life-echo/api/app/features/evaluation/judge_service.py

"""智谱 GLM-5 评审调用（结构化 JSON）。"""

from __future__ import annotations

from collections.abc import AsyncIterator
from dataclasses import dataclass
from typing import Any, Generic, TypeVar

from app.core.config import settings
from app.core.llm_call import LLMCallError, allm_json_call
from app.core.logging import get_logger
from app.features.evaluation.judge_schemas import (
    ConversationJudgeOutput,
    MemoirJudgeOutput,
    TurnJudgeOutput,
)
from app.features.evaluation.rubrics.conversation_v1 import (
    COMPARE_CONV_STREAM_HINT,
    CONV_JUDGE_INSTRUCTIONS,
    TURN_JUDGE_INSTRUCTIONS,
)
from app.features.evaluation.rubrics.memoir_v1 import MEMOIR_JUDGE_INSTRUCTIONS

logger = get_logger(__name__)

TJudgeOutput = TypeVar(
    "TJudgeOutput", TurnJudgeOutput, ConversationJudgeOutput, MemoirJudgeOutput
)

_TURN_MAX = 768
_CONV_JUDGE_JSON_MAX = 2048
_CONV_HEADER = "【完整对话】（每轮以 `[Turn k]` 开头）\n\n"
_MEMOIR_MAX = 12000
_MEMOIR_JSON_MAX = 1536
_COMPARE_STREAM_MAX = 6144
_MEMOIR_EVIDENCE_MAX = 12000


def _eval_judge_prompt_char_pool() -> int:
    """整段请求的字符预算（由评审模型 context window 推导，保守）。"""
    toks = (
        settings.eval_judge_context_window_tokens
        - settings.eval_judge_completion_reserve_tokens
        - settings.eval_judge_prompt_budget_safety_tokens
    )
    toks = max(1, toks)
    return max(1, int(toks / settings.eval_judge_approx_tokens_per_char))


def eval_judge_conversation_transcript_max_chars() -> int:
    """整段对话评审：【完整对话】transcript 最大字符数。"""
    if settings.eval_judge_max_transcript_chars > 0:
        return settings.eval_judge_max_transcript_chars
    overhead = len(CONV_JUDGE_INSTRUCTIONS) + len(_CONV_HEADER) + 32
    return max(1, _eval_judge_prompt_char_pool() - overhead)


def eval_judge_turn_prior_transcript_max_chars() -> int:
    """逐轮评审：截至上一轮的 transcript 节选上限（含与用户/助手正文头的固定开销）。"""
    if settings.eval_judge_max_transcript_chars > 0:
        return settings.eval_judge_max_transcript_chars
    static = len(TURN_JUDGE_INSTRUCTIONS) + 8800
    return max(1, _eval_judge_prompt_char_pool() - static)


def eval_judge_compare_transcript_each_max_chars() -> int:
    """A/B 两段 transcript 同 prompt 时，每条 transcript 的上限（均分字符预算）。"""
    if settings.eval_judge_max_compare_transcript_chars_each > 0:
        return settings.eval_judge_max_compare_transcript_chars_each
    pool = _eval_judge_prompt_char_pool() - settings.eval_judge_compare_prompt_overhead_chars
    return max(1, pool // 2)


@dataclass(slots=True)
class JudgeCallResult(Generic[TJudgeOutput]):
    output: TJudgeOutput | None
    error: str | None = None


def _judge_error_message(e: LLMCallError) -> str:
    prefix = {
        "invoke": "模型调用失败",
        "decode": "JSON 解析失败",
        "validation": "结果校验失败",
    }.get(e.kind, "评审失败")
    detail = str(e).strip()
    return f"{prefix}: {detail}" if detail else prefix


def _build_memoir_judge_prompt(
    *,
    memoir_markdown: str,
    source_transcript: str = "",
    reference_memoir_markdown: str = "",
    evidence_notes: str = "",
) -> str:
    """Assemble an evidence-aware memoir judging prompt."""
    source = (source_transcript or "").strip()
    reference = (reference_memoir_markdown or "").strip()
    notes = (evidence_notes or "").strip()
    sections = [
        MEMOIR_JUDGE_INSTRUCTIONS,
        "",
        "【证据与输入顺序】以下区块按优先级给出：评审说明（若有）→ 原始访谈证据 → 参考基线（若有）→ 待评成稿。**真实性相关细项必须以原始访谈证据为准。**",
        "",
    ]
    if notes:
        sections.extend(["【评审说明】", notes[:1200], ""])
    if source:
        sections.extend(["【原始访谈/证据】", source[:_MEMOIR_EVIDENCE_MAX], ""])
    else:
        sections.extend(
            [
                "【原始访谈/证据】",
                "无可用原始访谈证据。对于记忆忠实度、事实准确性、事实覆盖率、记忆可追溯性，必须保守打分，不得凭空高分。",
                "",
            ]
        )
    if reference:
        sections.extend(["【参考基线/导出成稿】", reference[:_MEMOIR_EVIDENCE_MAX], ""])
    sections.extend(["【当前回忆录正文】", memoir_markdown[:_MEMOIR_MAX]])
    return "\n".join(sections)


class EvalJudgeService:
    def __init__(self, judge_llm: Any | None) -> None:
        self._llm = judge_llm

    async def judge_turn(
        self,
        *,
        prior_transcript: str,
        user_utterance: str,
        assistant_reply: str,
        turn_index_0: int = 0,
    ) -> TurnJudgeOutput | None:
        if not self._llm:
            return None
        t = max(0, int(turn_index_0))
        prompt = f"""{TURN_JUDGE_INSTRUCTIONS}

【本轮位置】完整对话中当前轮次为 Turn {t + 1}（与下方节选及全量 transcript 的 `[Turn ...]` 编号一致）。evidence_refs.turn_index 请使用该编号。

【截至上一轮的对话节选】（含 `[Turn k]` 标签）
{prior_transcript[: eval_judge_turn_prior_transcript_max_chars()]}

【本轮用户】
{user_utterance[:4000]}

【本轮 AI】
{assistant_reply[:4000]}
"""
        try:
            return await allm_json_call(
                self._llm,
                prompt,
                TurnJudgeOutput,
                max_tokens=_TURN_MAX,
                agent="EvalJudgeService.judge_turn",
            )
        except LLMCallError as e:
            logger.warning("turn judge failed: {}", e)
            return None

    async def judge_conversation_result(
        self, *, full_transcript: str
    ) -> JudgeCallResult[ConversationJudgeOutput]:
        if not self._llm:
            return JudgeCallResult(output=None, error="评审模型未配置")
        prompt = f"""{CONV_JUDGE_INSTRUCTIONS}

【完整对话】（每轮以 `[Turn k]` 开头）
{full_transcript[: eval_judge_conversation_transcript_max_chars()]}
"""
        try:
            out = await allm_json_call(
                self._llm,
                prompt,
                ConversationJudgeOutput,
                max_tokens=_CONV_JUDGE_JSON_MAX,
                agent="EvalJudgeService.judge_conversation",
            )
            return JudgeCallResult(output=out)
        except LLMCallError as e:
            error = _judge_error_message(e)
            logger.warning("conversation judge failed: {}", error)
            return JudgeCallResult(output=None, error=error)

    async def judge_conversation(
        self, *, full_transcript: str
    ) -> ConversationJudgeOutput | None:
        result = await self.judge_conversation_result(full_transcript=full_transcript)
        return result.output

    async def stream_conversation_compare(
        self,
        *,
        baseline_transcript: str,
        replay_transcript: str,
        baseline_judge: ConversationJudgeOutput | None,
        replay_judge: ConversationJudgeOutput | None,
    ) -> AsyncIterator[str]:
        """流式输出中文对比与建议（非 JSON）。"""
        if not self._llm:
            yield "[错误] 未配置评审模型 API Key（eval_judge_api_key / zhipu_api_key）"
            return
        cap_each = eval_judge_compare_transcript_each_max_chars()
        cap_single = eval_judge_conversation_transcript_max_chars()
        b_tr = (baseline_transcript or "").strip()[:cap_each]
        r_tr = (replay_transcript or "").strip()[:cap_each]
        b_json = (
            baseline_judge.model_dump_json(ensure_ascii=False)
            if baseline_judge
            else "null"
        )
        r_json = (
            replay_judge.model_dump_json(ensure_ascii=False) if replay_judge else "null"
        )
        if baseline_judge and replay_judge:
            prompt = f"""你是访谈对话评测专家。下面给出两份完整对话 transcript 及各自的整体打分（JSON）。请用中文直接写正文（不要用 JSON、不要用 Markdown 代码块）：

【A：导出基准对话】（历史快照：用户与当时导出的线上 AI，多轮合并为一篇）
{b_tr}

【B：本次回放/新测对话】（用户句与基准对齐，AI 为当前后端重新生成）
{r_tr}

【A 的整体评分 JSON】
{b_json}

【B 的整体评分 JSON】
{r_json}

请依次撰写：
1) 两段对话在整体体验上的主要差异（情绪承接、信息挖掘、人物建模、访谈结构、提问质量、上下文与重复盘问等）；
2) B 相对 A 的优点与不足；
3) 若 B 在关键维度明显弱于 A，给出可操作的改进方向（系统提示、访谈策略、模型或温度等）。

笔调简洁、偏执行清单。"""
        elif replay_judge:
            r_one = (replay_transcript or "").strip()[:cap_single]
            prompt = f"""{COMPARE_CONV_STREAM_HINT}

【回放/新测 transcript】
{r_one}

【整体评分 JSON】
{r_json}
"""
        else:
            yield "[错误] 缺少回放对话评分，无法生成建议"
            return

        llm = self._llm
        if hasattr(llm, "bind"):
            llm = llm.bind(max_tokens=_COMPARE_STREAM_MAX)
        try:
            async for chunk in llm.astream(prompt):
                piece = getattr(chunk, "content", None)
                if piece:
                    yield piece
        except Exception as e:
            logger.warning("conversation compare stream failed: {}", e)
            yield f"\n\n[流式输出中断：{e}]"

    async def judge_memoir(
        self,
        *,
        memoir_markdown: str,
        source_transcript: str = "",
        reference_memoir_markdown: str = "",
        evidence_notes: str = "",
    ) -> MemoirJudgeOutput | None:
        result = await self.judge_memoir_result(
            memoir_markdown=memoir_markdown,
            source_transcript=source_transcript,
            reference_memoir_markdown=reference_memoir_markdown,
            evidence_notes=evidence_notes,
        )
        return result.output

    async def judge_memoir_result(
        self,
        *,
        memoir_markdown: str,
        source_transcript: str = "",
        reference_memoir_markdown: str = "",
        evidence_notes: str = "",
    ) -> JudgeCallResult[MemoirJudgeOutput]:
        if not self._llm:
            return JudgeCallResult(output=None, error="评审模型未配置")
        prompt = _build_memoir_judge_prompt(
            memoir_markdown=memoir_markdown,
            source_transcript=source_transcript,
            reference_memoir_markdown=reference_memoir_markdown,
            evidence_notes=evidence_notes,
        )
        try:
            out = await allm_json_call(
                self._llm,
                prompt,
                MemoirJudgeOutput,
                max_tokens=_MEMOIR_JSON_MAX,
                agent="EvalJudgeService.judge_memoir",
            )
            return JudgeCallResult(output=out)
        except LLMCallError as e:
            error = _judge_error_message(e)
            logger.warning("memoir judge failed: {}", error)
            return JudgeCallResult(output=None, error=error)