from crewai_tools import ParallelSearchTool

tool = ParallelSearchTool()
resp_json = tool.run(
  objective="When was the United Nations established? Prefer UN's websites.",
  search_queries=["Founding year UN", "Year of founding United Nations"],
  processor="base",
  max_results=5,
  max_chars_per_result=1500,
)
print(resp_json)  # => {"search_id": ..., "results": [{"url", "title", "excerpts": [...]}, ...]}

tool.run(
  objective: str | None = None,                 # ≤ 5000 chars
  search_queries: list[str] | None = None,      # up to 5 items, each ≤ 200 chars
  processor: str = "base",                      # "base" (fast) or "pro" (freshness/quality)
  max_results: int = 10,                        # ≤ 40 (processor limits apply)
  max_chars_per_result: int = 6000,             # ≥ 100 (values > 30000 not guaranteed)
  source_policy: dict | None = None,            # optional SourcePolicy config
)

source_policy = {
  "allow": {"domains": ["un.org"]},
  # "deny": {"domains": ["example.com"]},  # optional
}

resp_json = tool.run(
  objective="When was the United Nations established?",
  processor="base",
  max_results=5,
  max_chars_per_result=1500,
  source_policy=source_policy,
)

import os
from crewai import Agent, Task, Crew, LLM, Process
from crewai_tools import ParallelSearchTool

# LLM
llm = LLM(
  model="gemini/gemini-2.0-flash",
  temperature=0.5,
  api_key=os.getenv("GEMINI_API_KEY")
)

# Parallel Search
search = ParallelSearchTool()

# User query
query = "find all the recent concerns about AI evals? please cite the sources"

# Researcher agent 
researcher = Agent(
  role="Web Researcher",
  backstory="You are an expert web researcher",
  goal="Find cited, high-quality sources and provide a brief answer.",
  tools=[search],
  llm=llm,
  verbose=True,
)

# Research task
task = Task(
  description=f"Research the {query} and produce a short, cited answer.",
  expected_output="A concise, sourced answer to the question. The answer should be in this format: [query]: [answer] - [source]",
  agent=researcher,
  output_file="answer.mdx",
)

# Crew
crew = Crew(
    agents=[researcher], 
    tasks=[task], 
    verbose=True,
    process=Process.sequential,
)

# Run the crew
result = crew.kickoff(inputs={'query': query})
print(result)

Recent concerns about AI evaluations include: the rise of AI-related incidents alongside a lack of standardized Responsible AI (RAI) evaluations among major industrial model developers - [https://hai.stanford.edu/ai-index/2025-ai-index-report]; flawed benchmark datasets that fail to account for critical factors, leading to unrealistic estimates of AI model abilities - [https://www.nature.com/articles/d41586-025-02462-5]; the need for multi-metric, context-aware evaluations in medical imaging AI to ensure reliability and clinical relevance - [https://www.sciencedirect.com/science/article/pii/S3050577125000283]; challenges related to data sets (insufficient, imbalanced, or poor quality), communication gaps, and misaligned expectations in AI model training - [https://www.oracle.com/artificial-intelligence/ai-model-training-challenges/]; the argument that LLM agents should be evaluated primarily on their riskiness, not just performance, due to unreliability, hallucinations, and brittleness - [https://www.technologyreview.com/2025/06/24/1119187/fix-ai-evaluation-crisis/]; the fact that the AI industry's embraced benchmarks may be close to meaningless, with top makers of AI models picking and choosing different responsible AI benchmarks, complicating efforts to systematically compare risks and limitations - [https://themarkup.org/artificial-intelligence/2024/07/17/everyone-is-judging-ai-by-these-tests-but-experts-say-theyre-close-to-meaningless]; and the difficulty of building robust and reliable model evaluations, as many existing evaluation suites are limited in their ability to serve as accurate indicators of model capabilities or safety - [https://www.anthropic.com/research/evaluating-ai-systems].