diff --git a/backend/app/look/service.py b/backend/app/look/service.py index 8a0229f6..769daeee 100644 --- a/backend/app/look/service.py +++ b/backend/app/look/service.py @@ -4,7 +4,9 @@ 1. need 没写：回「look 要写 need」，不算次数；次数用完：回「按已查看的结构完成」，不再查看。 2. targets 按类型分组（旧写法 tables:[{source, table}] 读进来换成 kind=table）。写了 name 的点名直取，不调挑选模型； 只写了 kind（和 source）的、或者整个没写 targets 的，按 need 在那一类里挑（没写时挑这个入口的第一类）。 - 3. 挑选失败或挑出来是空的：交回那一类的目录（status=index），写明原因，请模型点名再看。 + 3. 挑选失败或挑出来是空的：交回那一类的目录（status=index），写明原因，请模型点名再看。点名的一个都没认出来时， + 按 need 在那几个来源里挑（同只写类别），没认出来的照样写「没找到」；候选里一个能查看的都没有（库还没探查表 + 结构）时没有目录可交，status=failed，照实说原因。 4. 这一批对象这一轮都已经看过：回「已在上面」，不重新排一遍。 5. 各类给详情，合计受 LOOK_RESULT_CHARS 约束；记一条记录（id、by、need、status、objects、用时）。 @@ -36,6 +38,7 @@ INDEX_PROMPT = ("没能按需求挑出{what}：{reason}。下面是候选的{wha EMPTY_PROMPT = ("按这个需求没有挑出相关的{what}。这一步确实要用到{what}的话，从下面的目录里点名再 look；" "不需要就接着写：\n{index}") FAILED_PROMPT = "这一次查看出错了：{reason}。按已有的信息接着写" +UNAVAILABLE_PROMPT = "没有可查看的{what}：{reason}。按已有的信息接着写，不要猜{what}的名字和字段" #: 按需求挑出来是空的：写进 look 事件的 reason，界面上照原话显示 EMPTY_REASON = "按这个需求没有挑出相关的{what}" UNKNOWN_KIND_PROMPT = "这一轮不能查看这一类对象：{kinds}（能查看的：{allowed}）" @@ -184,21 +187,33 @@ class LookService: keys: set[str] = set() indexed: list[str] = [] empty: list[str] = [] + unavailable: list[str] = [] budget = LOOK_RESULT_CHARS for kind_name in dict.fromkeys([*named, *picks]): kind = self.kinds[kind_name] found, missing = await kind.resolve(named.get(kind_name, [])) if kind_name in named else ([], []) + fallback = bool(missing) and not found and kind_name not in picks + if fallback: + # 点名的一个都没认出来（常见的是拿说明里的中文词当表名）：按 need 在那几个来源里挑，和只写类别时一样， + # 不让模型为了拿到真名再 look 一次。没认出来的照样写进「没找到」，不拿挑出来的冒充点名的 + picks[kind_name] = list(dict.fromkeys(str(t["source"]) for t in missing if t.get("source"))) or None if kind_name in picks: within = picks[kind_name] or sources chosen, reason = await kind.pick(need, within) if reason is not None or not chosen: index = await kind.index(within) - indexed.append(INDEX_PROMPT.format(what=kind.label, reason=reason, index=index) if reason - else EMPTY_PROMPT.format(what=kind.label, index=index)) - if reason is not None: + if not index and reason is not None: + # 候选里一个能查看的都没有（库还没探查表结构）：没有目录可交，照实说原因，算查看没完成 + unavailable.append(UNAVAILABLE_PROMPT.format(what=kind.label, reason=reason)) record["reason"] = reason - else: - empty.append(kind.label) + elif index or not fallback: + # 点名的来源本身不在这一轮里（目录是空的）：「没找到数据源」已经说清楚了，不再补一段空目录 + indexed.append(INDEX_PROMPT.format(what=kind.label, reason=reason, index=index) if reason + else EMPTY_PROMPT.format(what=kind.label, index=index)) + if reason is not None: + record["reason"] = reason + else: + empty.append(kind.label) found += [o for o in chosen if kind.key(o) not in {kind.key(f) for f in found}] if not found and not missing: continue @@ -210,8 +225,10 @@ class LookService: if unknown: parts.append(UNKNOWN_KIND_PROMPT.format(kinds="、".join(dict.fromkeys(unknown)), allowed="、".join(self.kinds))) - parts += indexed - if indexed and not keys: + parts += indexed + unavailable + if unavailable and not indexed and not keys: + record["status"] = "failed" + elif indexed and not keys: # 一个对象都没给出详情、交回的是目录：挑选失败（有原因）或按需求没挑出来（返回空） record["status"] = "index" record.setdefault("reason", EMPTY_REASON.format(what="、".join(empty))) diff --git a/backend/app/look/tables.py b/backend/app/look/tables.py index 68c526c4..25fe1e5c 100644 --- a/backend/app/look/tables.py +++ b/backend/app/look/tables.py @@ -236,6 +236,12 @@ class TableKind: async def pick(self, need: str, sources: list[str] | None) -> tuple[list[TableRef], str | None]: candidates, _ = self._candidates(sources) + if not any(self.plans[i].tables for i in candidates): + # 候选库一张表都没探查到：不是「按需求没挑出来」，照实说为什么是空的（和概览行同一句） + bare = [self.plans[i] for i in candidates] if sources else self.plans + if bare: + return [], "；".join(f"数据源「{p.name}」{introspect.why_empty(p.source.schema_cache or {})}" + for p in bare) pool = [TableRef(i, key, "picked") for i in candidates for key in self.plans[i].tables] if len(pool) <= PICK_LIMIT: return pool, None diff --git a/backend/tests/test_copilot_table_pick.py b/backend/tests/test_copilot_table_pick.py index 1bdf3077..d8b39e55 100644 --- a/backend/tests/test_copilot_table_pick.py +++ b/backend/tests/test_copilot_table_pick.py @@ -458,3 +458,49 @@ async def test_the_old_tables_field_still_looks(scenic_source): result = await looks.look({"op": "look", "need": "门店", "tables": [{"source": scenic_source.name, "table": "stores"}]}, LookBudget()) assert result.record["objects"][0]["named"] == ["stores"] and "store_code TEXT" in result.text + + +async def test_names_that_match_nothing_fall_back_to_picking_by_need(monkeypatch, small_source): + """模型拿说明里的中文词点名（「订单」「客户」）一个都没认出来：同一次查看里按需求挑，不让它为了拿到真名再看一次。 + 小库不调挑选模型、全部给出；没认出来的照样写进「没找到」，不拿挑出来的冒充点名的。""" + fake = FakePicker(error=AssertionError("小库不该调挑选模型")) + _use_picker(monkeypatch, fake) + looks = await _service([small_source]) + result = await looks.look({"op": "look", "need": "每位客户的订单金额", "targets": [ + {"kind": "table", "source": small_source.name, "name": "订单"}, + {"kind": "table", "source": small_source.name, "name": "客户"}]}, LookBudget()) + record = result.record + assert record["status"] == "ok" and not fake.calls + (block,) = record["objects"] + assert block["named"] == [] and sorted(block["picked"]) == ["customers", "orders"] + assert block["missing"] == ["订单", "客户"] + assert "customer_id" in result.text and f"没找到：{small_source.name} 的 订单" in result.text + + +async def test_names_in_an_unknown_source_do_not_hand_back_an_empty_index(small_source): + looks = await _service([small_source]) + result = await looks.look({"op": "look", "need": "退款", "targets": [ + {"kind": "table", "source": "没有这个库", "name": "refunds"}]}, LookBudget()) + assert "没找到数据源「没有这个库」" in result.text + assert "目录" not in result.text and result.record["status"] == "ok" + + +async def test_a_source_never_introspected_says_so_instead_of_an_empty_pick(): + """库接上了但还没探查表结构：不是「按需求没挑出来」，照实说为什么是空的，查看算没完成（不交空目录）。""" + async with SessionLocal() as session: + row = DataSource(name=f"look_bare_{uuid.uuid4().hex[:8]}", kind="sqlite", database="/nonexistent.db", + description="门店、订单、客户") + session.add(row) + await session.commit() + await session.refresh(row) + session.expunge(row) + try: + looks = await _service([row]) + for line in ({"op": "look", "need": "各地区销售额"}, + {"op": "look", "need": "各地区销售额", "targets": [{"kind": "table", "name": "订单"}]}): + result = await looks.look(line, LookBudget()) + assert result.record["status"] == "failed", line + assert result.record["reason"] == f"数据源「{row.name}」尚未探查结构" + assert "尚未探查结构" in result.text and "目录" not in result.text + finally: + await drop_source(row.id)