life-echo/api/app/agents/memoir/batch_phase1_prep.py

"""
Phase1 批处理：一次 LLM 调用完成多段的抽取 + 章节分类（与逐段循环语义对齐）。
"""

from __future__ import annotations

import math
from dataclasses import dataclass
from typing import Any, Callable, Dict, List

from app.agents.memoir.prompts import get_batch_memoir_phase1_prep_prompt
from app.agents.memoir.schemas import BatchPhase1LLMOutput
from app.agents.state_schema import MemoirStateSchema
from app.core.config import settings
from app.core.llm_call import LLMCallError, llm_json_call
from app.core.logging import get_logger
from app.features.conversation.models import Segment

logger = get_logger(__name__)

def _slots_snapshot(state: MemoirStateSchema) -> dict:
    snap: dict = {}
    for stage, buckets in (state.slots or {}).items():
        snap[stage] = {}
        for k, v in (buckets or {}).items():
            if hasattr(v, "snippet"):
                sn = getattr(v, "snippet", None) or ""
            elif isinstance(v, dict):
                sn = (
                    (v.get("snippet") or "")
                    if isinstance(v.get("snippet"), str)
                    else ""
                )
            else:
                sn = ""
            snap[stage][k] = (sn or "")[:120]
    return snap


@dataclass(frozen=True)
class BatchPhase1SegmentRow:
    detected_stage: str
    slots: Dict[str, str]
    chapter_category_raw: str


def run_batch_phase1_prep(
    segments: List[Segment],
    state: MemoirStateSchema,
    llm: Any,
    *,
    language: str = "zh",
) -> Dict[str, BatchPhase1SegmentRow]:
    """对 segments 顺序批量调用 LLM；返回 id → 行。id 集合必须与入参完全一致。"""
    if not llm:
        raise ValueError("batch phase1 requires llm")
    if not segments:
        return {}
    items = [(str(s.id), (s.user_input_text or "").strip()) for s in segments]
    prompt = get_batch_memoir_phase1_prep_prompt(
        system_current_stage=state.current_stage or "childhood",
        slots_snapshot=_slots_snapshot(state),
        segment_items=items,
        language=language,
    )
    try:
        parsed = llm_json_call(
            llm,
            prompt,
            BatchPhase1LLMOutput,
            max_tokens=int(settings.memoir_phase1_batch_llm_max_tokens),
            agent="BatchPhase1Prep.run",
        )
    except LLMCallError as e:
        logger.warning("batch phase1 LLM 解析失败: {}", e)
        raise ValueError("batch phase1: llm parse failed") from e

    rows = parsed.segments
    if not rows:
        raise ValueError("batch phase1: segments must be a non-empty list")

    by_id: Dict[str, BatchPhase1SegmentRow] = {}
    for row in rows:
        sid = str(row.id).strip()
        if not sid:
            continue
        ds = str(row.detected_stage or "").strip().lower()
        slots_raw = row.slots or {}
        slots = {
            k: v if isinstance(v, str) else str(v)
            for k, v in slots_raw.items()
            if k and isinstance(k, str)
        }
        cat_raw = str(row.chapter_category or "")
        by_id[sid] = BatchPhase1SegmentRow(
            detected_stage=ds or (state.current_stage or "childhood"),
            slots=slots,
            chapter_category_raw=cat_raw,
        )

    expected = {str(s.id) for s in segments}
    if by_id.keys() != expected:
        missing = expected - by_id.keys()
        extra = by_id.keys() - expected
        logger.warning("batch phase1 id mismatch missing={} extra={}", missing, extra)
        raise ValueError("batch phase1 response segment ids do not match input")
    return by_id


def _run_batch_phase1_prep_chunk_with_bisect(
    segments: List[Segment],
    state: MemoirStateSchema,
    llm: Any,
    *,
    language: str = "zh",
) -> Dict[str, BatchPhase1SegmentRow]:
    """单块 LLM；失败时（如输出截断）将块二等分重试直至单段。"""
    try:
        return run_batch_phase1_prep(segments, state, llm, language=language)
    except ValueError:
        if len(segments) <= 1:
            raise
        mid = len(segments) // 2
        if mid < 1:
            raise
        left = _run_batch_phase1_prep_chunk_with_bisect(
            segments[:mid], state, llm, language=language
        )
        right = _run_batch_phase1_prep_chunk_with_bisect(
            segments[mid:], state, llm, language=language
        )
        merged = {**left, **right}
        expected = {str(s.id) for s in segments}
        if merged.keys() != expected:
            raise ValueError(
                "batch phase1 chunked bisect merge: segment ids do not match input"
            ) from None
        return merged


def run_batch_phase1_prep_chunked(
    segments: List[Segment],
    state: MemoirStateSchema,
    llm: Any,
    *,
    chunk_size: int,
    on_chunk: Callable[[int, int], None] | None = None,
    language: str = "zh",
) -> Dict[str, BatchPhase1SegmentRow]:
    """
    将 segments 按 chunk_size 切片多次调用 Phase1 批处理 LLM，合并 by_id。
    单块仍失败时在块内二分回退（最后回退到单段），与 orchestrator 外层逐段回退衔接。
    """
    if not segments:
        return {}
    if chunk_size < 1:
        chunk_size = 1
    n = len(segments)
    total_chunks = max(1, math.ceil(n / chunk_size))
    merged: Dict[str, BatchPhase1SegmentRow] = {}
    for i in range(0, n, chunk_size):
        chunk_idx = i // chunk_size + 1
        sub = segments[i : i + chunk_size]
        logger.info(
            "event=batch_phase1_chunk chunk_idx={}/{} segment_count={} batch_path=chunked "
            "msg=Phase1 批处理分块调用",
            chunk_idx,
            total_chunks,
            len(sub),
        )
        part = _run_batch_phase1_prep_chunk_with_bisect(
            sub, state, llm, language=language
        )
        merged.update(part)
        if on_chunk is not None:
            on_chunk(chunk_idx, total_chunks)
    expected = {str(s.id) for s in segments}
    if merged.keys() != expected:
        missing = expected - merged.keys()
        extra = merged.keys() - expected
        logger.warning(
            "batch phase1 chunked id mismatch missing={} extra={}",
            missing,
            extra,
        )
        raise ValueError("batch phase1 chunked: merged segment ids do not match input")
    return merged