"""Financial analyst agent.

Interprets retrieved evidence — trends, liquidity, leverage, profitability,
risk — and can call tools (ratio engine, calculator, filings search,
currency conversion) via a bounded tool-calling loop.
Default model: GPT (strong numerical reasoning + tool use).
"""

from __future__ import annotations

from langchain_core.messages import HumanMessage, SystemMessage, ToolMessage

from src.agents.qa_agent import format_evidence
from src.llm import get_llm
from src.retrieval.hybrid import RetrievedChunk
from src.tools.langchain_tools import ANALYST_TOOLS

SYSTEM = """You are a senior financial analyst. Interpret the evidence excerpts to answer
the question: explain trends, liquidity, leverage, profitability, and risks.

Rules:
- Ground every quantitative claim in the evidence and cite it, e.g. [balance_sheet_2024 p.12].
- Use the calculate_ratio / calculator tools for ALL arithmetic — never compute in your head.
- Make your reasoning explicit as a short chain: observation -> ratio/figure -> implication.
- Distinguish clearly between what the documents state and your inference.
- If evidence is insufficient for a firm conclusion, say what additional data you would need."""

MAX_TOOL_ROUNDS = 6


def answer(question: str, retrieved: list[RetrievedChunk],
           history: str = "") -> tuple[str, list[dict]]:
    """Returns (answer_text, tool_trace). tool_trace records each tool call
    for the explainability panel in the UI."""
    llm = get_llm("analyst").bind_tools(ANALYST_TOOLS)
    tools_by_name = {t.name: t for t in ANALYST_TOOLS}

    context = format_evidence(retrieved) if retrieved else "(no evidence retrieved)"
    prompt = ""
    if history:
        prompt += f"Conversation so far:\n{history}\n\n"
    prompt += f"Evidence excerpts:\n\n{context}\n\nQuestion: {question}"

    messages = [SystemMessage(content=SYSTEM), HumanMessage(content=prompt)]
    tool_trace: list[dict] = []

    for _ in range(MAX_TOOL_ROUNDS):
        resp = llm.invoke(messages)
        messages.append(resp)
        if not getattr(resp, "tool_calls", None):
            return resp.content, tool_trace
        for call in resp.tool_calls:
            tool = tools_by_name.get(call["name"])
            try:
                result = tool.invoke(call["args"]) if tool else f"Unknown tool {call['name']}"
            except Exception as e:
                result = f"Tool error: {e}"
            tool_trace.append({"tool": call["name"], "args": call["args"], "result": str(result)})
            messages.append(ToolMessage(content=str(result), tool_call_id=call["id"]))

    # ran out of tool rounds — ask for a final synthesis without tools
    final = get_llm("analyst").invoke(
        messages + [HumanMessage(content="Provide your final answer now without further tool calls.")]
    )
    return final.content, tool_trace