MetaGPT

English

中文

English

中文

Appearance

Researcher: Search Web and Write Reports ​

Introduction ​

Background ​

In MetaGPT, the role of a researcher allows users to conduct research by summarizing information gathered from the internet based on user queries. This document will cover the researcher role from aspects such as design philosophy, code implementation, usage examples, etc.

Objective ​

Through this document, you can learn how to use the MetaGPT researcher role to search the internet and summarize reports. Additionally, you can leverage MetaGPT's networking capabilities to develop new intelligent agents.

Source Code ​

Design Overview ​

Design Philosophy ​

Before developing the Researcher role in MetaGPT, it's essential to consider how one would conduct research on the internet. Generally, the process involves the following steps:

Analyze the research question and break it down into several sub-questions suitable for searching with a search engine. Use a search engine to search for sub-questions, review the search results with titles, original URLs, abstracts, etc., and determine the relevance and reliability of each result. Decide whether to further browse the web using the provided URLs. Click on the web pages that need further exploration, assess whether the content is helpful for the research question, extract relevant information, and record it. Aggregate all recorded relevant information and write a report addressing the research question. Therefore, we aim to simulate the above research process using GPT. The overall steps are as follows:

User inputs the research question. Researcher generates a set of research questions using GPT, forming an objective opinion on any given task. Upon receiving the decomposed questions from GPT, the researcher, for each research question, searches through a search engine to obtain initial search results. Retrieve web page content using a browser for the URLs and summarize the web page content. Consolidate all summarized content and track their sources. Finally, instruct GPT to generate the final research report based on the consolidated content. The following is a flowchart illustrating the Researcher role architecture:

Researcher Role Architecture Diagram

Based on this process, we can abstract three Actions and define a Role as follows:

NameClassDescription
CollectLinksActionCollect links from a search engine
WebBrowseAndSummarizeActionExplore the web and provide summaries of articles and webpages
ConductResearchActionConduct research and generate a research report
ResearcherRoleSearch Web and Write Reports

Action Definitions ​

The CollectLinks Action is used to search the internet for relevant questions and retrieve a list of URL addresses. Since user-input questions may not be directly suitable for search engine queries, the CollectLinks Action first breaks down the user's question into multiple sub-questions suitable for search. It then uses a search engine for this purpose. The implementation utilizes the SearchEngine in the tools module, supporting searches through serpapi/google/serper/ddg. The implementation details can be found in metagpt/actions/research.py, and the following provides a basic explanation of the CollectLinks.run method:

python
class CollectLinks(Action):

    async def run(
        self,
        topic: str,
        decomposition_nums: int = 4,
        url_per_query: int = 4,
        system_text: str | None = None,
    ) -> dict[str, list[str]]:
        """Run the action to collect links.

        Args:
            topic: The research topic.
            decomposition_nums: The number of search questions to generate.
            url_per_query: The number of URLs to collect per search question.
            system_text: The system text.

        Returns:
            A dictionary containing the search questions as keys and the collected URLs as values.
        """
        system_text = system_text if system_text else RESEARCH_TOPIC_SYSTEM.format(topic=topic)
        # Decompose the research question into multiple sub-problems
        keywords = await self._aask(SEARCH_TOPIC_PROMPT, [system_text])
        try:
            keywords = OutputParser.extract_struct(keywords, list)
            keywords = parse_obj_as(list[str], keywords)
        except Exception as e:
            logger.exception(f"fail to get keywords related to the research topic \"{topic}\" for {e}")
            keywords = [topic]

        # Search the sub-problems using the search engine
        results = await asyncio.gather(*(self.search_engine.run(i, as_string=False) for i in keywords))

        # Browse through the search engine results and filter out those relevant to the research question
        def gen_msg():
            while True:
                search_results = "\n".join(f"#### Keyword: {i}\n Search Result: {j}\n" for (i, j) in zip(keywords, results))
                prompt = SUMMARIZE_SEARCH_PROMPT.format(decomposition_nums=decomposition_nums, search_results=search_results)
                yield prompt
                remove = max(results, key=len)
                remove.pop()
                if len(remove) == 0:
                    break
        prompt = reduce_message_length(gen_msg(), self.llm.model, system_text, CONFIG.max_tokens_rsp)
        logger.debug(prompt)
        queries = await self._aask(prompt, [system_text])
        try:
            queries = OutputParser.extract_struct(queries, list)
            queries = parse_obj_as(list[str], queries)
        except Exception as e:
            logger.exception(f"fail to break down the research question due to {e}")
            queries = keywords
        ret = {}

        # Sort and take the TopK URLs from the search results
        for query in queries:
            ret[query] = await self._search_and_rank_urls(topic, query, url_per_query)
        return ret
class CollectLinks(Action):

    async def run(
        self,
        topic: str,
        decomposition_nums: int = 4,
        url_per_query: int = 4,
        system_text: str | None = None,
    ) -> dict[str, list[str]]:
        """Run the action to collect links.

        Args:
            topic: The research topic.
            decomposition_nums: The number of search questions to generate.
            url_per_query: The number of URLs to collect per search question.
            system_text: The system text.

        Returns:
            A dictionary containing the search questions as keys and the collected URLs as values.
        """
        system_text = system_text if system_text else RESEARCH_TOPIC_SYSTEM.format(topic=topic)
        # Decompose the research question into multiple sub-problems
        keywords = await self._aask(SEARCH_TOPIC_PROMPT, [system_text])
        try:
            keywords = OutputParser.extract_struct(keywords, list)
            keywords = parse_obj_as(list[str], keywords)
        except Exception as e:
            logger.exception(f"fail to get keywords related to the research topic \"{topic}\" for {e}")
            keywords = [topic]

        # Search the sub-problems using the search engine
        results = await asyncio.gather(*(self.search_engine.run(i, as_string=False) for i in keywords))

        # Browse through the search engine results and filter out those relevant to the research question
        def gen_msg():
            while True:
                search_results = "\n".join(f"#### Keyword: {i}\n Search Result: {j}\n" for (i, j) in zip(keywords, results))
                prompt = SUMMARIZE_SEARCH_PROMPT.format(decomposition_nums=decomposition_nums, search_results=search_results)
                yield prompt
                remove = max(results, key=len)
                remove.pop()
                if len(remove) == 0:
                    break
        prompt = reduce_message_length(gen_msg(), self.llm.model, system_text, CONFIG.max_tokens_rsp)
        logger.debug(prompt)
        queries = await self._aask(prompt, [system_text])
        try:
            queries = OutputParser.extract_struct(queries, list)
            queries = parse_obj_as(list[str], queries)
        except Exception as e:
            logger.exception(f"fail to break down the research question due to {e}")
            queries = keywords
        ret = {}

        # Sort and take the TopK URLs from the search results
        for query in queries:
            ret[query] = await self._search_and_rank_urls(topic, query, url_per_query)
        return ret

WebBrowseAndSummarize ​

The WebBrowseAndSummarize Action is responsible for browsing web pages and summarizing their content. MetaGPT provides the WebBrowserEngine in the tools module, which supports web browsing through playwright/selenium. The WebBrowseAndSummarize Action uses the WebBrowserEngine for web browsing. The implementation details can be found in metagpt/actions/research.py, and the following provides a basic explanation of the WebBrowseAndSummarize.run method:

python
class WebBrowseAndSummarize(Action):
    async def run(
        self,
        url: str,
        *urls: str,
        query: str,
        system_text: str = RESEARCH_BASE_SYSTEM,
    ) -> dict[str, str]:
        """Run the action to browse the web and provide summaries.

        Args:
            url: The main URL to browse.
            urls: Additional URLs to browse.
            query: The research question.
            system_text: The system text.

        Returns:
            A dictionary containing the URLs as keys and their summaries as values.
        """
        # Web page browsing and content extraction
        contents = await self.web_browser_engine.run(url, *urls)
        if not urls:
            contents = [contents]

        # Web page content summarization
        summaries = {}
        prompt_template = WEB_BROWSE_AND_SUMMARIZE_PROMPT.format(query=query, content="{}")
        for u, content in zip([url, *urls], contents):
            content = content.inner_text
            chunk_summaries = []
            for prompt in generate_prompt_chunk(content, prompt_template, self.llm.model, system_text, CONFIG.max_tokens_rsp):
                logger.debug(prompt)
                summary = await self._aask(prompt, [system_text])
                if summary == "Not relevant.":
                    continue
                chunk_summaries.append(summary)

            if not chunk_summaries:
                summaries[u] = None
                continue

            if len(chunk_summaries) == 1:
                summaries[u] = chunk_summaries[0]
                continue

            content = "\n".join(chunk_summaries)
            prompt = WEB_BROWSE_AND_SUMMARIZE_PROMPT.format(query=query, content=content)
            summary = await self._aask(prompt, [system_text])
            summaries[u] = summary
        return summaries
class WebBrowseAndSummarize(Action):
    async def run(
        self,
        url: str,
        *urls: str,
        query: str,
        system_text: str = RESEARCH_BASE_SYSTEM,
    ) -> dict[str, str]:
        """Run the action to browse the web and provide summaries.

        Args:
            url: The main URL to browse.
            urls: Additional URLs to browse.
            query: The research question.
            system_text: The system text.

        Returns:
            A dictionary containing the URLs as keys and their summaries as values.
        """
        # Web page browsing and content extraction
        contents = await self.web_browser_engine.run(url, *urls)
        if not urls:
            contents = [contents]

        # Web page content summarization
        summaries = {}
        prompt_template = WEB_BROWSE_AND_SUMMARIZE_PROMPT.format(query=query, content="{}")
        for u, content in zip([url, *urls], contents):
            content = content.inner_text
            chunk_summaries = []
            for prompt in generate_prompt_chunk(content, prompt_template, self.llm.model, system_text, CONFIG.max_tokens_rsp):
                logger.debug(prompt)
                summary = await self._aask(prompt, [system_text])
                if summary == "Not relevant.":
                    continue
                chunk_summaries.append(summary)

            if not chunk_summaries:
                summaries[u] = None
                continue

            if len(chunk_summaries) == 1:
                summaries[u] = chunk_summaries[0]
                continue

            content = "\n".join(chunk_summaries)
            prompt = WEB_BROWSE_AND_SUMMARIZE_PROMPT.format(query=query, content=content)
            summary = await self._aask(prompt, [system_text])
            summaries[u] = summary
        return summaries

ConductResearch ​

The ConductResearch Action is responsible for writing a research report. It is implemented by using the summarized data from the WebBrowseAndSummarize Action as context and then generating the research report. The implementation details can be found in metagpt/actions/research.py, and the following provides a basic explanation of the ConductResearch.run method:

python
class ConductResearch(Action):
    async def run(
        self,
        topic: str,
        content: str,
        system_text: str = RESEARCH_BASE_SYSTEM,
    ) -> str:
        """Run the action to conduct research and generate a research report.

        Args:
            topic: The research topic.
            content: The content for research.
            system_text: The system text.

        Returns:
            The generated research report.
        """
        prompt = CONDUCT_RESEARCH_PROMPT.format(topic=topic, content=content)
        logger.debug(prompt)
        self.llm.auto_max_tokens = True
        return await self._aask(prompt, [system_text])
class ConductResearch(Action):
    async def run(
        self,
        topic: str,
        content: str,
        system_text: str = RESEARCH_BASE_SYSTEM,
    ) -> str:
        """Run the action to conduct research and generate a research report.

        Args:
            topic: The research topic.
            content: The content for research.
            system_text: The system text.

        Returns:
            The generated research report.
        """
        prompt = CONDUCT_RESEARCH_PROMPT.format(topic=topic, content=content)
        logger.debug(prompt)
        self.llm.auto_max_tokens = True
        return await self._aask(prompt, [system_text])

Role (Researcher) ​

The Researcher role combines the CollectLinks, WebBrowseAndSummarize, and ConductResearch Actions to enable the capability of searching the internet and summarizing reports. Therefore, these three Actions need to be added to the role during initialization using the _init_actions method. Since these Actions are executed in the order of CollectLinks -> WebBrowseAndSummarize -> ConductResearch, the execution logic for these Actions needs to be defined in the react/_act methods. The implementation details can be found in metagpt/roles/researcher.py, and the following provides a basic explanation of the Researcher class:

python

class Researcher(Role):
    def __init__(
        self,
        name: str = "David",
        profile: str = "Researcher",
        goal: str = "Gather information and conduct research",
        constraints: str = "Ensure accuracy and relevance of information",
        language: str = "en-us",
        **kwargs,
    ):
        super().__init__(name, profile, goal, constraints, **kwargs)

        # Add the `CollectLinks`, `WebBrowseAndSummarize`, and `ConductResearch` actions
        self._init_actions([CollectLinks(name), WebBrowseAndSummarize(name), ConductResearch(name)])

        # Set to execute in order
        self._set_react_mode(react_mode="by_order")
        self.language = language
        if language not in ("en-us", "zh-cn"):
            logger.warning(f"The language `{language}` has not been tested, it may not work.")

    async def _act(self) -> Message:
        logger.info(f"{self._setting}: ready to {self._rc.todo}")
        todo = self._rc.todo
        msg = self._rc.memory.get(k=1)[0]
        if isinstance(msg.instruct_content, Report):
            instruct_content = msg.instruct_content
            topic = instruct_content.topic
        else:
            topic = msg.content

        research_system_text = get_research_system_text(topic, self.language)
        # Search the internet and retrieve URL information
        if isinstance(todo, CollectLinks):
            links = await todo.run(topic, 4, 4)
            ret = Message("", Report(topic=topic, links=links), role=self.profile, cause_by=type(todo))
        # Browse web pages and summarize their content
        elif isinstance(todo, WebBrowseAndSummarize):
            links = instruct_content.links
            todos = (todo.run(*url, query=query, system_text=research_system_text) for (query, url) in links.items())
            summaries = await asyncio.gather(*todos)
            summaries = list((url, summary) for i in summaries for (url, summary) in i.items() if summary)
            ret = Message("", Report(topic=topic, summaries=summaries), role=self.profile, cause_by=type(todo))
        # Generate a research report
        else:
            summaries = instruct_content.summaries
            summary_text = "\n---\n".join(f"url: {url}\nsummary: {summary}" for (url, summary) in summaries)
            content = await self._rc.todo.run(topic, summary_text, system_text=research_system_text)
            ret = Message("", Report(topic=topic, content=content), role=self.profile, cause_by=type(self._rc.todo))
        self._rc.memory.add(ret)
        return ret

    async def react(self) -> Message:
        msg = await super().react()
        report = msg.instruct_content
        # Output the report
        self.write_report(report.topic, report.content)
        return msg

class Researcher(Role):
    def __init__(
        self,
        name: str = "David",
        profile: str = "Researcher",
        goal: str = "Gather information and conduct research",
        constraints: str = "Ensure accuracy and relevance of information",
        language: str = "en-us",
        **kwargs,
    ):
        super().__init__(name, profile, goal, constraints, **kwargs)

        # Add the `CollectLinks`, `WebBrowseAndSummarize`, and `ConductResearch` actions
        self._init_actions([CollectLinks(name), WebBrowseAndSummarize(name), ConductResearch(name)])

        # Set to execute in order
        self._set_react_mode(react_mode="by_order")
        self.language = language
        if language not in ("en-us", "zh-cn"):
            logger.warning(f"The language `{language}` has not been tested, it may not work.")

    async def _act(self) -> Message:
        logger.info(f"{self._setting}: ready to {self._rc.todo}")
        todo = self._rc.todo
        msg = self._rc.memory.get(k=1)[0]
        if isinstance(msg.instruct_content, Report):
            instruct_content = msg.instruct_content
            topic = instruct_content.topic
        else:
            topic = msg.content

        research_system_text = get_research_system_text(topic, self.language)
        # Search the internet and retrieve URL information
        if isinstance(todo, CollectLinks):
            links = await todo.run(topic, 4, 4)
            ret = Message("", Report(topic=topic, links=links), role=self.profile, cause_by=type(todo))
        # Browse web pages and summarize their content
        elif isinstance(todo, WebBrowseAndSummarize):
            links = instruct_content.links
            todos = (todo.run(*url, query=query, system_text=research_system_text) for (query, url) in links.items())
            summaries = await asyncio.gather(*todos)
            summaries = list((url, summary) for i in summaries for (url, summary) in i.items() if summary)
            ret = Message("", Report(topic=topic, summaries=summaries), role=self.profile, cause_by=type(todo))
        # Generate a research report
        else:
            summaries = instruct_content.summaries
            summary_text = "\n---\n".join(f"url: {url}\nsummary: {summary}" for (url, summary) in summaries)
            content = await self._rc.todo.run(topic, summary_text, system_text=research_system_text)
            ret = Message("", Report(topic=topic, content=content), role=self.profile, cause_by=type(self._rc.todo))
        self._rc.memory.add(ret)
        return ret

    async def react(self) -> Message:
        msg = await super().react()
        report = msg.instruct_content
        # Output the report
        self.write_report(report.topic, report.content)
        return msg

Usage Instructions ​

Dependencies and Configuration ​

The Researcher role depends on SearchEngine and WebBrowserEngine. Below are brief instructions for installing and configuring these components.

SearchEngine ​

Supports serpapi/google/serper/ddg search engines. They differ as follows:

NameDefault EngineAdditional Dependency PackagesInstallation
serpapi√\\
google×google-api-python-clientpip install metagpt\[search-google\]
serper×\\
ddg×duckduckgo-searchpip install metagpt\[search-ddg\]

Configuration:

WebBrowserEngine ​

Supports playwright/selenium engines. To use them, additional dependencies must be installed. They differ as follows:

NameDefault EngineAdditional Dependency PackagesInstallationAsynchronousSupported Platforms
playwright√playwright beautifulsoup4pip install metagpt\[playwright\]NativePartially supported platforms
selenium×selenium webdriver_manager beautifulsoup4pip install metagpt\[selenium\]Thread PoolAlmost all platforms

Configuration:

  • playwright
    • WEB_BROWSER_ENGINE: Set to playwright
    • PLAYWRIGHT_BROWSER_TYPE: Supports chromium/firefox/webkit; defaults to chromium. More information: Playwright BrowserType
  • selenium
    • WEB_BROWSER_ENGINE: Set to selenium
    • SELENIUM_BROWSER_TYPE: Supports chrome/firefox/edge/ie; defaults to chrome. More information: Selenium BrowserTypes

Additionally, the Researcher provides the following configurations:

  • MODEL_FOR_RESEARCHER_SUMMARY: GPT model for content summarization; defaults to gpt-3.5-turbo
  • MODEL_FOR_RESEARCHER_REPORT: GPT model for report generation; defaults to gpt-3.5-turbo-16k

Running Examples and Results ​

The metagpt.roles.researcher module provides a command-line interface for executing the functionalities of the Researcher. An example is as follows:

bash
python -m metagpt.roles.researcher "dataiku vs. datarobot"
python -m metagpt.roles.researcher "dataiku vs. datarobot"

Log output: log.txt Report output: dataiku vs. datarobot.md

Released under the MIT License.

Copyright © 2023-present alexanderwu