mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-20 01:02:00 +00:00
c3123e3818
对账卡的「口径」、召回卡的「旧版 / 新版」写成 nodes.<口径卡>.metrics[i].value,出处停在「来自口径卡指标」,对账运行、
召回运行一律降档出具(「输入没追到原始数据」);口径卡自己的指标明明追得到取数查询。
- engine/lineage:metric_copy 认出转抄一个指标的值的写法;traced() 对 via=metric 的输入看 upstream.traced;
MAX_COPY_DEPTH = 8。
- engine/nodes/metrics:转抄的输入记 upstream {metric, artifact, depth, traced}——那张卡的指标集落了工件、那个指标的输入
都追到了原始数据(它自己转抄的已经一层层追过,记在它的输入上)、层数没超上限才算追到;超了就算没追到(防环:循环里
抄自己上一轮的卡每一轮多一层)。那个指标不完整、所依据的查询没通过 SQL 检查的,转抄出来的同样不完整、存疑,出具照样
降档:信任按追到底的结果给,不多给。
- api/evidence:转抄的输入后面接一步 upstream(那张卡的那个指标:值、原式、代入式、它的输入、追没追到,取值按同一套遮罩),
转抄的转抄接着追(同一个指标只接一次、最多 8 层),它的输入落在查询快照上的照常给查询步骤;出处链里遮着的列把这几步
的输入也算上。reconcile.py、recall.py 生成的式子不用改。
- 前端证据面板:新的一节「转抄的口径卡指标」,输入来源里转抄的那一项写清追没追到、点一下跳到那一步,那一步的输入点一下
跳到口径的取数查询(文案进 lib/terms.ts)。check-ledger 加一段:对账结果里点「按口径」的数,一路点到口径的取数查询。
测试:tests/test_metric_copy_lineage.py(转抄、转抄的转抄完整出具;转抄没追到的照样停在那张卡;上游不完整的转抄也不完整;
层数上限;真跑对账、召回都完整出具,出处链接上口径卡那一步、给出口径的取数查询)。去掉 upstream 再跑,5 条失败。
159 lines
12 KiB
JavaScript
159 lines
12 KiB
JavaScript
// 口径账本的检查夹具(check-ledger、check-evidence、check-chat 共用):GET /api/runs/{id}/ledger 的答复形状照服务端
|
||
// app/ledger/view.py 的 run_view(labels.view_label、issue.Plan.action、reconcile.actions_of / reconciled_view),
|
||
// 挂在可点击证据的夹具(frontend/src/run/__tests__/evidence-doc.json,compose_doc 真跑出来的周报,只用通用名)上:
|
||
// s6「45,678.5元」来自受管口径「周报口径」v2,等价已证明;s10「1,234单」来自已发布口径「订单口径」v1,助手选用;
|
||
// s8「8.7%」相关口径「订单口径」v1(差在哪:少了一个筛选条件),可以对账;s12「37.0元」这次的推理;
|
||
// s23「2.35%」没有比对(比对超时)。
|
||
// 口径名、表名、数都是虚构的通用名。
|
||
|
||
/** 口径的几份(calibers) */
|
||
export const WEEKLY = { caliber: 'wf-weekly@2', workflow_id: 'wf-weekly', version: 2, level: 'governed', level_text: '受管',
|
||
workflow_name: '周报口径', graph_hash: 'aa11' }
|
||
export const ORDERS = { caliber: 'wf-orders@1', workflow_id: 'wf-orders', version: 1, level: 'published', level_text: '已发布',
|
||
workflow_name: '订单口径', graph_hash: 'bb22' }
|
||
|
||
const caliberLabel = (c, how, extra = {}) => ({
|
||
kind: 'caliber', how, caliber: c.caliber, workflow_id: c.workflow_id, version: c.version, level: c.level,
|
||
level_text: c.level_text, workflow_name: c.workflow_name, ...extra,
|
||
})
|
||
|
||
export const DIFF = { items: [{ aspect: 'filter', a: null, b: "status = '[已遮罩]'", text: '口径只保留状态为已支付的行,这次没有这个条件' }],
|
||
fallback: false, text_diff: [] }
|
||
|
||
/** 一份比对过的账本。whole:整题判断(缺省不等价);actions、issued、reconciled 可以换 */
|
||
export function ledgerOf(runId, { whole, actions, issued = [], reconciled = [], retro = false, labels } = {}) {
|
||
return {
|
||
run_id: runId, state: 'matched', note: '封存后比对,不在封存范围内', post_seal: true, artifact: 'cm-1', algo: 'v1',
|
||
trigger: retro ? 'on_demand' : 'sealed', retro, matched_at: '2026-10-15T01:00:00+00:00',
|
||
...(retro ? { retro_note: '按现在的口径回看' } : {}),
|
||
calibers: [WEEKLY, ORDERS],
|
||
queries: [
|
||
{ artifact: 'q-gmv', node_id: 'ag', source: 'shop', tables: ['orders'], status: 'proven',
|
||
candidates: [{ ...WEEKLY, node_id: 'fetch', status: 'proven', how: 'independent' }] },
|
||
{ artifact: 'q-orders', node_id: 'ag', source: 'shop', tables: ['orders'], status: 'related',
|
||
candidates: [{ ...ORDERS, node_id: 'fetch', status: 'related', comparable: true, via: 'heuristic', diff: DIFF,
|
||
binding: [{ field: 'start', value: '2026-09-01' }] }] },
|
||
],
|
||
reports: { write: { labels: labels ?? {
|
||
s6: caliberLabel(WEEKLY, 'independent', { binding: [{ field: 'week', value: '2026-W37' }], card: 'caliber', metric: 'gmv',
|
||
queries: ['q-gmv'] }),
|
||
s10: caliberLabel(ORDERS, 'composed', { binding: [{ field: 'store', masked: true, value: '已遮罩' }], card: 'caliber',
|
||
metric: 'orders', queries: ['q-gmv'] }),
|
||
s8: { kind: 'related', related: [{ caliber: ORDERS.caliber, workflow_id: ORDERS.workflow_id, version: 1, level: 'published',
|
||
level_text: '已发布', workflow_name: '订单口径', node_id: 'fetch' }], queries: ['q-orders'] },
|
||
s12: { kind: 'reasoning', query: 'q-gmv' },
|
||
s23: { kind: 'unknown', reason: '比对超过时限,没有做完', query: 'q-gmv' },
|
||
} } },
|
||
primary: 'write',
|
||
whole: whole ?? { equivalent: false, reason: '有的数是这次的推理', counts: { caliber: 2, related: 1, reasoning: 1, unknown: 1 } },
|
||
actions: actions ?? [RECONCILE],
|
||
issued, reconciled,
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 答案身上的召回标记(服务端 ledger/recall.answer_marks):周报口径升到 v3 重算过、数没变(同一份口径核对过两次,只写最近
|
||
* 那一版);目录里影响数字的两项被驳回、按新目录检查新出现 1 条错误级问题,另有一次出现了新码值(同归「数据目录」那一类,
|
||
* 项数相加);看的这个人手上有一条说的是这个答案的 recall 待办(订单口径升到 v2,数变了)
|
||
*/
|
||
export const RECALL = {
|
||
checked: [
|
||
{ workflow_id: 'wf-weekly', workflow_name: '周报口径', from_version: 1, to_version: 2, recall_run: 'run-rc0', changed: false,
|
||
at: '2026-10-14T01:00:00+00:00' },
|
||
{ workflow_id: 'wf-weekly', workflow_name: '周报口径', from_version: 2, to_version: 3, recall_run: 'run-rc1', changed: false,
|
||
at: '2026-10-15T01:00:00+00:00' },
|
||
],
|
||
basis: [
|
||
{ kind: 'catalog', items: 2, new_errors: 1, at: '2026-10-15T02:00:00+00:00' },
|
||
{ kind: 'values', items: 1, new_errors: 0, at: '2026-10-15T03:00:00+00:00' },
|
||
],
|
||
tasks: [{ id: 'tk-recall-1', reason: 'caliber', to_version: 2 }],
|
||
}
|
||
|
||
/** 看差多少(s8 那条查询、订单口径):参数是推断的,要先核对 */
|
||
export const RECONCILE = {
|
||
kind: 'reconcile', query: 'q-orders', caliber: ORDERS.caliber, workflow_id: ORDERS.workflow_id, version: 1, node_id: 'fetch',
|
||
workflow_name: '订单口径', level_text: '已发布', needs_check: true, check_reason: '参数按这次的 SQL 推断,请核对',
|
||
binding: [{ field: 'start', value: '2026-09-01' }], fallback: false,
|
||
fields: [{ name: 'start', required: true, has_default: false }, { name: 'store', required: false, has_default: true }],
|
||
}
|
||
|
||
/** 按口径正式出具(整题等价于周报口径):搭建者能一键,不自动的原因写在 auto_reason */
|
||
export const ISSUE = {
|
||
kind: 'issue', workflow_id: WEEKLY.workflow_id, version: 2, caliber: WEEKLY.caliber, workflow_name: '周报口径', how: 'independent',
|
||
level: 'governed', level_text: '受管', available: true, auto: false, auto_reason: '设置里关了自动按口径出具',
|
||
}
|
||
|
||
/** 整题等价时的标签(三个数都来自周报口径) */
|
||
export const WHOLE_LABELS = {
|
||
s6: caliberLabel(WEEKLY, 'independent', { binding: [{ field: 'week', value: '2026-W37' }], queries: ['q-gmv'] }),
|
||
s10: caliberLabel(WEEKLY, 'independent', { binding: [{ field: 'week', value: '2026-W37' }], queries: ['q-gmv'] }),
|
||
s12: caliberLabel(WEEKLY, 'independent', { binding: [{ field: 'week', value: '2026-W37' }], queries: ['q-gmv'] }),
|
||
}
|
||
export const WHOLE = { equivalent: true, caliber: WEEKLY.caliber, workflow_id: WEEKLY.workflow_id, version: 2, level: 'governed',
|
||
level_text: '受管', workflow_name: '周报口径', how: 'independent', binding: [{ field: 'week', value: '2026-W37' }],
|
||
missing: [], missing_required: [], counts: { caliber: 3, related: 0, reasoning: 0, unknown: 0 } }
|
||
|
||
/** 对过的一次账(单行:一组对照)。status 可以换 */
|
||
export const reconciledOf = (status = 'succeeded') => ({
|
||
run_id: 'run-rec-1', status, query: 'q-orders', caliber: ORDERS.caliber, workflow_id: ORDERS.workflow_id, version: 1,
|
||
inputs_from: 'checked', rows: null, card: 'reconcile',
|
||
groups: [{ label: '订单数', ours: 'ours_0', caliber: 'cal_0', diff: 'diff_0', rel: 'rel_0' }], created_at: '2026-10-15T02:00:00Z',
|
||
})
|
||
|
||
/** 对账卡上四个指标的展示值(/evidence/metrics/reconcile/{id} 的 rendered) */
|
||
export const RECONCILE_VALUES = { ours_0: '1,234单', cal_0: '1,180单', diff_0: '54单', rel_0: '4.58%' }
|
||
|
||
/** 对账运行的指标出处(形状同片段接口:chain、seal),给 EvidenceDrawer 画 */
|
||
export const metricDetail = (runId, metric) => ({
|
||
run_id: runId, node: 'reconcile', metric, rendered: RECONCILE_VALUES[metric] ?? '—', eid: `ev:metric:${metric}`,
|
||
chain: [{ step: 'metric', alias: 'm', metric, name: metric, value: 1, rendered: RECONCILE_VALUES[metric], node_id: 'reconcile',
|
||
caliber: '口径对账', version: 'v1', expression: 'cell(nodes.ours_q1, 0, \'orders\')', sealed: true, inputs: [] }],
|
||
seal: { sealed: true, ok: true, covered: true }, redacted: { columns: [] },
|
||
})
|
||
|
||
/**
|
||
* 对账卡上「按口径」那个数的出处(服务端 api/evidence._upstream_steps 的形状):它转抄口径卡的指标
|
||
* (nodes.card.metrics[0].value),出处跨卡接上口径卡「订单数」那一步,再到口径的取数查询(q-cal-fetch)
|
||
*/
|
||
export const copiedMetricDetail = (runId, metric) => ({
|
||
run_id: runId, node: 'reconcile', metric, rendered: RECONCILE_VALUES[metric] ?? '—', eid: `ev:metric:${metric}`,
|
||
chain: [
|
||
{ step: 'metric', alias: 'm', metric, name: '订单数 · 口径', value: 1180, rendered: RECONCILE_VALUES[metric], node_id: 'reconcile',
|
||
caliber: '口径对账', version: 'v1', expression: 'nodes.card.metrics[0].value', substituted: '1180', recompute_ok: true,
|
||
sealed: true, hash_ok: true, untraced: [],
|
||
inputs: [{ path: 'nodes.card.metrics[0].value', value: 1180, node_id: 'card', via: 'metric', field: 'metrics[0].value',
|
||
status: 'ok', upstream: { metric: 'orders', artifact: 'ms-card', depth: 1, traced: true } }] },
|
||
{ step: 'upstream', of: 'nodes.card.metrics[0].value', node_id: 'card', label: '口径卡', artifact: 'ms-card', metric: 'orders',
|
||
traced: true, depth: 1, hash_ok: true, sealed: true, name: '订单数', value: 1180, unit: '单', rendered: '1,180单',
|
||
caliber: '订单口径', version: 'v1', expression: "cell(nodes.fetch, 0, 'orders')", substituted: '1180', recompute_ok: true,
|
||
status: 'ok', untraced: [],
|
||
inputs: [{ step: 'input', path: "cell(nodes.fetch, 0, 'orders')", value: 1180, node_id: 'fetch', via: 'tool_cell',
|
||
status: 'verified', artifact: 'q-cal-fetch', locator: { row: 0, column: 'orders' }, eid: 'ev:cell:q-cal-fetch' }] },
|
||
{ step: 'query', alias: null, artifact: 'q-cal-fetch', node_id: 'fetch', tool: 'db_query__shop', source: 'shop',
|
||
sql: "SELECT COUNT(*) AS orders FROM orders WHERE status = 'paid'", columns: ['orders'], rows: [[1180]], row_offset: 0,
|
||
row_index: [0], total_rows: 1, highlight: { rows: [0], cols: ['orders'], cells: [[0, 'orders']] }, sealed: true, hash_ok: true },
|
||
],
|
||
seal: { sealed: true, ok: true, covered: true }, redacted: { columns: [] },
|
||
})
|
||
|
||
/** 认证队列:三类来源各一项 */
|
||
export const QUEUE = {
|
||
items: [
|
||
{ id: 'shape:aa11', kind: 'shape', source: 'shop', shape: 'SELECT SUM(amount) FROM orders WHERE paid_at >= ? AND status = ?',
|
||
runs: 4, conversations: 3, question: '上周销售额是多少',
|
||
turns: [{ run_id: 'run-s1', conversation_id: 'c-1', turn_id: 't-1', question: '上周销售额是多少' },
|
||
{ run_id: 'run-s2', conversation_id: 'c-2', turn_id: 't-2', question: '这周销售额是多少' }],
|
||
params: [{ index: 0, distinct: 3 }], constants: 1, related: [{ caliber: ORDERS.caliber, workflow_id: ORDERS.workflow_id, version: 1, level: 'published' }],
|
||
representative: { run_id: 'run-s1', query: 'q-1' }, peers: ['run-s2', 'run-s3'], frozen_into: [] },
|
||
{ id: 'cluster:bb22', kind: 'cluster', question: '各门店的退款率', count: 3,
|
||
turns: [{ run_id: 'run-c1', turn_id: 't-c1', question: '各门店的退款率' }, { run_id: 'run-c2', turn_id: 't-c2', question: '门店退款率怎么样' }],
|
||
frozen_into: [] },
|
||
{ id: 'reused:cc33', kind: 'reused', question: '本月新客数', uses: { exports: 2, app: false, extracted: false, canvas: true },
|
||
turns: [{ run_id: 'run-r1', conversation_id: 'c-3', turn_id: 't-r1', question: '本月新客数' }],
|
||
representative: { run_id: 'run-r1' }, frozen_into: [{ workflow_id: 'wf-new-old', name: '本月新客数(固化的口径草稿)', status: 'draft' }] },
|
||
],
|
||
considered: { runs: 120, matched: 100, questions: 90 },
|
||
threshold: { runs: 3, conversations: 2, exports: 2 },
|
||
}
|