How to Build a Multi-Agent AI App with AutoGen

1# Check if the database is running on a shared tier2shared_tier_check = %sql show variables like 'is_shared_tier'3
4# If not on a shared tier, or if the shared tier is turned off, drop the existing database and create a new one5if not shared_tier_check or shared_tier_check[0][1] == 'OFF':6    %sql DROP DATABASE IF EXISTS autogen7    %sql CREATE DATABASE autogen

1!pip install --quiet langchain_community pyautogen langchain_openai langchain_text_splitters unstructured

1!pip install --quiet markdown

1import requests2
3r = requests.get("https://raw.githubusercontent.com/microsoft/FLAML/main/website/docs/Examples/Integrate%20-%20Spark.md")4open('example.md', 'wb').write(r.content)

1from langchain_community.vectorstores import SingleStoreDB2from langchain_openai import OpenAIEmbeddings3from langchain_community.document_loaders import UnstructuredMarkdownLoader4from langchain_text_splitters import CharacterTextSplitter5from typing import List, Dict, Union6import os7
8loader = UnstructuredMarkdownLoader("./example.md")9
10os.environ["OPENAI_API_KEY"] = "api-key"11
12data = loader.load()13
14text_splitter = CharacterTextSplitter()15
16docs = text_splitter.split_documents(data)17
18embeddings = OpenAIEmbeddings()19
20os.environ["SINGLESTOREDB_URL"] = "admin:pass@host:3306/db"

1singlestore_db = SingleStoreDB.from_documents(2    docs,3    embeddings,4    table_name="notebook2",  # use table with a custom name5)

1!pip install --quiet pyautogen[retrievechat]

1import autogen2from autogen.agentchat.contrib.retrieve_assistant_agent import RetrieveAssistantAgent3from autogen.agentchat.contrib.retrieve_user_proxy_agent import RetrieveUserProxyAgent4from autogen import config_list_from_json5from autogen import AssistantAgent

1class SingleStoreRetrieveUserProxyAgent(RetrieveUserProxyAgent):2    def __init__(self, singlestore_db: SingleStoreDB, **kwargs):3        super().__init__(**kwargs)4        self.singlestore_db = singlestore_db5
6    def query_vector_db(7        self,8        query_texts: List[str],9        n_results: int = 10,10        search_string: str = "",11        **kwargs,12    ) -> Dict[str, List[List[str]]]:13        documents = []14        ids = []15        for query_index, query_text in enumerate(query_texts):16            searched_docs = self.singlestore_db.similarity_search(17                query=query_text,18                k=n_results,19            )20            # Assuming searched_docs is a list of documents with only 'page_content' property21            batch_documents = [doc.page_content for doc in searched_docs]22            documents.append(batch_documents)23
24            # Generate a unique ID for each document based on enumeration25            batch_ids = [f"{query_index}-{i}" for i in range(len(batch_documents))]26            ids.append(batch_ids)27
28        return {29            "ids": ids,30            "documents": documents,31        }32
33    def retrieve_docs(self, problem: str, n_results: int = 20, search_string: str = "", **kwargs):34        results = self.query_vector_db(35            query_texts=[problem],36            n_results=n_results,37            search_string=search_string,38            **kwargs,39        )40
41        self._results = results

1import os2os.environ["AUTOGEN_USE_DOCKER"] = "False"

1llm_config = {2    "config_list": [{"model": "gpt-3.5-turbo", "api_key": os.environ["OPENAI_API_KEY"]}],3    }4
5def termination_msg(x):6    return isinstance(x, dict) and "TERMINATE" == str(x.get("content", ""))[-9:].upper()7
8
9boss = autogen.UserProxyAgent(10    name="Boss",11    is_termination_msg=termination_msg,12    human_input_mode="NEVER",13    code_execution_config=False,  # we don't want to execute code in this case.14    default_auto_reply="Reply `TERMINATE` if the task is done.",15    description="The boss who ask questions and give tasks.",16)17
18boss_aid = SingleStoreRetrieveUserProxyAgent(19    name="Boss_Assistant",20    is_termination_msg=termination_msg,21    human_input_mode="NEVER",22    max_consecutive_auto_reply=3,23    retrieve_config={24        "task": "code",25    },26    code_execution_config=False,  # we don't want to execute code in this case.27    description="Assistant who has extra content retrieval power for solving difficult problems.",28    singlestore_db=singlestore_db29)30
31coder = autogen.AssistantAgent(32    name="Senior_Python_Engineer",33    is_termination_msg=termination_msg,34    system_message="You are a senior python engineer, you provide python code to answer questions. Reply `TERMINATE` in the end when everything is done.",35    llm_config=llm_config,36    description="Senior Python Engineer who can write code to solve problems and answer questions.",37)38
39pm = autogen.AssistantAgent(40    name="Product_Manager",41    is_termination_msg=termination_msg,42    system_message="You are a product manager. Reply `TERMINATE` in the end when everything is done.",43    llm_config=llm_config,44    description="Product Manager who can design and plan the project.",45)46
47reviewer = autogen.AssistantAgent(48    name="Code_Reviewer",49    is_termination_msg=termination_msg,50    system_message="You are a code reviewer. Reply `TERMINATE` in the end when everything is done.",51    llm_config=llm_config,52    description="Code Reviewer who can review the code.",53)54
55PROBLEM = "How to use spark for parallel training in FLAML? Give me sample code."56
57
58def _reset_agents():59    boss.reset()60    boss_aid.reset()61    coder.reset()62    pm.reset()63    reviewer.reset()64
65
66def rag_chat():67    _reset_agents()68    groupchat = autogen.GroupChat(69        agents=[boss_aid, pm, coder, reviewer], messages=[], max_round=12, speaker_selection_method="round_robin"70    )71    manager = autogen.GroupChatManager(groupchat=groupchat, llm_config=llm_config)72
73    # Start chatting with boss_aid as this is the user proxy agent.74    boss_aid.initiate_chat(75        manager,76        problem=PROBLEM,77        n_results=3,78    )79
80
81def norag_chat():82    _reset_agents()83    groupchat = autogen.GroupChat(84        agents=[boss, pm, coder, reviewer],85        messages=[],86        max_round=12,87        speaker_selection_method="auto",88        allow_repeat_speaker=False,89    )90    manager = autogen.GroupChatManager(groupchat=groupchat, llm_config=llm_config)91
92    # Start chatting with the boss as this is the user proxy agent.93    boss.initiate_chat(94        manager,95        message=PROBLEM,96    )97
98
99def call_rag_chat():100    _reset_agents()101
102    # In this case, we will have multiple user proxy agents and we don't initiate the chat103    # with RAG user proxy agent.104    # In order to use RAG user proxy agent, we need to wrap RAG agents in a function and call105    # it from other agents.106    def retrieve_content(107        message: Annotated[108            str,109            "Refined message which keeps the original meaning and can be used to retrieve content for code generation and question answering.",110        ],111        n_results: Annotated[int, "number of results"] = 3,112    ) -> str:113        boss_aid.n_results = n_results  # Set the number of results to be retrieved.114        # Check if we need to update the context.115        update_context_case1, update_context_case2 = boss_aid._check_update_context(message)116        if (update_context_case1 or update_context_case2) and boss_aid.update_context:117            boss_aid.problem = message if not hasattr(boss_aid, "problem") else boss_aid.problem118            _, ret_msg = boss_aid._generate_retrieve_user_reply(message)119        else:120            ret_msg = boss_aid.generate_init_message(message, n_results=n_results)121        return ret_msg if ret_msg else message122
123    boss_aid.human_input_mode = "NEVER"  # Disable human input for boss_aid since it only retrieves content.124
125    for caller in [pm, coder, reviewer]:126        d_retrieve_content = caller.register_for_llm(127            description="retrieve content for code generation and question answering.", api_style="function"128        )(retrieve_content)129
130    for executor in [boss, pm]:131        executor.register_for_execution()(d_retrieve_content)132
133    groupchat = autogen.GroupChat(134        agents=[boss, pm, coder, reviewer],135        messages=[],136        max_round=12,137        speaker_selection_method="round_robin",138        allow_repeat_speaker=False,139    )140
141    manager = autogen.GroupChatManager(groupchat=groupchat, llm_config=llm_config)142
143    # Start chatting with the boss as this is the user proxy agent.144    boss.initiate_chat(145        manager,146        message=PROBLEM,147    )

1norag_chat()

1rag_chat()

1shared_tier_check = %sql show variables like 'is_shared_tier'2if not shared_tier_check or shared_tier_check[0][1] == 'OFF':3    %sql DROP DATABASE IF EXISTS autogen