lhl
3f4d17cf69
真实 ja 试运行暴露：auto 模式 fallback 误用 impact/data 源数据（含中文元素名 止损风控等 + 汉字标签无假名）→ 把日文文档误判为期望 zh → LLM 正常日文输出被 判违规 → 硬失败。修复：fallback 改用 write_rules/design_rules（RAG 自日文作成 说明书/记入规则检索，含假名可靠判 ja）；加回归测试 test_generate_chapter_auto_with_ja_rules_but_chinese_source_data_enforces_ja。 修复后 ja 真实试运行成功（7章）；zh 真实试运行成功（7章）；425 passed / 99.15%
180 lines
8.5 KiB
Python
180 lines
8.5 KiB
Python
"""Writer Agent：调用推理引擎生成单章内容（Phase 5）。
|
||
|
||
接入真实 InferenceEngine.chat_structured（session_id/prompt/variables/schema/retry_count），
|
||
并对章节级失败做有限重试；token 估算分块（真实拼回留待后续并发实现）。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
|
||
from genesis.inference.engine import InferenceEngine
|
||
from genesis.inference.prompt_registry import PromptRegistry
|
||
from genesis.inference.types import Prompt, StructuredResult
|
||
from genesis.writer.models import ChapterContent, GenerationContext
|
||
from genesis.writer.writer_state import WriterState
|
||
from genesis.writer.exceptions import WriterGenerationError
|
||
from genesis.writer.language import (
|
||
find_language_violations,
|
||
resolve_expected_language,
|
||
)
|
||
|
||
|
||
WRITER_PROMPT_TEMPLATE = (
|
||
"你是概要设计书撰写专家。\n"
|
||
"章节: {{chapter_id}} {{title}}\n"
|
||
"本章小节结构:\n{{sub_headings}}\n"
|
||
"写入规则:\n{{write_rules}}\n"
|
||
"设计规则:\n{{design_rules}}\n"
|
||
"模板样式:\n{{template_styles}}\n"
|
||
"影响调查上下文:\n{{impact}}\n"
|
||
"参考资料（本章对应数据）:\n{{data}}\n"
|
||
"请输出符合 schema 的章节内容 JSON。\n"
|
||
"【小节约束】若上方「本章小节结构」非空，必须按该小节顺序组织内容，"
|
||
"每个小节以 type=heading、level=2 的内容块开头（标题使用小节原文），随后为该小节的内容块；"
|
||
"不得遗漏或新增小节。若「本章小节结构」为空，则不得输出任何 type=heading 的内容块"
|
||
"（章节标题已由模板提供），仅以 paragraph/table/list/note 块组织内容。\n"
|
||
"【主题约束】本章必须且仅围绕标题「{{title}}」所对应的主题撰写，"
|
||
"严格以「参考资料（本章对应数据）」中的要件定义数据和「影响调查上下文」为核心依据；"
|
||
"禁止输出与本章无关的系统整体架构、通用设计说明等内容，禁止套用其他章节的主题。"
|
||
"若本章数据为空，则基于规则与影响调查上下文简要撰写，不得虚构数据。\n"
|
||
"【语言约束】章节正文（所有 block 的 text 字段）所使用的自然语言：{{language_instruction}}\n"
|
||
)
|
||
|
||
|
||
CHAPTER_OUTPUT_SCHEMA = {
|
||
"type": "object",
|
||
"properties": {
|
||
"title": {"type": "string"},
|
||
"blocks": {
|
||
"type": "array",
|
||
"items": {
|
||
"type": "object",
|
||
"properties": {
|
||
"type": {"type": "string"},
|
||
"text": {"type": "string"},
|
||
"level": {"type": "integer"},
|
||
"headers": {"type": "array", "items": {"type": "string"}},
|
||
"rows": {"type": "array", "items": {"type": "array", "items": {"type": "string"}}},
|
||
"items": {"type": "array", "items": {"type": "string"}},
|
||
"source_uris": {"type": "array", "items": {"type": "string"}},
|
||
},
|
||
"required": ["type"],
|
||
},
|
||
},
|
||
},
|
||
"required": ["title", "blocks"],
|
||
}
|
||
|
||
|
||
class WriterAgent:
|
||
def __init__(
|
||
self,
|
||
session_id,
|
||
engine: InferenceEngine,
|
||
prompt_registry: PromptRegistry,
|
||
state: WriterState,
|
||
max_retries: int = 2,
|
||
) -> None:
|
||
self.session_id = session_id
|
||
self.engine = engine
|
||
self.prompt_registry = prompt_registry
|
||
self.state = state
|
||
self.max_retries = max_retries
|
||
|
||
def _chunk_source(self, source) -> list[dict]:
|
||
if source is None:
|
||
return [{"index": 0, "text": ""}]
|
||
src = source if isinstance(source, str) else str(source)
|
||
n = max(1, len(src) // 1800 + 1)
|
||
return [{"index": i, "text": src[i * 1800:(i + 1) * 1800]} for i in range(n)]
|
||
|
||
def _resolve_prompt(self) -> Prompt:
|
||
"""取用/注册 writer.chapter 模板，并保证返回 Prompt 对象。
|
||
|
||
优先使用 get_or_create（与测试 Fake 兼容）；真实 PromptRegistry 无该方法时，
|
||
回退为 register + get。注意：真实 get() 返回模板字符串而非 Prompt，
|
||
而引擎 _render_prompt 仅对 Prompt 对象做变量渲染、str 原样发送
|
||
（否则 {{占位符}} 不被替换直接进 LLM）→ 此处统一包装为 Prompt。
|
||
"""
|
||
get_or_create = getattr(self.prompt_registry, "get_or_create", None)
|
||
if get_or_create is not None:
|
||
resolved = get_or_create("writer.chapter", WRITER_PROMPT_TEMPLATE)
|
||
else:
|
||
self.prompt_registry.register("writer.chapter", "1", WRITER_PROMPT_TEMPLATE)
|
||
resolved = self.prompt_registry.get("writer.chapter", "1")
|
||
if isinstance(resolved, Prompt):
|
||
return resolved
|
||
template = getattr(resolved, "template", None) or str(resolved)
|
||
version = str(getattr(resolved, "version", "1") or "1")
|
||
return Prompt(name="writer.chapter", version=version, template=template)
|
||
|
||
def _call_llm(self, context: GenerationContext) -> dict:
|
||
prompt = self._resolve_prompt()
|
||
result = self.engine.chat_structured(
|
||
session_id=self.session_id,
|
||
prompt=prompt,
|
||
variables=context.to_vars(),
|
||
schema=CHAPTER_OUTPUT_SCHEMA,
|
||
retry_count=2,
|
||
)
|
||
# 真实 InferenceEngine.chat_structured 为 async；测试用同步 FakeEngine 返回普通对象。
|
||
# 兼容两者：若返回协程则通过 asyncio.run 驱动（调用方 orchestrator/qa_loop/脚本均为同步上下文）。
|
||
if asyncio.iscoroutine(result):
|
||
result = asyncio.run(result)
|
||
if result.status not in ("ok", "fallback"):
|
||
# 透传底层错误详情（如 401 鉴权失败原因），便于人工门禁定位
|
||
err = getattr(result, "error", None)
|
||
code = getattr(result, "error_code", None)
|
||
suffix = ""
|
||
if err:
|
||
suffix += f"; {err}"
|
||
if code:
|
||
suffix += f" (code={code})"
|
||
raise WriterGenerationError(f"引擎返回异常状态: {result.status}{suffix}")
|
||
return result.data
|
||
|
||
def generate_chapter(self, context: GenerationContext) -> ChapterContent:
|
||
self._chunk_source(context.structured_source) # 分块可用性验证（真实拼回留待后续）
|
||
|
||
# 步骤 A：推导本章期望输出语言（仅依赖 context，循环前置，稳定）
|
||
# fallback 用「规则文档」（write_rules/design_rules，RAG 自日文作成说明书/记入规则检索，
|
||
# 含假名可判日文）。不能用 impact/data（源数据）：样本含中文元素名（止损风控等）、
|
||
# 影响标签为汉字无假名 → 会把日文文档误判为期望 zh（真实试运行暴露）。
|
||
expected = resolve_expected_language(
|
||
explicit=context.output_language,
|
||
title=context.title,
|
||
fallback_texts=[
|
||
"\n".join(context.write_rules or []),
|
||
"\n".join(context.design_rules or []),
|
||
],
|
||
)
|
||
|
||
last_err: Exception | None = None
|
||
for _ in range(max(1, self.max_retries)):
|
||
try:
|
||
data = self._call_llm(context)
|
||
except Exception as e: # 引擎可能抛出任意异常，统一按章节级失败重试
|
||
last_err = e
|
||
continue
|
||
try:
|
||
content = ChapterContent.from_llm(context.chapter_id, context.title, data)
|
||
except (KeyError, TypeError, ValueError) as e:
|
||
last_err = e
|
||
continue
|
||
# 模板结构为准（design §6.5）：无子节结构的章，剔除 LLM 自造的
|
||
# heading 块（prompt 约束为尽力而为，此处程序化强制）
|
||
if not (getattr(context.template_marker, "sub_headings", None) or []):
|
||
content.blocks = [b for b in content.blocks if b.type != "heading"]
|
||
# 步骤 A：输出语言一致性强制（期望语言可推导时，违规按失败重试）
|
||
if expected:
|
||
viol = find_language_violations(content.blocks, expected)
|
||
if viol:
|
||
last_err = WriterGenerationError(
|
||
f"章节 {context.chapter_id} 语言不一致（期望 {expected}，"
|
||
f"发现 {len(viol)} 处违规正文）"
|
||
)
|
||
continue
|
||
self.state.record_success(content)
|
||
return content
|
||
raise WriterGenerationError(f"章节 {context.chapter_id} 重试耗尽: {last_err}")
|