Files
AgentLab/backend/tests/test_definition_batches.py
T
ThearchyHelios e15e94f93e 口径工作台:结构化语义导入、起草续跑恢复与本体关系编排
- 新增口径结构化语义定义、分批解析与 SQL 生成管道(semantic / semantic_import / semantic_sql)
- 支持起草断点续跑恢复、指标自动匹配对齐与存量口径语义迁移面板(MigrationPanel / SemanticPanel)
- 口径草稿与变更流程接入候选本体边提议、审阅确认与一键修改保护,同步更新可移植 SQLite 基线校验
2026-11-12 12:58:12 +08:00

250 lines
13 KiB
Python

"""Default-all implementation, bounded internal batches and durable resume without user batching."""
import json
import pytest
from app.api import ledger
from app.db.base import SessionLocal
from app.db.models import Chunk, Document
from tests.test_ledger_docdraft_api import client, datasource, fake_model, fake_trial # noqa: F401
async def imported(client, datasource, monkeypatch, count=31, conflicts=False):
source = await datasource()
text = '# 合成计数定义\n| 指标 | 公式 | 口径/说明 |\n|---|---|---|\n'
text += ''.join(f'| 计数指标{i} | COUNT(*) | 按日汇总计数,不取明细第一行 |\n' for i in range(count))
if conflicts:
text += '| 无来源指标 | 无数据源 | 不推算 |\n| 同名指标 | COUNT(*) | 范围甲 |\n| 同名指标 | SUM(x) | 范围乙 |\n'
async with SessionLocal() as session:
doc = Document(title='自动处理定义', content=text, status='ready', chunk_count=1)
session.add(doc)
await session.flush()
chunk = Chunk(document_id=doc.id, ordinal=1, content=text)
session.add(chunk)
await session.commit()
doc_id, chunk_id = doc.id, chunk.id
response = await client.post('/api/ledger/drafts/definitions', json={'documents': [doc_id]})
assert response.status_code == 201, response.text
draft_id = response.json()['draft_id']
view = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
async def enqueue(_):
pass
monkeypatch.setattr(ledger.draft_jobs, 'start', enqueue)
return draft_id, source, chunk_id, view['semantic_inventory']['items']
def reply(items, chunk):
return json.dumps({'sql': 'SELECT COUNT(*) AS count FROM orders', 'entry_fields': [],
'metrics': [{'id': f'm{index}', 'name': item['name'], 'source_definition_id': item['id'],
'expression': "cell(nodes.fetch, 0, 'count')",
'definition': {'text': item['name'], 'citations': [{'chunk': chunk, 'quote': '按日汇总计数,不取明细第一行'}]}}
for index, item in enumerate(items)], 'uncovered': [], 'open_questions': []}, ensure_ascii=False)
@pytest.mark.parametrize('count', [31, 67])
async def test_all_definitions_run_past_four_rounds_without_manual_batches(client, datasource, monkeypatch, fake_trial, count):
draft_id, source, chunk, items = await imported(client, datasource, monkeypatch, count=count, conflicts=True)
response = await client.post(f'/api/ledger/drafts/{draft_id}/implement', json={'datasource_id': source})
assert response.status_code == 202, response.text
eligible = items[:count]
model = fake_model(monkeypatch, *[reply(eligible[i:i+6], chunk) for i in range(0, count, 6)])
await ledger._extract_in_background(draft_id)
made = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
assert made['metric_count'] == count, made
assert len(model.calls) == (count + 5) // 6 and made['progress']['round'] == len(model.calls)
assert made['implementation_mode'] == 'all' and len(made['definition_issues']) == 3
assert made['implementation_progress'] == {'total': count, 'completed': count, 'pending': 0, 'blocked': 0, 'active_names': []}
assert made['status'] == 'questions_open' and not made['uncovered']
assert [p['definition_ids'] for p in made['prompts'] if p['purpose'] == 'plan'] == [
[i['id'] for i in eligible[n:n+6]] for n in range(0, count, 6)]
async def test_partial_selection_is_optional_and_can_exceed_six(client, datasource, monkeypatch, fake_trial):
draft_id, source, chunk, items = await imported(client, datasource, monkeypatch, count=15)
chosen = items[2:15]
response = await client.post(f'/api/ledger/drafts/{draft_id}/implement', json={
'datasource_id': source, 'selected': [i['id'] for i in chosen]})
assert response.status_code == 202, response.text
model = fake_model(monkeypatch, *[reply(chosen[n:n+6], chunk) for n in range(0, 13, 6)])
await ledger._extract_in_background(draft_id)
made = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
assert made['status'] == 'ready' and made['metric_count'] == 13 and len(model.calls) == 3
assert {m['source_definition_id'] for m in made['metrics']} == {i['id'] for i in chosen}
async def test_unresolved_batch_does_not_stop_other_definitions(client, datasource, monkeypatch, fake_trial):
draft_id, source, chunk, items = await imported(client, datasource, monkeypatch, count=13)
await client.post(f'/api/ledger/drafts/{draft_id}/implement', json={'datasource_id': source})
empty = json.dumps({'groups': [], 'open_questions': ['首组缺少字段映射'], 'uncovered': []})
model = fake_model(monkeypatch, empty, reply(items[6:12], chunk),
reply(items[12:] + items[:5], chunk), empty)
await ledger._extract_in_background(draft_id)
made = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
assert made['status'] == 'questions_open' and len(model.calls) == 4
assert made['metric_count'] == 12 and made['uncovered'] == [items[5]['name']]
assert made['implementation_progress']['blocked'] == 1 and made['implementation_progress']['pending'] == 0
assert any(items[5]['name'] in q for q in made['open_questions'])
async def test_failed_job_resumes_only_unfinished_definitions(client, datasource, monkeypatch, fake_trial):
from app.ledger import draftplan, docdraft
draft_id, source, chunk, items = await imported(client, datasource, monkeypatch, count=13)
await client.post(f'/api/ledger/drafts/{draft_id}/implement', json={'datasource_id': source})
fake_model(monkeypatch, reply(items[:6], chunk))
original = draftplan.plan_groups
calls = 0
async def fail_second(*args, **kwargs):
nonlocal calls
calls += 1
if calls == 2:
raise docdraft.DraftUnavailable('模型连接中断')
return await original(*args, **kwargs)
monkeypatch.setattr(draftplan, 'plan_groups', fail_second)
await ledger._extract_in_background(draft_id)
failed = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
assert failed['status'] == 'failed' and failed['metric_count'] == 6
async with SessionLocal() as session:
await ledger.draft_jobs.finish(session, draft_id, 'failed', '模型连接中断')
response = await client.post(f'/api/ledger/drafts/{draft_id}/retry')
assert response.status_code == 200, response.text
monkeypatch.setattr(draftplan, 'plan_groups', original)
model = fake_model(monkeypatch, reply(items[6:12], chunk), reply(items[12:], chunk))
await ledger._extract_in_background(draft_id)
made = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
assert made['status'] == 'ready' and made['metric_count'] == 13 and len(model.calls) == 2
assert made['implementation_progress']['completed'] == 13
assert len({m['source_definition_id'] for m in made['metrics']}) == 13
async def test_oee_overlapping_chunks_and_wide_prefixed_catalog_keep_visible_citations_and_diagnostics(
client, monkeypatch, fake_trial):
from pathlib import Path
import uuid
from app.db.models import DataSource
from app.ledger import authoring
fixture_path = Path(__file__).parent / 'fixtures' / 'oee_metric_definitions.md'
text = fixture_path.read_text(encoding='utf-8')
lines = text.splitlines(keepends=True)
# Build 16 overlapping chunks similar to production KB chunking (~16k chars total).
overlapping_chunks: list[str] = []
stride = max(1, len(lines) // 16)
for idx in range(16):
start = max(0, idx * stride - 6)
end = len(lines) if idx == 15 else min(len(lines), (idx + 2) * stride + 6)
overlapping_chunks.append(''.join(lines[start:end]))
tables: dict[str, dict] = {}
for idx in range(42):
tables[f'vi_ike_sap_wide_{idx:02d}'] = {
'columns': [{'name': f'sap_col_{c:02d}', 'type': 'VARCHAR2'} for c in range(60)]
}
tables['vi_ike_mes_bhgsj'] = {'columns': [{'name': 'scrap_qty', 'type': 'NUMBER'}]}
tables['vi_ike_mes_jzhmjxs'] = {'columns': [{'name': 'mjxs', 'type': 'NUMBER'}]}
tables['vi_ike_iot_dn_productionline'] = {'columns': [{'name': 'beat', 'type': 'NUMBER'}, {'name': 'line_code', 'type': 'VARCHAR2'}]}
tables['vi_ike_iot_kpi_info'] = {
'columns': [{'name': 'line_code', 'type': 'VARCHAR2'}, {'name': 'shift_date', 'type': 'DATE'}]
+ [{'name': f'attribute{c:02d}', 'type': 'NUMBER'} for c in range(1, 56)]
}
source_id = f'ds-{uuid.uuid4().hex[:10]}'
async with SessionLocal() as session:
src = DataSource(
id=source_id, name=f'bi_{uuid.uuid4().hex[:6]}', kind='oracle',
description='制造数据仓库', schema_cache={'schema': 'OWODS', 'tables': tables, 'total': len(tables)},
)
session.add(src)
doc = Document(title='指标公式总表.md', content=text, status='ready', chunk_count=len(overlapping_chunks))
session.add(doc)
await session.flush()
for ordinal, content in enumerate(overlapping_chunks, start=1):
session.add(Chunk(document_id=doc.id, ordinal=ordinal, content=content))
await session.commit()
doc_id = doc.id
focused = authoring._focus_tables(tables, text, source=src)
assert 'vi_ike_mes_bhgsj' in focused
assert 'vi_ike_iot_dn_productionline' in focused
assert 'vi_ike_iot_kpi_info' in focused
async def enqueue(_):
pass
monkeypatch.setattr(ledger.draft_jobs, 'start', enqueue)
created = await client.post('/api/ledger/drafts/definitions', json={'documents': [doc_id]})
assert created.status_code == 201, created.text
draft_id = created.json()['draft_id']
view = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
metric_items = [i for i in view['semantic_inventory']['items'] if i['kind'] == 'metric' and i['status'] != 'unavailable']
assert len(metric_items) >= 40
target = next(i for i in metric_items if i['name'] == '装配线 OEE')
chosen = [target, *[i for i in metric_items if i['id'] != target['id']][:11]]
# First batch implements `target` even if model omits citations (verified provenance fallback attaches it),
# while remaining 11 items stay unresolved with a model question (testing >10 unresolved non-spam behavior).
first_reply = json.dumps({
'groups': [{
'sql': 'SELECT AVG(attribute36) AS oee_assembly FROM OWODS.vi_ike_iot_kpi_info',
'entry_fields': [],
'metrics': [{
'id': 'm1', 'name': target['name'], 'source_definition_id': target['id'],
'expression': "cell(nodes.fetch, 0, 'oee_assembly')",
'definition': {'text': '装配线直接取源系统 OEE', 'citations': []},
}],
}],
'open_questions': ['其余指标需确认产线过滤编码与早晚班归属字段'],
'uncovered': [],
}, ensure_ascii=False)
later_reply = json.dumps({
'groups': [],
'open_questions': ['其余指标需确认产线过滤编码与早晚班归属字段'],
'uncovered': [],
}, ensure_ascii=False)
fake_model(monkeypatch, first_reply, later_reply)
started = await client.post(f'/api/ledger/drafts/{draft_id}/implement', json={
'datasource_id': source_id, 'selected': [i['id'] for i in chosen],
})
assert started.status_code == 202, started.text
await ledger._extract_in_background(draft_id)
made = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
assert made['metric_count'] == 1
assert made['metrics'][0]['source_definition_id'] == target['id']
assert made['metrics'][0]['definition']['citations']
for trace in made['prompts']:
assert trace['visible_chars'] and all(chars > 0 for chars in trace['visible_chars'].values())
assert target['id'] not in made['definition_diagnostics']
unresolved_ids = {i['id'] for i in chosen if i['id'] != target['id']}
assert set(made['definition_diagnostics']) == unresolved_ids
assert not any('尚未生成可用实现,请补充字段映射或确认口径' in q for q in made['open_questions'])
async def test_pure_validation_failure_without_model_questions_marks_draft_failed(
client, datasource, monkeypatch, fake_trial):
draft_id, source, _chunk, items = await imported(client, datasource, monkeypatch, count=2)
bad = json.dumps({
'groups': [{
'sql': 'SELECT COUNT(*) AS n FROM nonexistent_table',
'entry_fields': [],
'metrics': [{
'id': 'm1', 'name': items[0]['name'], 'source_definition_id': items[0]['id'],
'expression': "cell(nodes.fetch, 0, 'n')",
'definition': {'text': items[0]['name'], 'citations': []},
}],
}],
'open_questions': [],
'uncovered': [],
}, ensure_ascii=False)
fake_model(monkeypatch, bad)
await client.post(f'/api/ledger/drafts/{draft_id}/implement', json={'datasource_id': source})
await ledger._extract_in_background(draft_id)
made = (await client.get(f'/api/ledger/drafts/{draft_id}')).json()
assert made['status'] == 'failed'
assert 'nonexistent_table' in (made['error'] or '')
assert not made['open_questions']
assert made['definition_diagnostics'][items[0]['id']]['kind'] == 'validation_failed'