backend/knowledge_loader.py

"""加载 skus 和 skills 知识库 + 周礼原文，构建周礼知识上下文。"""

import json
from pathlib import Path


PROJECT_ROOT = Path(__file__).resolve().parent.parent
SKUS_DIR = PROJECT_ROOT / "skus"
SKILLS_DIR = PROJECT_ROOT / "generated_skills"
RAW_TEXT_FILE = PROJECT_ROOT / "zhouli_raw_texts.json"

# 原文分块：相邻短段合并，目标每块 300-500 字
RAW_CHUNK_TARGET = 400


def load_all_skus() -> list[dict]:
    """读取 skus/ 下所有 JSON 文件并返回列表。"""
    skus = []
    for f in sorted(SKUS_DIR.glob("*.json")):
        with open(f, "r", encoding="utf-8") as fh:
            skus.append(json.load(fh))
    return skus


def load_all_skills() -> list[dict]:
    """读取 generated_skills/ 下所有 SKILL.md，返回 {name, content} 列表。"""
    skills = []
    for skill_dir in sorted(SKILLS_DIR.iterdir()):
        skill_file = skill_dir / "SKILL.md"
        if skill_file.is_file():
            content = skill_file.read_text(encoding="utf-8")
            skills.append({"name": skill_dir.name, "content": content})
    return skills


def _merge_paragraphs(paragraphs: list[str], target: int = RAW_CHUNK_TARGET) -> list[str]:
    """将短段落合并为接近 target 字数的块。"""
    merged = []
    buf = ""
    for p in paragraphs:
        if buf and len(buf) + len(p) > target:
            merged.append(buf)
            buf = p
        else:
            buf = buf + p if not buf else buf + "\n" + p
    if buf:
        merged.append(buf)
    return merged


def load_raw_text_chunks() -> list[dict]:
    """读取周礼原文 JSON，按章节合并短段落后返回 chunk 列表。"""
    if not RAW_TEXT_FILE.exists():
        return []

    with open(RAW_TEXT_FILE, "r", encoding="utf-8") as f:
        data = json.load(f)

    chunks = []
    for chapter, paragraphs in data.items():
        merged = _merge_paragraphs(paragraphs)
        for i, text in enumerate(merged):
            chunks.append({
                "id": f"raw:{chapter}:{i}",
                "title": f"《周礼·{chapter}》原文（{i+1}）",
                "text": f"《周礼·{chapter}》原文：\n{text}",
            })
    return chunks


def build_chunks() -> list[dict]:
    """将 SKU、Skill 和原文拆分为检索用的 chunk 列表。

    返回: [{"id": str, "title": str, "text": str}, ...]
    """
    chunks = []

    for sku in load_all_skus():
        uid = sku.get("metadata", {}).get("uuid", "")
        name = sku.get("metadata", {}).get("name", "未命名")
        logic = sku.get("core_logic", {}).get("execution_body", "")
        schema = sku.get("schema_explanation", "")
        tags = ", ".join(sku.get("custom_attributes", {}).get("domain_tags", []))
        trigger = sku.get("trigger", {}).get("condition_logic", "")

        text = f"【{name}】\n领域: {tags}\n触发条件: {trigger}\n{logic}\n{schema}"
        chunks.append({"id": f"sku:{uid}", "title": name, "text": text})

    for skill in load_all_skills():
        chunks.append({
            "id": f"skill:{skill['name']}",
            "title": skill["name"],
            "text": skill["content"],
        })

    chunks.extend(load_raw_text_chunks())

    return chunks


def build_knowledge_text() -> str:
    """（保留兼容）将所有知识拼接为全量文本。"""
    parts = []
    for chunk in build_chunks():
        parts.append(chunk["text"])
    return "\n\n---\n\n".join(parts)