✨ feat(tsl-syntax-reference): harden retrieval and restructure pages
- flag weak candidates (no intent/heading/identifier/tag hit) and exit 2 when every candidate is weak: mis-hits used to be indistinguishable from real hits, so the retry-with-better-terms loop never fired - accept multiple ids per --section for batch fetch, failing atomically on any unknown id so a partial fetch cannot pass as complete - move query synonyms and page intent aliases to data/lexicon.json and enforce alias/page correspondence in --check; curation data no longer lives in the engine - document the weak-hit rule, batch fetch and prelude-once guidance in SKILL.md, with curation discipline in data/README.md - drop 11_pitfalls.md, renumber the trailing pages and spread retrieval tags across topics; lexicon keys are page filenames, so the renumbering and the new --check rule cannot land in separate commits
This commit is contained in:
@@ -0,0 +1,33 @@
|
||||
# 检索词表(lexicon.json)
|
||||
|
||||
`lexicon.json` 是 `scripts/lookup.py` 的策展检索数据,与引擎逻辑分离维护。
|
||||
两张表都只影响检索,不是事实正文。
|
||||
|
||||
## query_synonyms —— 口语同义词
|
||||
|
||||
中文口语词 → TSL 术语的展开(如「打印」→ 输出 / writeLn)。策展纪律:
|
||||
|
||||
- 同义词必须落在**目标章节已有的词面**上,否则不产生区分度。用户说的词在
|
||||
参考页正文里出现 0 次时,页级 alias 只能把整页抬起来,页内仍需要词法
|
||||
信号才能选中正确章节——节级词汇桥(如「交集」→ `intersect`)补的就是
|
||||
这一层。
|
||||
- 不要展开成泛化的中文概念词(如「集合运算」会命中别页交叉引用小节标题)。
|
||||
|
||||
## page_intent_aliases —— 页级意图短语
|
||||
|
||||
整句口语短语 → 参考页文件名的路由(命中一条 +80 分,是页级路由的唯一强
|
||||
信号)。约束:
|
||||
|
||||
- 键必须与 `references/` 下的文件名完全一致;每个参考页至少一条短语。
|
||||
键写错或参考页改名后 `_intent_score` 会静默返回 0 分,该页失去自然语言
|
||||
入口——`lookup.py --check` 和回归测试会拦住这两种漂移。
|
||||
|
||||
## 改动后的校验
|
||||
|
||||
```bash
|
||||
python scripts/lookup.py --check
|
||||
python -m unittest test.test_tsl_syntax_lookup_recall # 在仓库根目录运行
|
||||
```
|
||||
|
||||
新增参考页时,除补词表外还须在回归测试里补一条真实问法
|
||||
(`test/test_tsl_syntax_lookup_recall.py` 的 `RECALL_CASES`)。
|
||||
Reference in New Issue
Block a user