Repository navigation
RepoWiki v0.4.2: knowledge cards, symbol skeletons for big files
A feature release: every generated wiki now opens with a knowledge-cards page, and oversized Python files are read by structure instead of by head.
What's new
- Knowledge cards page — each module gets one compact card: its purpose, file count, the symbols and concepts it exposes, its internal links, and a jump to the full module page. It is the fastest way to pick a starting point before reading anything longer, and it costs no extra LLM calls because it derives entirely from the existing analysis output. The incremental pipeline rewrites it whenever module data changes.
- Symbol skeletons for oversized files — a Python module too big for the per-file context budget used to contribute only its first 4,096 characters. The analyzer now sees an ast-derived skeleton instead: every top-level class and function with its signature and docstring, so a 2,000-line module is read by structure.
- Retrieval eval fixtures — a fixture suite gates retrieval quality as a regression check for the RAG index.
- SPA fallback no longer swallows API misses on Windows — unknown
/apipaths return real errors instead of the app shell.
Live demo: the wiki for this very repo, generated by repowiki scan . --site: https://he-yufeng.github.io/RepoWiki/
126 tests green. pip install -U repowiki.
这个版本让每个生成的 wiki 一打开就是知识卡片页,超大 Python 文件也终于按结构被读完。
更新内容
- 知识卡片页:每个模块压成一张卡片——用途、文件数、暴露的符号与概念、模块内链接、通往完整模块页的跳转。挑入口最快的方式,而且完全从既有分析结果派生,零额外 LLM 调用;模块数据一变增量管线就重写它。
- 超大文件的符号骨架:过去超出单文件上下文预算的 Python 模块只贡献前 4096 个字符,现在分析器看到的是 ast 提取的符号骨架(顶层类和函数的签名与 docstring),2000 行的模块按结构被读完。
- 检索评测夹具:一套评测夹具作为 RAG 索引质量的回归门。
- SPA 回退不再吞掉 API miss(Windows):未知
/api路径返回真实错误而不是应用外壳。
126 个测试全绿。pip install -U repowiki 升级。