bridge_service.py

"""
Core bridge logic.
- Level inference (rule-based + free-text regex)
- Session context accumulation
- Response composition
"""

from __future__ import annotations
import os
import re
from bridge_models import (
    BridgeRequest,
    BridgeResponse,
    BridgeErrorResponse,
    DashboardContext,
)
from session_store import AccumulatedContext, HistoryEntry, SessionState, store
import mock_responses

MOCK_MODE = os.environ.get("MOCK_MODE", "true").lower() == "true"

# Regex patterns for free-text metric extraction (case-insensitive)
_METRIC_PATTERNS = {
    "cvr": re.compile(r"cvr\s*[:=は]?\s*(\d+\.?\d*)\s*%?", re.IGNORECASE),
    "ctr": re.compile(r"ctr\s*[:=は]?\s*(\d+\.?\d*)\s*%?", re.IGNORECASE),
    "cpa": re.compile(r"cpa\s*[:=は]?\s*(\d+\.?\d*)\s*円?", re.IGNORECASE),
}


def _infer_level(ctx: AccumulatedContext) -> tuple[str, str, str]:
    """
    Returns (level, confidence, reason).
    Level only goes upward -- caller ensures accumulated context is monotonic.
    """
    if ctx.image_base64:
        return "level3", "high", "image_base64 present"
    if ctx.cvr is not None or ctx.ctr is not None or ctx.cpa is not None:
        return "level2", "high", "at least one metric (cvr/ctr/cpa) present"
    if ctx.campaign_name or ctx.industry:
        return "level1", "high", "campaign_name or industry present"
    return "level1", "low", "no context available, defaulting to level1"


def _extract_metrics_from_text(text: str) -> dict[str, float]:
    """Extract metric values from free-text user message."""
    extracted: dict[str, float] = {}
    for key, pattern in _METRIC_PATTERNS.items():
        m = pattern.search(text)
        if m:
            try:
                extracted[key] = float(m.group(1))
            except ValueError:
                pass
    return extracted


def _context_from_dashboard(dc: DashboardContext | None) -> AccumulatedContext:
    if dc is None:
        return AccumulatedContext()
    metrics = dc.metrics or {}
    if hasattr(metrics, "cvr"):
        cvr, ctr, cpa = metrics.cvr, metrics.ctr, metrics.cpa
    else:
        cvr = ctr = cpa = None
    return AccumulatedContext(
        campaign_name=dc.campaign_name,
        industry=dc.industry,
        cvr=cvr,
        ctr=ctr,
        cpa=cpa,
        image_base64=dc.image_base64,
    )


def _call_level_api(level: str, ctx: AccumulatedContext, history: list[HistoryEntry]):
    if MOCK_MODE:
        return mock_responses.get_mock_best_now(level, ctx)
    if level == "level1":
        return mock_responses.level1_propose(ctx, history)
    if level == "level2":
        return mock_responses.level2_propose(ctx, history)
    return mock_responses.level3_propose(ctx, history)


def process_request(req: BridgeRequest) -> BridgeResponse | BridgeErrorResponse:
    """Main entry point. Stateless from caller's perspective."""

    # --- Session resolution ---
    if req.session_id:
        state = store.get(req.session_id)
        if state is None:
            return BridgeErrorResponse(
                session_id=req.session_id,
                error_code="INVALID_SESSION",
                message="セッションが見つからないか期限切れです。ページを再読み込みしてください。",
            )
    else:
        try:
            state = store.create()
        except RuntimeError:
            return BridgeErrorResponse(
                error_code="SERVER_ERROR",
                message="サーバーが混雑しています。しばらくしてから再度お試しください。",
            )

    # --- Merge dashboard_context (Turn 1 or later) ---
    new_ctx = _context_from_dashboard(req.dashboard_context)
    state.accumulated_context.merge(new_ctx)

    # --- Extract metrics from free-text message ---
    if req.message.strip():
        extracted = _extract_metrics_from_text(req.message)
        if "cvr" in extracted:
            state.accumulated_context.cvr = extracted["cvr"]
        if "ctr" in extracted:
            state.accumulated_context.ctr = extracted["ctr"]
        if "cpa" in extracted:
            state.accumulated_context.cpa = extracted["cpa"]

        state.history.append(HistoryEntry(role="user", content=req.message))

    # --- Level inference ---
    level, confidence, reason = _infer_level(state.accumulated_context)
    # Level only goes upward
    level_order = {"level1": 1, "level2": 2, "level3": 3}
    if level_order.get(level, 0) > level_order.get(state.current_level, 0):
        state.current_level = level
    else:
        level = state.current_level

    state.turn_count += 1

    # Guard: max turns
    if state.turn_count > 50:
        return BridgeErrorResponse(
            session_id=state.session_id,
            error_code="MAX_TURNS_EXCEEDED",
            message="会話の最大ターン数（50）に達しました。新しいセッションを開始してください。",
        )

    # --- Call level API ---
    try:
        best_now = _call_level_api(level, state.accumulated_context, state.history)
    except Exception as e:
        fallback_preview = mock_responses.get_mock_next_level_preview(level)
        return BridgeErrorResponse(
            session_id=state.session_id,
            error_code="DOWNSTREAM_TIMEOUT",
            message=f"提案APIの呼び出しに失敗しました: {e}",
            fallback={"next_level_preview": fallback_preview.model_dump()},
        )

    next_preview = mock_responses.get_mock_next_level_preview(level)
    follow_up = mock_responses.get_follow_up_question(level)

    # --- Record assistant response in history ---
    state.history.append(HistoryEntry(role="assistant", content=best_now.summary))

    # --- Persist session ---
    store.save(state)

    return BridgeResponse(
        session_id=state.session_id,
        turn_number=state.turn_count,
        inferred_level=level,
        level_confidence=confidence,
        level_reason=reason,
        best_now=best_now,
        next_level_preview=next_preview,
        follow_up_question=follow_up,
    )