **Clinical Decision Support Systems:**\\n - Integrating GPT-4 into clinical decision support systems could provide healthcare workers with insights and suggestions based on a patient\\'s medical history, potentially improving diagnosis and treatment planning.\\n\\n3. **Patient Education and Communication:**\\n - Software applications could leverage GPT-4 to translate complex medical information into patient-friendly summaries, enhancing patient understanding of their health conditions and treatments.\\n\\n4. **Medical Training and Simulation:**\\n - GPT-4 could be used to create realistic medical scenarios for training medical students and professionals, simulating patient interactions and generating case studies.\\n\\n5. **Data Quality Assurance:**\\n - The paper suggests that GPT-4 can be used for automatic hallucination detection, which refers to the identification of inaccuracies or fabrications in generated text. This could be applied to software that ensures the quality and reliability of medical documentation.\\n\\n6. **Research and Development:**\\n - GPT-4 could assist researchers in summarizing and synthesizing large volumes of medical literature, aiding in the discovery of new insights and the development of novel treatments.\\n\\n7. **Personalized Health Monitoring:**\\n - Software applications could use GPT-4 to provide personalized health monitoring and advice by analyzing user input, such as symptoms or lifestyle factors, and generating tailored health recommendations.\\n\\nThese potential applications highlight the versatility of GPT-4 in the realm of healthcare software, offering opportunities to enhance patient care, improve healthcare workflows, and support medical education and research.', 'name': 'Scientist', 'role': 'user'}, {'content': 'TERMINATE', 'role': 'assistant'}], summary='', cost=({'total_cost': 0}, {'total_cost': 0}), human_input=['Approve', 'TERMINATE'])"
|
|
]
|
|
},
|
|
"execution_count": 8,
|
|
"metadata": {},
|
|
"output_type": "execute_result"
|
|
}
|
|
],
|
|
"source": [
|
|
"user_proxy.initiate_chat(\n",
|
|
" manager, message=\"Find a latest paper about gpt-4 on arxiv and find its potential applications in software.\"\n",
|
|
")\n",
|
|
"# type exit to terminate the chat"
|
|
]
|
|
}
|
|
],
|
|
"metadata": {
|
|
"front_matter": {
|
|
"description": "Introduce Group Chat with Customized Speaker Selection Method",
|
|
"tags": [
|
|
"orchestration",
|
|
"group chat"
|
|
]
|
|
},
|
|
"kernelspec": {
|
|
"display_name": "flaml",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.9.18"
|
|
},
|
|
"orig_nbformat": 4
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 2
|
|
}
|