| """训练和推理共用的一套 prompt，两边必须一模一样，不然分数都是假的。 |
| |
| 可以附上「部分字段的取值」：状态、等级、渠道这种分类字段，把库里实际有的值列出来， |
| 模型照着抄，不用猜是写 '已完成' 还是 'completed'。推理时用 values_from_db() 从真实库里取； |
| 不给也能用（训练时有两成样本不带），只是分类值要靠猜。 |
| """ |
|
|
| import re |
| import sqlite3 |
|
|
| SYSTEM = ( |
| "你是 SQLite 查询专家。用户会给你建表语句和一句中文问题，" |
| "你要写出一条能回答它的 SQLite 查询。只输出 SQL，不要解释。" |
| ) |
|
|
| USER = """建表语句： |
| {ddl} |
| |
| 问题：{question}""" |
|
|
| USER_WITH_VALUES = """建表语句： |
| {ddl} |
| |
| 部分字段的取值： |
| {values} |
| |
| 问题：{question}""" |
|
|
| DATE_LIKE = re.compile(r"^\d{4}-\d{2}(-\d{2})?") |
|
|
|
|
| def build(ddl: str, question: str, values: dict[str, list[str]] | None = None) -> list[dict]: |
| """values 形如 {"orders.status": ["已取消", "已完成"]}，按给的顺序列出。""" |
| if values: |
| lines = "\n".join(f"- {col}：{'、'.join(vs)}" for col, vs in values.items()) |
| user = USER_WITH_VALUES.format(ddl=ddl.strip(), values=lines, question=question.strip()) |
| else: |
| user = USER.format(ddl=ddl.strip(), question=question.strip()) |
| return [ |
| {"role": "system", "content": SYSTEM}, |
| {"role": "user", "content": user}, |
| ] |
|
|
|
|
| def values_from_db(conns, max_distinct: int = 12) -> dict[str, list[str]]: |
| """从库里找分类字段：文本列、不同值不超过 max_distinct 个、值有重复（不是名字、编号）、不是日期。 |
| conns 可以是同一个库的几份数据，取并集。返回 {"表.列": [排好序的值]}，表和列按建表顺序。""" |
| if isinstance(conns, sqlite3.Connection): |
| conns = [conns] |
| out = {} |
| tables = [t for (t,) in conns[0].execute( |
| "select name from sqlite_master where type='table' and name not like 'sqlite_%' order by rowid")] |
| for t in tables: |
| for col in conns[0].execute(f'pragma table_info("{t}")').fetchall(): |
| name, decl = col[1], (col[2] or "").upper() |
| if decl and "TEXT" not in decl and "CHAR" not in decl: |
| continue |
| vals, rows = set(), 0 |
| for c in conns: |
| vals.update(v for (v,) in c.execute( |
| f'select distinct "{name}" from "{t}" where "{name}" is not null limit {max_distinct + 1}')) |
| rows += c.execute(f'select count("{name}") from "{t}"').fetchone()[0] |
| if (not vals or len(vals) > max_distinct or len(vals) * 2 > rows |
| or not all(isinstance(v, str) and not DATE_LIKE.match(v) for v in vals)): |
| continue |
| out[f"{t}.{name}"] = sorted(vals) |
| return out |
|
|
