I am feeling frustrated because I’ve spent over three weeks trying to implement a simple solution. While I’ve managed to achieve the desired output twice, the results are inconsistent—when I run the crew again, the output doesn’t meet the requirements. I am seriously considering abandoning CrewAI and switching to a different framework, I need to finish this project.
Here’s what my crew is supposed to do:
- Create 3 web search queries related some keywords. Status: Working correctly
- Search for each query received 15 articles using SerperDevTool. Status: Working correctly
- Select the 25 most relevant and impactful articles for the user. Issue: It’s only returning 4 articles, and the last one is incomplete.
- Summarize the selected articles based on specific criteria using OpenAI. Status: Working correctly
- Return 10 correctly summarized articles. Issue: It’s only returning 1 summarized article.
I have tried executing the crew activating and deactivating the memory, using managers and not.
crewai version: 0.80.0
This is my code:
- crew.yaml:
@CrewBase
class ArticleResearchCrew:
“”“NexagoAgentsProject crew”“”agents_config = "config/agents.yaml" tasks_config = "config/tasks.yaml" def prepare_data(self, data): start_date = (datetime.today() - timedelta(days=7)).strftime('%Y-%m-%d') end_date = datetime.today().strftime('%Y-%m-%d') keywords = " and ".join(data.get("keywords", [])) return { "keywords": keywords, "startDate": start_date, "endDate": end_date, "numQueries": data.get("numQueries"), "userQuery": data.get("userQuery") } def kickoff(self, data): try: prepared_data = self.prepare_data(data) crew_instance = self.crew() result = crew_instance.kickoff(inputs=prepared_data) logger.debug("Kickoff result: %s", result) return result except Exception as e: logger.error(f"Error in kickoff: {e}") return {"error": str(e), "status": "kickoff_failed"} @agent def query_generator(self) -> Agent: return Agent( config=self.agents_config['query_generator'], llm=llm, verbose = False, allow_delegation=False ) @agent def article_researcher(self) -> Agent: serper_tool = SerperDevTool(n_results=15, tbs="qdr:w", url="https://google.serper.dev/news") return Agent( config=self.agents_config['article_researcher'], tools=[serper_tool], verbose = True, llm=llm ) @agent def article_summarizer(self) -> Agent: article_validation = ArticleValidatorExtractorTool() summarizer_tool = ArticleSummarizerTool() return Agent( config=self.agents_config['article_summarizer'], tools=[article_validation, summarizer_tool], verbose = True, llm=llm ) @task def query_generator_task(self) -> Task: return Task( config=self.tasks_config['query_generator_task'], expected_output=self.tasks_config['query_generator_task'].get('expected_output'), agent=self.query_generator() ) @task def article_research_task(self) -> Task: return Task( config=self.tasks_config['article_research_task'], expected_output=self.tasks_config['article_research_task'].get('expected_output'), agent=self.article_researcher() ) @task def article_summarizer_task(self) -> Task: return Task( config=self.tasks_config['article_summarizer_task'], expected_output=self.tasks_config['article_summarizer_task'].get('expected_output'), agent=self.article_summarizer() ) @crew def crew(self) -> Crew: """Creates the NexagoAgentsProject crew""" logger.debug("Initializing agents and tasks in crew") return Crew( agents=self.agents, # Automatically created by the @agent decorator tasks=self.tasks, # Automatically created by the @task decorator #manager_llm=llm, process=Process.sequential, #process=Process.hierarchical, verbose=True, memory=True )
- agents.yaml:
query_generator:
role: >
Expert Web Queries Generator
goal: >
Generate {numQueries} unique, targeted search queries related to {keywords}.
These queries should be designed to retrieve high-quality articles or news published recently.
Prioritize queries that are likely to yield the most relevant and engaging content for users searching on {keywords}.
backstory: >
As a seasoned researcher with deep expertise in {keywords}, you stay informed about the most recent projects, developments, innovations, and potential impacts.
You use your knowledge to develop queries that expose cutting-edge research and new developments in {keywords}.article_researcher:
role: >
Senior Articles Researcher specializing in {keywords}
goal: >
Do thorough research using SerperDevTool to find all recent articles about queries provided to you.
Your goal is to create a list of the 20 most relevant and up-to-date articles about {keywords}.
backstory: >
You are a seasoned {keywords} researcher with a knack for putting together the most relevant and recent articles about queries provided.
Your expertise lies to turn raw research into clear and comprehensive articles list.
You are very strict with the final output, you meet all requirements.article_summarizer:
role: >
Senior Articles Summarizer specializing in {keywords}
goal: >
You will receive a list of articles, you will:
1. Validate each link using the ArticleValidatorExtractorToolValidate to ensure it leads to a full-length article and, if so, extract its content.
2. Pass each content extracted as a block to ArticleSummarizerTool to create comprehensive, detailed, and cohesive summary for each article.
3. Create a final list of articles about {keywords}.
backstory: >
You are a seasoned expert in evaluating URLs to confirm they lead to high-quality, full-length articles.
Your expertise lies in crafting well-structured, detailed summaries that provide valuable insights into {keywords} for industry professionals and researchers.
You are very strict with the final output, you meet all requirements.
- tasks.yaml:
query_generator_task:
description: >
Generate {numQueries} distinct, high-quality web search queries and identify the most promising full articles published recently about {keywords}.
Follow these guidelines:
- Query variety:
- For the first query, use phrases that directly relate to “{userQuery}”, if it is not null. If query user doesn’t define a site for search, don’t include it. If userQuery is null, don’t define a query pass to second query.
- The second query must be: “Latest news, trends, new developments, latest releases and new studies in {keywords}”. Don’t define in this query a “site”
- Subsequent queries must differ significantly from each other, each addressing distinct aspects of the topic, such as social or business impacts, trends, new studies, or applications.
Ensure diverse perspectives are explored by using specific and relevant keywords.
- Article focus:
- Use “full-length articles” to avoid listing pages, blog posts, or video content.
- Use always “filetype:html” to ensure article pages only.
- Emphasize up-to-date and relevant information specifiying the timeframe: “after:{startDate} before:{endDate}”.
- Clarity and relevance:
- Ensure each query targets the retrieval of insightful, engaging, and high-quality content that aligns with the specified {keywords}.
- Avoid ambiguous terms and focus on precision to enhance search accuracy.
- Web search format:
- Queries must have a structured web search format, varying in structure to ensure diverse results.
expected_output: >
A single list of {numQueries} well-structured web search queries with the following structure:
- queryText (str format)
agent: query_generatorarticle_research_task:
description: >
Conduct thorough research using each query received and SerperDevTool. Ensure the following:
- Use all {numQueries} queries for articles searching.
- The URLs should link to full-length articles, avoiding listing pages.
- Select the 25 most relevant articles found, taking in consideration:
- Each article found is directly related to the search queries and provides insightful, relevant, and engaging content published between {startDate} and {endDate}.
- Relevance score criteria:
- society impact level: up to 3 points
- business impact level: society impact level: up to 3 points
- audience engagement level: society impact level: up to 4 points
- Content:
- 40% of articles directly address {userQuery}.
- 40% focus on the latest news, developments, or research.
- 20% cover other relevant or related topics.
expected_output: >
- A structured single list containing the 25 most relevant articles, each one with following data:
- title
- link (raw URL)
- snippet
- date (date of publication, format: YYYY-MM-DD)
- source
- query (query used to find this article)
- All information should be factual, and no content should be invented.
- MANDATORY:
- Return 25 articles, don’t return less.
- Don’t return a list with articles without complete data.
agent: article_researcherarticle_summarizer_task:
description: >
- For each article of the provided list, perform the following steps in order:
- Use the ArticleValidatorExtractorTool to validate its link and then, extract its content.
If the link does not lead to a valid article, exclude it from extraction and other further processing.
- Use the ArticleSummarizerTool to generate a summary of the content extracted.
ArticleSummarizerTool must receive the content extracted complete, don’t truncate it.
If you cannot summarize a content article, exclude it from further processing.
- Create a final list of 10 articles that meet these criteria:
- Information is accurate and directly sourced from the original articles.
- Articles are relevant and engaging for {keywords} experts.
- Articles that fail any step are excluded from the output.
- Ensure that the full text summary generated by ArticleSummarizerTool is included without modification or truncation, mantain paragraph separation.
- MANDATORY:
- Process all articles of the list.
- If some article retrieves error, pass to following article, until have finished all the list.
expected_output: >
- A single structured list of the 10 most relevance articles found in JSON format.
- Each article must contain:
- title
- link (only raw URL)
- date (date of publication, format: YYYY-MM-DD)
- source
- summarize (Ensure full text generated by ArticleSummarizerTool is included without truncation, mantain also paragraph separation)
- MANDATORY: return 10 articles, don’t return less. If you cannot do it, explain why.
agent: article_summarizer
I will appreciate any help.
