paeg-lang-style

Chinese | English

What This Is

paeg-lang-style is a Chinese language standards module — three-layer architecture (user requirements):

Originated from the PAEG educational agent (v0.12-v0.71 iterations), refactored into a zero-host-dependency standalone plugin — any Python project can integrate it.

Related MCP server: sensitive-lexicon-mcp

Core Features

• Extensible rule set: rules externalized to data/rules.json; appending hot-loads them (RuleRegistry)

• General principles direct the LLM: lexical completeness/syntactic completeness/sufficient-adverbial principles let the LLM generalize rather than memorize word-by-word (avoiding the narrowness of "only optimizing 倦 into 疲倦")

• Deterministic fallback: enumeration-layer rules ("我在这里听着你。"→"我就在这里听你说说。") guard against LLM disobedience

• Rule-ID feedback loop: rewrite feedback carries "violates #rule-lx-001", forming a rule-generation-feedback loop

• Dynamic forbidden-word lexicon: runtime add/remove + external JSON hot-loading

• Injection-based design: chat_fn is force-injected — plug in your own LLM call, zero host coupling

• Three profile tiers: general / teaching / confessional — assemble system prompts by scenario

• AI-taste detection: 5-dimensional signals — sentence-length variance/transition-word density/three-part structure/dashes/paragraph symmetry

• 8+ grammar rules: GB/T 15834 punctuation standards + six categories of faulty sentences + teaching language standards

• 75 tests all green + 20 behavioral consistency checks (string-equal vs. the original PAEG implementation)

Installation

# 方式 1：pip 安装（推荐）
pip install -e /path/to/paeg-lang-style-plugin

# 方式 2：直接引用（零安装）
# 把 src/ 加入 sys.path 即可
import sys
sys.path.insert(0, "/path/to/paeg-lang-style-plugin/src")

Requires Python 3.9+. Zero third-party runtime dependencies.

Integrating as an MCP Server (install and use directly, just like MCP)

Accessibility standard (§3.109): after pip install in any project, declare it in the MCP client config to integrate — zero code bridge, zero host dependency.

# 方式 1：console_scripts 入口（pip install 后）
paeg-lang-style-mcp

# 方式 2：python -m 入口（源码运行）
python -m paeg_lang_style.mcp_server

MCP client config declaration (e.g. config/mcp_servers.json):

{
  "mcpServers": {
    "paeg-lang-style": {
      "command": "python",
      "args": ["-m", "paeg_lang_style.mcp_server"],
      "cwd": "D:/wbo-workspace/paeg_project/paeg-lang-style-plugin"
    }
  }
}

Exposed MCP tools (7):

Quick Start

Three steps: install → assemble the prompt → process the output.

from paeg_lang_style import RuleRegistry, make_refiner, gate_content

# Step 1: 语法规则拼进你的系统提示词（谁用都拼）
system_prompt = "你是教育智能体，负责讲解数学概念。"
system_prompt += RuleRegistry().build_prompt("teaching")   # 通则层指挥 LLM

# Step 2: 注入你的 LLM 调用（改写脚本）
def my_llm(system, user, max_tokens=800, **kw):
    return call_my_llm_api(system, user, max_tokens=max_tokens)

refiner = make_refiner(chat_fn=my_llm)

# Step 3: LLM 输出后处理（L0 规则 + L2 重写）
raw_output = "总的来说，让我们一起赋能这个时代！"
clean = gate_content(raw_output, refiner=refiner)
# → "我们把这个时代里的每一个孩子，把他们的潜能一步步唤起。"

Table of Contents

• Core Concepts

• External Project Integration Guide

• Extensibility

• Maintainability

• Built-in Grammar Rules

• Forbidden-Word Lexicon

• API Reference

• Configuration Reference

• Architecture Design

• Integration with the PAEG Main Project

• Testing

• Contribution Guide

• Changelog

• License

Core Concepts

Three-layer architecture + rule-set data model — extensibility runs through all three places: rule definition/detection/prompt generation:

graph LR
    A[LLM 生成文本] --> B[gate_content 守门]
    B --> C{L0 规则检测}
    C -->|命中列举层| D[确定性替换<br/>听着你→听你说说]
    C -->|通则触发| E[L2 refiner.refine<br/>chat_fn 注入 LLM]
    E --> F[反馈带规则 ID<br/>违反 #rule-lx-001]
    F --> G[多轮 Self-Refine]
    D --> H[输出]
    G --> H
    H --> I[收口: 规则再跑一遍]

Rule data model (Rule — extensible via external JSON):

{
  "id": "rule-lx-general-001",
  "type": "general",
  "category": "lexical",
  "pattern": "(倦|乏|沉|累|苦|慌|虚|弱|低|烦|闷|困|急|乱)",
  "replacement": null,
  "message": "存在单字状态词——应扩展为完整双字词形",
  "prompt_block": "### 词法完整通则（指挥 LLM 泛化）...",
  "severity": "high",
  "enabled": true,
  "source": "builtin",
  "profile_tags": ["general", "teaching", "confessional"]
}

• type: "general" (general-principle layer): prompt_block is assembled into the system prompt, directing the LLM to generalize — "any single-character adjective expressing state/feeling must be expanded to its full two-character form (倦/乏/沉/累/苦/慌/虚/弱/低/烦/闷/困/急/乱)". The LLM generalizes to unlisted words on its own, rather than only fixing "倦→疲倦".

• type: "explicit" (enumeration layer): pattern + replacement deterministic fallback — the last line of defense when the LLM disobeys.

External Project Integration Guide

User requirement: any project/agent that wants to use our grammar rules module — how?

Scenario A: Only want "grammar rule constraints" (system prompt)

from paeg_lang_style import RuleRegistry

# 语法规则拼进自己的系统提示词（谁用都拼）
system = "你是我的客服机器人。"
system += RuleRegistry().build_prompt("general")   # 或 "teaching" / "confessional"

The rule fragments returned by build_prompt(profile) (lexical-completeness principle/syntactic-completeness principle/sufficient-adverbial principle/punctuation standards) can be concatenated directly into your LLM system prompt. This is the essence of "directing the LLM to use complete words" — no dependency on our rewriter.

Scenario B: Want to use the "rewriting script" to process LLM output

from paeg_lang_style import make_refiner, gate_content

# 注入你自己的 LLM 调用包装（chat_fn 强制注入，零宿主耦合）
def my_chat(system, user, max_tokens=800, **kw):
    return call_your_llm(system, user, max_tokens=max_tokens)

refiner = make_refiner(chat_fn=my_chat)
clean = gate_content(your_llm_output, refiner=refiner)   # L0 规则 + L2 重写
# 纯规则（不调 LLM）：
clean = gate_content(your_llm_output)                    # 病句/违禁词确定性修正

Scenario C: Want to use the "forbidden-word lexicon"

from paeg_lang_style import ForbiddenWords

fb = ForbiddenWords()                     # 内置违禁词（AI 腔/空洞大词/伪共情/网络用语）
fb.load_json("my_words.json")             # 合并你自己的词库（动态扩充）
fb.add("你们公司的禁词")                  # 运行时新增
hits = fb.detect(text)                    # → ["禁词1", "禁词2"]

Scenario D: Want to extend the rules (extensibility)

Edit data/rules.json and append a rule — it hot-loads:

{
  "rules": [
    {
      "id": "rule-my-001",
      "type": "explicit",
      "category": "lexical",
      "pattern": "你们行业的黑话",
      "replacement": "规范说法",
      "message": "这是行业黑话，应改规范",
      "severity": "medium",
      "enabled": true,
      "source": "user",
      "profile_tags": ["general"]
    }
  ]
}

reg = RuleRegistry()
reg.load("data/rules.json")    # 合并（追加即生效）
reg.watch("data/rules.json")   # mtime 变更自动热重载
reg.check_reload()             # 每次调用前检查

Scenario E: Full integration (including the rule-ID feedback loop)

from paeg_lang_style import RuleRegistry, make_refiner, gate_content, get_style_prompt

# 1. 系统提示词（规则 + 风格）
system = get_style_prompt("all") + "\n" + RuleRegistry().build_prompt("general")

# 2. 改写器（规则检测 → 反馈带 ID → 重写）
refiner = make_refiner(chat_fn=my_chat)

# 3. 守门（L0 规则 + L2 重写 + 收口）
final = gate_content(raw, refiner=refiner)

Extensibility

Maintainability

• Zero host dependency: imports no host project modules, independently testable

• Legacy API compatibility: rules.py/rules_enhanced.py kept as thin wrappers, backward compatible

• 75 tests: full coverage of rule-set loading/hot-reload/detection/assembly/user extension/corruption tolerance

• Behavioral consistency: 20 samples string-equal vs. the original PAEG implementation (zero drift)

• Corruption tolerance: keeps the previous rule set when JSON is corrupted — never "runs empty"

• Anti-bloat: token_budget controls system prompt length (default 800)

• Clear separation of modes: general-principle layer (directs the LLM) vs. enumeration layer (deterministic fallback) with distinct responsibilities

Built-in Grammar Rules

Sufficient-adverbial principle details (rule-sx-general-002, user-added):

Every action/judgment should be clearly qualified with sufficient adverbials — time, place, manner, condition, object, purpose. "复习单词。" → "你可以在每天睡前用十分钟复习单词。" "使用这个软件。" → "你可以在每天固定的时间使用这个软件。"

Forbidden-Word Lexicon

How to extend: ForbiddenWords().load_json("path.json") — JSON structure {"extra_forbidden": [...], "ai_tells_extra": [...]}.

API Reference

RuleRegistry

LanguageRefiner / make_refiner

gate_content / gate_short

ForbiddenWords

get_style_prompt

Configuration Reference

Environment Variables

data/ Files

Architecture Design

宿主系统（任何 Python 项目 / 智能体）
  system_prompt += RuleRegistry().build_prompt()    <- 语法规则拼系统提示词（谁用都拼）
  gate_content(output, refiner=make_refiner(chat))  <- 输出后处理
        |
        | 零宿主依赖（不 import 宿主任何模块）
        v
paeg_lang_style（独立插件）
  +------------------+  +-----------------+  +-----------------+
  | rule_registry    |  | forbidden.py    |  | ai_taste.py     |
  | 可扩充规则集      |  | 动态违禁词库     |  | AI 味检测        |
  +--------+---------+  +--------+--------+  +--------+--------+
           |                    |                     |
           v                    v                     v
  +-----------------------------------------------------------+
  | refiner.py（改写脚本：chat_fn 注入 + 规则 ID 闭环）           |
  | gate.py（守门入口：L0+L2 编排）                              |
  +-----------------------------------------------------------+
  data/（rules.json / forbidden_words.json / 语料）

Integration with the PAEG Main Project

The PAEG educational agent integrates through the single adapter layer infra/lang_plugin_bridge.py (R18/R20 zero-breakage iron law):

from infra.lang_plugin_bridge import gate_content, get_style_prompt, make_refiner
# 插件挂载 → 走插件；插件未挂载 → 静默回退 PAEG 原实现（旧文件永不删除）

See docs/integration_paeg.md for details.

Testing

python -m pytest tests/ -q
# 75 项：规则集加载/热重载/检测/拼装/用户扩充/损坏容错 + 通则 + 充分状语 + 行为一致性

Contribution Guide

Contributions welcome! Please read CONTRIBUTING.md (to be created) to learn about:

• Add new grammar rules: edit data/rules.json to append a Rule (with tests)

• Add forbidden words: ForbiddenWords.load_json or submit a PR directly to the built-in lexicon

• Code style: follow the existing module structure + comment conventions

Changelog

See CHANGELOG.md for details.

Acknowledgements

• PAEG Education Agent (v0.12-v0.71 iterations) — this plugin is extracted from its language specification module

• LanguageTool — declarative rule engine paradigm (dev.languagetool.org)

• textstat — readability measurement paradigm (github.com/textstat/textstat)

• GB/T 15834-2011《Usage of Punctuation Marks》 — national standard for punctuation rules

• Agent Skills Standard — progressive disclosure paradigm (agentskills.io)

License

MIT © 2026 PAEG Team — see the LICENSE file for details.