AI导哈

产品名称规范化工具

一个基于字典驱动的产品/工具名称规范化工具，用于修正AI生成文本中的错误拼写。

  • by serejaris
  • Python
  • 本地部署
  • 更新于
在 GitHub 查看 serejaris/product-name-normalizer

使用说明

product-name-normalizer

Коротко: Рабочий репозиторий: Dictionary-driven product/tool name normalization for AI text (Python + MCP server).

Что здесь

  • Назначение: Рабочий репозиторий: Dictionary-driven product/tool name normalization for AI text (Python + MCP server).
  • Основной стек: Python.
  • Видимость: публичный репозиторий.
  • Статус: активный репозиторий; актуальность проверять по issues и последним коммитам.

Где смотреть работу

  • Задачи и текущие решения: GitHub Issues этого репозитория.
  • Код и материалы: файлы в корне и профильные папки проекта.
  • Связь с HQ: если проект влияет на продукт, контент или воронку, сверяйте канон в 0_hq и репозитории-владельце.

Для агентов

  • Сначала прочитайте этот README и открытые issues.
  • Не переносите сюда канон соседних проектов без ссылки на источник.
  • Перед правками проверьте существующие scripts, package.json/pyproject и локальные инструкции.

Исходный README

product-name-normalizer

Local (user-scope) normalization of product/tool names for AI-generated text.

This exists because LLMs (and chat participants) routinely misspell tool names and the mistakes propagate into digests/posts: Cloudcode -> Claude Code, Antygravity -> Antigravity, Wisprflow -> Wispr Flow, etc.

The goal is to make this a single, reusable component:

  • as a Python module (stdlib-only at import time)
  • as a user-scope MCP server (for Claude Code / skills / other MCP clients)

Context / Why

In tg-digest-opus the digest for 2026-02-08 was produced with incorrect product names. Fixing this inside prompts is fragile; it keeps leaking in from:

  • raw chat messages (people write names wrong)
  • LLM completions (the model "learns" the wrong spelling from context)

So the fix is centralized and deterministic: a dictionary-driven normalizer that any consumer can call.

Python usage (zero deps)

from term_fixer import fix_terms

text = fix_terms("Cloudcode vs Cursor")

Dictionary is stored in data/product-terms.json (override with TERM_FIXER_TERMS_PATH).

Dictionary format

product-terms.json is a JSON map:

{
  "Claude Code": ["Cloudcode", "ClaudeCode"],
  "Antigravity": ["Antygravity", "Anti-gravity"]
}

Add / extend a term

from term_fixer import add_term

add_term("Claude Code", ["Cloudcode", "ClaudeCode"])

Or edit data/product-terms.json manually (it is a simple JSON map).

How It Works

  • Dictionary format: { "Correct Name": ["wrong1", "wrong2"] }
  • Replacements: case-insensitive regex with word boundaries (\\b...\\b)
  • HTML safety: it avoids touching tag/attribute content by only rewriting text segments outside <...>
  • It applies longest variants first to reduce partial-overlap issues.
  • Compiled regex rules are cached and invalidated automatically when product-terms.json changes (mtime-based).

Files In This Repo

  • term_fixer.py: stdlib-only core + MCP entrypoint
  • data/product-terms.json: the dictionary used by default
  • scripts/smoke_mcp.py: end-to-end MCP smoke test

Non-goals (for now):

  • fuzzy matching / spelling suggestions
  • language-aware morphology

MCP server (Claude Code)

Run locally via stdio:

uv -q --directory ~/Documents/GitHub/product-name-normalizer run term_fixer.py

Add to Claude Code (user scope):

claude mcp add product-name-normalizer -s user -- uv -q --directory ~/Documents/GitHub/product-name-normalizer run term_fixer.py

Tools:

  • fix_terms(text: str) -> str
  • add_term(correct: str, wrong_variants: list[str]) -> str

Integration (tg-digest-opus)

tg-digest-opus loads this repo in best-effort mode (no hard dependency).

  • Default lookup path: ~/Documents/GitHub/product-name-normalizer
  • Override: set TERM_FIXER_REPO=/path/to/product-name-normalizer

Development

Run tests:

uv run pytest

Smoke test MCP (spawns the server and calls tools over stdio):

uv run python scripts/smoke_mcp.py

相关 MCP