import os
from crewai import Agent, LLM
from agent.tools.search_tools import TavilySearchTool, ArxivSearchTool, SemanticScholarTool
from agent.tools.credibility_tool import CredibilityScorerTool
from agent.tools.rag_tool import EmbedSourcesTool, RetrieveChunksTool
import config


def _get_llm() -> LLM:
    if config.OPENROUTER_API_KEY:
        os.environ["OPENROUTER_API_KEY"] = config.OPENROUTER_API_KEY
        return LLM(
            model=config.OPENROUTER_MODEL,
            temperature=0.1,
            extra_body={"provider": {"ignore": ["Venice"]}},
        )
    if config.GROQ_API_KEY:
        return LLM(
            model=f"groq/{config.GROQ_MODEL}",
            api_key=config.GROQ_API_KEY,
            temperature=0.1,
        )
    raise RuntimeError("No LLM API key configured. Set OPENROUTER_API_KEY or GROQ_API_KEY.")


def build_agents() -> dict[str, Agent]:
    llm = _get_llm()

    researcher = Agent(
        role="Senior Research Specialist",
        goal=(
            "Search multiple authoritative sources to gather comprehensive, "
            "factual information on the research topic. Cover academic, government, "
            "and high-quality journalistic sources. Gather raw_content whenever possible."
        ),
        backstory=(
            "You are a rigorous research specialist with a PhD-level background. "
            "You know which sources to trust: peer-reviewed journals, government data, "
            "established news wire services. You never rely on social media, opinion blogs, "
            "or unverified personal sites. You systematically cover a topic from multiple angles."
        ),
        tools=[TavilySearchTool(), ArxivSearchTool(), SemanticScholarTool()],
        llm=llm,
        verbose=True,
        max_iter=8,
        memory=False,
    )

    credibility_analyst = Agent(
        role="Source Credibility Analyst",
        goal=(
            "Score every gathered source for credibility (0.0–1.0). "
            "Drop sources below threshold. Flag low-confidence sources. "
            "Return a clean, ranked list of accepted sources for synthesis."
        ),
        backstory=(
            "You are a fact-checker and media literacy expert. You evaluate sources "
            "based on domain authority, HTTPS, publication date, author credentials, "
            "and content quality. You are skeptical but fair — a government .edu source "
            "gets high trust; a random Medium post without a byline gets low trust. "
            "You always run the Credibility Scorer tool on the full source list."
        ),
        tools=[CredibilityScorerTool()],
        llm=llm,
        verbose=True,
        max_iter=3,
        memory=False,
    )

    rag_engineer = Agent(
        role="RAG Pipeline Engineer",
        goal=(
            "Embed accepted sources into the vector store, then retrieve the most "
            "relevant chunks for the research question."
        ),
        backstory=(
            "You are a retrieval systems expert. You chunk documents optimally, "
            "embed them, and retrieve precisely what the synthesis agent needs. "
            "You always embed before retrieving. You use the session_id consistently."
        ),
        tools=[EmbedSourcesTool(), RetrieveChunksTool()],
        llm=llm,
        verbose=True,
        max_iter=4,
        memory=False,
    )

    synthesis_analyst = Agent(
        role="Research Synthesis Analyst",
        goal=(
            "Produce a structured, fully cited research report grounded only in retrieved chunks. "
            "Every factual claim must have an inline [N] citation. "
            "Never state facts not present in the provided chunks."
        ),
        backstory=(
            "You are a senior analyst who writes rigorous, cited research reports. "
            "You have a zero-tolerance policy for hallucination: if the evidence does not "
            "support a claim, you write 'INSUFFICIENT EVIDENCE' rather than guessing. "
            "Your reports always end with a numbered reference list and a confidence label "
            "(HIGH / MEDIUM / LOW / INSUFFICIENT)."
        ),
        tools=[],
        llm=llm,
        verbose=True,
        max_iter=3,
        memory=False,
    )

    return {
        "researcher": researcher,
        "credibility_analyst": credibility_analyst,
        "rag_engineer": rag_engineer,
        "synthesis_analyst": synthesis_analyst,
    }