# crew_deep_research.py — Deep Research (CrewAI) # Safety: local/OpenAI-compatible endpoint preferred. No keys in source. # pip install crewai (pin versions yourself) from crewai import Agent, Task, Crew, Process SYSTEM = """You are a careful research desk. Goal: produce a brief, not a verdict. RULES: - Use only facts present in the user paste or clearly marked as [UNVERIFIED]. - Separate: Known / Inferred / Unknown / Next checks. - Label every finding Known vs Inferred. Never state Inferred as fact. - Do not invent causes (trust, reputation, criteria changes, market effects) not in the paste. - Do not invent metrics, percentages, or dates. - If recommending a priority, phrase as hypothesis under Open questions / Next checks — not as decided scope. - Prefer quoting pasted labels/counts exactly. - Cite sources as the user labeled them (URL, doc name, date). Do not invent citations. - No spending, no shell commands, no credential requests. - Stop after the brief. Ask one clarifying question only if blocking. OUTPUT: 1) One-sentence scope 2) Findings (bullets) 3) Open questions 4) Next checks (max 5) 5) Confidence: low|medium|high + why""" GOAL = "Turn a fuzzy question into a sourced research brief with open questions and next checks \u2014 not a final verdict." lead = Agent( role="Deep Research lead", goal=GOAL, backstory="Local BotShelf operator. One job, then stop. No spend. No destructive tools.", allow_delegation=False, verbose=False, ) reviewer = Agent( role="Quality gate", goal="Check the lead output against the required sections; list gaps only.", backstory="Pedantic reviewer. Does not invent facts.", allow_delegation=False, verbose=False, ) produce = Task( description=( "Follow SYSTEM strictly.\n\nSYSTEM:\n" + SYSTEM + "\n\nUSER INPUT:\n{user_input}" ), expected_output="Complete Deep Research structured result.", agent=lead, ) gate = Task( description=( "Verify the previous result has all required sections. " "Return Pass/Fail + missing sections. Do not rewrite creatively." ), expected_output="Pass/Fail and gap list.", agent=reviewer, ) crew = Crew( agents=[lead, reviewer], tasks=[produce, gate], process=Process.sequential, ) def run(user_input: str) -> str: return str(crew.kickoff(inputs={"user_input": user_input})) if __name__ == "__main__": print(run("PASTE_EXAMPLE_IN"))