life-echo/api/app/features/memory/extractor.py

"""从 transcript 块中抽取结构化事实（LLM + JSON）。"""

from __future__ import annotations

from typing import Any

from app.core.langchain_llm import ainvoke_json_object, invoke_json_object
from app.core.logging import get_logger
from app.features.memory.llm_schemas import (
    FactsExtractionPayload,
    facts_payload_to_dicts,
    parse_json_payload,
)

logger = get_logger(__name__)


def _max_transcript_chars() -> int:
    from app.core.config import settings

    return settings.memory_enrichment_max_chars


def _facts_extraction_instructions(narrator_label: str) -> str:
    return (
        "你是回忆录事实抽取助手。用户正在口述人生回忆，所有内容默认是**过去发生的事**，"
        "而非当前或未来计划（除非原文明确说「现在」「打算」「准备将要」等）。\n\n"
        "## 抽取规则\n"
        "1. subject 必须用明确的人名或固定称谓：\n"
        f"   - 叙述者本人统一用「{narrator_label}」\n"
        "   - 其他人用全名或稳定专名（如「王伟」），禁止用「他」「她」「我」「我们大伙」等代词作 subject；"
        "若代词在上下文中可唯一解析为某人，则 subject 写该人姓名/专名\n"
        "2. 事件、职务变动、地点迁移等一律按**过去回忆**理解；travel/调动/命令类表述勿写成「即将要做」"
        "除非原文明确为未来时态\n"
        "3. 若可推断大约年代或人生阶段，将 approximate_era 写入 object_json（与 value 等字段并存），"
        '例如 "1990年代"、"2001年"、"退休后"、"30岁前后"\n'
        "4. fact_type: person|event|relation|place|milestone\n"
        "5. predicate：简短中文谓语（如「出生地」「担任职务」「调往」）\n"
        "6. object_json：字符串或对象；可含 value、approximate_era 等\n"
        "7. confidence 0..1；source_chunk_id 必须等于某段 [chunk_id=...] 中的 id\n\n"
        '只输出 JSON：{"facts":[...]}，无事实则 {"facts":[]}。\n\n'
    )


def extract_facts_from_transcript_sync(
    llm: Any,
    numbered_blocks: str,
    *,
    narrator_name: str | None = None,
) -> list[dict]:
    """同步：带 chunk_id 标记的文本 → 事实列表。"""
    if not llm or not (numbered_blocks or "").strip():
        return []
    text = numbered_blocks.strip()[: _max_transcript_chars()]
    narrator_label = (narrator_name or "").strip() or "叙述者"
    prompt = _facts_extraction_instructions(narrator_label) + text
    try:
        raw = invoke_json_object(
            llm,
            prompt,
            max_tokens=4096,
            agent="memory.extract_facts_sync",
        )
        parsed = parse_json_payload(raw, FactsExtractionPayload)
        if parsed is None:
            return []
        return facts_payload_to_dicts(parsed)
    except (TypeError, ValueError) as e:
        logger.warning("extract_facts_from_transcript_sync 解析失败: {}", e)
        return []


async def extract_facts_from_transcript_async(
    llm: Any,
    numbered_blocks: str,
    *,
    narrator_name: str | None = None,
) -> list[dict]:
    """异步版。"""
    if not llm or not (numbered_blocks or "").strip():
        return []
    text = numbered_blocks.strip()[: _max_transcript_chars()]
    narrator_label = (narrator_name or "").strip() or "叙述者"
    prompt = _facts_extraction_instructions(narrator_label) + text
    try:
        raw = await ainvoke_json_object(
            llm,
            prompt,
            max_tokens=4096,
            agent="memory.extract_facts_async",
        )
        parsed = parse_json_payload(raw, FactsExtractionPayload)
        if parsed is None:
            return []
        return facts_payload_to_dicts(parsed)
    except (TypeError, ValueError) as e:
        logger.warning("extract_facts_from_transcript_async 解析失败: {}", e)
        return []


async def extract_facts(chunk_text: str, *, user_id: str) -> list[dict]:
    """兼容旧接口：单块文本（无 chunk id 时传空 source_chunk_id）。"""
    from app.core.db import AsyncSessionLocal
    from app.core.dependencies import get_llm_provider_fast
    from app.features.user.models import User

    llm = get_llm_provider_fast().langchain_llm
    narrator_name: str | None = None
    try:
        async with AsyncSessionLocal() as db:
            u = await db.get(User, user_id)
            if u and (u.nickname or "").strip():
                narrator_name = (u.nickname or "").strip()
    except Exception:
        pass

    blocks = f"[chunk_id=null]\n{chunk_text}"
    facts = await extract_facts_from_transcript_async(
        llm, blocks, narrator_name=narrator_name
    )
    for f in facts:
        if f.get("source_chunk_id") in (None, "null", ""):
            f["source_chunk_id"] = None
    return facts