@hedeqiang/dsh-nl2sql 0.0.0-stage → 0.10.0

package/service/schema.js

@@ -0,0 +1,507 @@
+/**
+ * Nl2SqlService methods: schema.
+ * Mixed into the service prototype by index.js; `this` is the service.
+ */
+import { DIALECTS } from '../drivers.js'
+import { annotateSchema, columnMasked, filterSchema, normalizePolicy, policyActive } from '../policy.js'
+import { admitSql } from '../sql-guard.js'
+import { readFile, writeFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import { DIALECT_HINTS, UserError, classifyColumn, classifyTable, inferRelations, maskValue, newId, requireText, singular, upsert } from './shared.js'
+
+/** Character bigrams (whitespace removed, works for Chinese) plus ASCII words of 3+ characters. */
+export function textGrams(text) {
+  const s = String(text ?? '').toLowerCase().replace(/\s+/g, '')
+  const out = new Set()
+  for (let i = 0; i < s.length - 1; i++) out.add(s.slice(i, i + 2))
+  for (const w of String(text ?? '').toLowerCase().match(/[a-z0-9_]{3,}/g) ?? []) out.add(w)
+  return out
+}
+
+export const methods = {
+  // --------------------------------------------------------------- schema
+
+  async analyzeSchema(id, { refresh = false } = {}) {
+    const ds = this.requireDatasource(id)
+    if (!refresh) {
+      const cached = this.schemas.get(ds.id) ?? await this.readSchemaFile(ds.id)
+      if (cached) {
+        this.schemas.set(ds.id, cached)
+        return cached
+      }
+    }
+    const settings = this.settingsFor(ds)
+    const raw = await this.withDriver(ds, d => d.introspect(Math.max(settings.timeoutMs, 30000)))
+    const tables = raw.tables.map(t => ({ ...t, columns: [], indexes: [] }))
+    const byName = new Map(tables.map(t => [t.name, t]))
+    for (const c of raw.columns) {
+      const table = byName.get(c.table)
+      if (!table) continue
+      const { table: _t, ...col } = c
+      col.roles = classifyColumn(col, table)
+      table.columns.push(col)
+    }
+    const indexMap = new Map()
+    for (const i of raw.indexes) {
+      const k = `${i.table}\u0000${i.name}`
+      if (!indexMap.has(k)) indexMap.set(k, { table: i.table, name: i.name, unique: i.unique, columns: [] })
+      indexMap.get(k).columns.push(i.column)
+    }
+    for (const idx of indexMap.values()) byName.get(idx.table)?.indexes.push({ name: idx.name, unique: idx.unique, columns: idx.columns })
+    const fkRows = raw.foreignKeys
+    const versionRows = [{ version: raw.version }]
+    const declared = fkRows.map(r => ({ ...r, kind: 'declared', confidence: 1 }))
+    const relations = [...declared, ...inferRelations(tables, declared)]
+    for (const t of tables) t.role = classifyTable(t, relations)
+
+    // Enum value sampling, bounded and opt-out through settings.
+    if (settings.sampleMode !== 'off') {
+      const jobs = []
+      for (const t of tables) {
+        if (t.type === 'VIEW' || (t.rows ?? 0) > 2_000_000) continue
+        for (const c of t.columns) {
+          const cheapText = ds.type === 'sqlite' && c.roles.includes('text') && (t.rows ?? 0) <= 200000
+          if ((!c.roles.includes('enum') && !cheapText) || c.roles.includes('sensitive')) continue
+          jobs.push({ t, c })
+        }
+      }
+      const driver = await this.driver(ds)
+      for (const { t, c } of jobs.slice(0, 80)) {
+        try {
+          const rows = await driver.sampleColumn(t.name, c.name)
+          c.distinct = rows.length > 12 ? '>12' : rows.length
+          c.values = rows.slice(0, 12)
+        } catch {}
+      }
+    }
+    const schema = {
+      datasourceId: ds.id,
+      dialect: ds.type ?? 'mysql',
+      database: ds.database,
+      version: versionRows[0]?.version ?? null,
+      analyzedAt: Date.now(),
+      tables,
+      relations,
+    }
+    const previous = this.schemas.get(ds.id) ?? await this.readSchemaFile(ds.id)
+    if (previous && settings.schemaWatch !== false) {
+      try { this.recordSchemaChange(ds, previous, schema) } catch (error) { console.warn('nl2sql: schema diff failed: %s', error.message) }
+    }
+    this.schemas.set(ds.id, schema)
+    ds.schemaAnalyzedAt = schema.analyzedAt
+    if (raw.version && !ds.serverVersion) ds.serverVersion = raw.version
+    await writeFile(join(this.dir, `schema-${ds.id}.json`), JSON.stringify(schema), { mode: 0o600 })
+    await this.save()
+    return schema
+  },
+
+  async readSchemaFile(id) {
+    try {
+      return JSON.parse(await readFile(join(this.dir, `schema-${id}.json`), 'utf8'))
+    } catch {
+      return undefined
+    }
+  },
+
+  policyFor(ds) {
+    return normalizePolicy(ds?.policy ?? null)
+  },
+
+  /** Schema restricted to what the access policy exposes (model, tools, knowledge drafts). */
+  async visibleSchema(ds, opts) {
+    return filterSchema(await this.analyzeSchema(ds.id, opts), this.policyFor(ds))
+  },
+
+  /** Full schema annotated with policy flags (page only). */
+  async annotatedSchema(id, opts) {
+    const ds = this.requireDatasource(id)
+    return annotateSchema(await this.analyzeSchema(ds.id, opts), this.policyFor(ds))
+  },
+
+  /**
+   * Rank tables by relevance to a question: character-bigram / word overlap
+   * with the table name, comment, note, columns, plus glossary / metric /
+   * verified-example knowledge that references the table, plus a prior for
+   * fact tables and large tables. Returns [{ table, score }] best first.
+   */
+  rankTables(schema, ds, question) {
+    // Ranking uses non-ASCII bigrams and whole ASCII words (singular), so short Latin fragments don't match everything.
+    const grams = (text) => {
+      const out = new Set()
+      for (const x of textGrams(text)) if (x.length > 2 || /[^\x00-\x7f]/.test(x)) out.add(x.length > 3 ? singular(x) : x)
+      return out
+    }
+    const q = grams(question)
+    const k = this.knowledgeFor(ds.id)
+    const cover = (grams) => {
+      if (!q.size) return 0
+      let hit = 0
+      for (const x of q) if (grams.has(x)) hit++
+      return hit / q.size
+    }
+    const mentions = (text, t) => {
+      const short = t.name.split('.').pop().toLowerCase()
+      return new RegExp(`(^|[^\\w])${short.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}([^\\w]|$)`).test(String(text ?? '').toLowerCase())
+    }
+    const found = question ? this.searchKnowledge(ds.id, question, 20) : { examples: [], glossary: [], metrics: [] }
+    const prior = { fact: 0.15, dimension: 0.08, view: 0.04, other: 0, log: -0.05, config: -0.05 }
+    return schema.tables.map(t => {
+      const label = `${t.name.replace(/[_.]+/g, ' ')} ${t.comment ?? ''} ${k.notes[t.name] ?? ''}`
+      const cols = t.columns.map(c => `${c.name.replace(/_+/g, ' ')} ${c.comment ?? ''} ${k.notes[`${t.name}.${c.name}`] ?? ''}`).join(' ')
+      let score = 2 * cover(grams(label)) + 0.7 * cover(grams(cols))
+      for (const m of found.metrics) if (m.table && (m.table === t.name || mentions(m.table, t))) score += 1.5 * m.score + 0.3
+      for (const g of found.glossary) if (mentions(`${g.sql} ${g.definition}`, t)) score += g.score + 0.2
+      for (const e of found.examples) if (mentions(e.sql, t)) score += e.score + 0.2
+      score += prior[t.role] ?? 0
+      score += 0.05 * Math.min(1, Math.log10((t.rows ?? 0) + 1) / 8)
+      return { table: t, score }
+    }).sort((a, b) => b.score - a.score || (b.table.rows ?? 0) - (a.table.rows ?? 0))
+  },
+
+  /** Compact schema text for the model, honoring the sampling setting. */
+  schemaForModel(schema, ds, tableNames, { question } = {}) {
+    const settings = this.settingsFor(ds)
+    const lines = []
+    if (!tableNames || tableNames.length === 0) {
+      const dialect = DIALECTS[ds.type ?? 'mysql']
+      lines.push(`${ds.type === 'elasticsearch' ? 'index pattern' : 'database'}: ${schema.database} (${dialect.label} ${schema.version ?? '?'}; write SQL in the ${dialect.sqlName} dialect), ${schema.tables.length} ${ds.type === 'elasticsearch' ? 'indices' : 'tables'}, analyzed ${new Date(schema.analyzedAt).toISOString()}`)
+      lines.push(DIALECT_HINTS[ds.type ?? 'mysql'])
+      if (policyActive(this.policyFor(ds))) lines.push('access control: only the tables and columns listed here may be queried; hidden columns are omitted and system catalogs are blocked. Do not try to work around it.')
+      const max = settings.schemaOverviewMax ?? 80
+      const large = schema.tables.length > max
+      let detailed = schema.tables
+      if (large) {
+        const top = Math.max(1, settings.schemaOverviewTop ?? 15)
+        const ranked = this.rankTables(schema, ds, question)
+        detailed = ranked.slice(0, top).map(r => r.table)
+        lines.push(`large schema (${schema.tables.length} tables > ${max}): every table is listed by name only; ${detailed.length} tables ${question ? 'most relevant to the question' : 'ranked by role and size (pass `question` for relevance)'} are described in full below. Call nl2sql_schema with \`tables\` to read any other table before using it.`)
+        lines.push('all tables (table | role | ~rows):')
+        for (const t of schema.tables) lines.push(`${t.name} | ${t.role} | ${t.rows ?? '?'}`)
+        lines.push(`most relevant tables${question ? ` for: ${String(question).slice(0, 200)}` : ''}:`)
+      }
+      lines.push('table | role | ~rows | comment | columns(name:type[roles])')
+      for (const t of detailed) {
+        const cols = t.columns.map(c => `${c.name}:${c.dataType}${c.roles.length ? `[${c.roles.join(',')}]` : ''}`).join(', ')
+        const tnote = this.knowledgeFor(ds.id).notes[t.name]
+        lines.push(`${t.name} | ${t.role} | ${t.rows ?? '?'} | ${[t.comment, tnote].filter(Boolean).join('; ') || '-'} | ${cols}`)
+      }
+      const shown = new Set(detailed.map(t => t.name))
+      const rels = large ? schema.relations.filter(r => shown.has(r.from.table) || shown.has(r.to.table)) : schema.relations
+      if (rels.length) {
+        lines.push('relations (from -> to, kind, confidence):')
+        for (const r of rels) lines.push(`${r.from.table}.${r.from.column} -> ${r.to.table}.${r.to.column} (${r.kind}, ${r.confidence})`)
+      }
+      const k = this.knowledgeFor(ds.id)
+      if (ds.description) lines.push(`business context: ${ds.description}`)
+      if (k.glossary.length) {
+        lines.push('glossary (business term => meaning / SQL hint):')
+        for (const g of k.glossary) lines.push(`- ${g.term}${g.synonyms?.length ? ` (${g.synonyms.join('/')})` : ''} => ${g.definition}${g.sql ? ` [SQL: ${g.sql}]` : ''}`)
+      }
+      if (k.metrics.length) {
+        lines.push('metrics (use these definitions exactly when the user asks for them):')
+        for (const m of k.metrics) lines.push(`- ${m.name}${m.unit ? ` (${m.unit})` : ''}: ${m.expression}${m.table ? ` FROM ${m.table}` : ''}${m.filter ? ` WHERE ${m.filter}` : ''}${m.timeColumn ? `; time column ${m.timeColumn}` : ''}${m.description ? ` — ${m.description}` : ''}`)
+      }
+      if (k.examples.length) lines.push(`${k.examples.length} verified question→SQL examples exist; call nl2sql_knowledge action=search with the user's question to reuse them.`)
+      lines.push('Call nl2sql_schema with `tables` for full column comments, indexes and enum values.')
+      return lines.join('\n')
+    }
+    const notes = this.knowledgeFor(ds.id).notes
+    const wanted = new Set(tableNames.map(n => n.toLowerCase()))
+    const missing = tableNames.filter(n => !schema.tables.some(t => t.name.toLowerCase() === n.toLowerCase()))
+    for (const t of schema.tables) {
+      if (!wanted.has(t.name.toLowerCase())) continue
+      lines.push(`## ${t.name} (${t.type.toLowerCase()}, role=${t.role}, ~${t.rows ?? '?'} rows)${t.comment ? ` — ${t.comment}` : ''}`)
+      if (notes[t.name]) lines.push(`note: ${notes[t.name]}`)
+      for (const c of t.columns) {
+        const note = notes[`${t.name}.${c.name}`]
+        let line = `- ${c.name} ${c.columnType}${c.nullable ? '' : ' NOT NULL'}${c.key ? ` ${c.key}` : ''}${c.roles.length ? ` [${c.roles.join(',')}]` : ''}${c.comment ? ` — ${c.comment}` : ''}${note ? ` (note: ${note})` : ''}`
+        if (c.values && settings.sampleMode !== 'off' && !(settings.maskSensitive && c.roles.includes('sensitive'))) {
+          line += ` values: ${c.values.map(v => `${JSON.stringify(v.value)}(${v.count})`).join(', ')}${c.distinct === '>12' ? ', …' : ''}`
+        }
+        lines.push(line)
+      }
+      if (t.indexes.length) lines.push(`indexes: ${t.indexes.map(i => `${i.unique ? 'UNIQUE ' : ''}${i.name}(${i.columns.join(',')})`).join('; ')}`)
+      const rel = schema.relations.filter(r => r.from.table === t.name || r.to.table === t.name)
+      if (rel.length) lines.push(`relations: ${rel.map(r => `${r.from.table}.${r.from.column}->${r.to.table}.${r.to.column}(${r.kind})`).join('; ')}`)
+      lines.push('')
+    }
+    if (missing.length) lines.push(`unknown tables: ${missing.join(', ')}`)
+    return lines.join('\n')
+  },
+
+  async sampleRows(ds, table, limit, actor) {
+    const settings = this.settingsFor(ds)
+    if (actor && actor.kind !== 'owner' && (this.rowRulesFor(actor, ds).length || actor.role === 'viewer')) throw new UserError('nl2sql/row-policy', '当前成员不能查看原始样例数据行')
+    if (settings.sampleMode !== 'all') throw new UserError('nl2sql/privacy', '当前隐私设置不允许把样例数据行发送给模型（需将“表结构抽样”设为“全部抽样”）。')
+    const schema = await this.visibleSchema(ds)
+    const t = schema.tables.find(x => x.name.toLowerCase() === table.toLowerCase())
+    if (!t) throw new UserError('nl2sql/unknown-table', `表不存在或无权访问：${table}`)
+    const res = await this.withDriver(ds, d => d.query(d.previewSql(t.name, Math.min(Math.max(limit, 1), 20)), { timeoutMs: settings.timeoutMs, maxRows: 20 }))
+    const keep = res.columns.map((c, i) => (t.columns.some(x => x.name === c.name) ? i : -1)).filter(i => i >= 0)
+    const columns = keep.map(i => res.columns[i].name)
+    const masked = res.rows.map(r => keep.map(i => r[i]).map((cell, i) => {
+      const col = t.columns.find(c => c.name === columns[i])
+      const forced = columnMasked(this.policyFor(ds), t.name, columns[i]) || col?.policyMasked
+      return forced || (settings.maskSensitive && col?.roles.includes('sensitive')) ? maskValue(columns[i], cell) : cell
+    }))
+    return { columns, rows: masked }
+  },
+
+  // ------------------------------------------------------ schema changes
+
+  /** Compare two analyzed schemas: added/removed tables, added/removed/changed columns. */
+  diffSchemas(before, after) {
+    const tb = new Map(before.tables.map(t => [t.name, t]))
+    const ta = new Map(after.tables.map(t => [t.name, t]))
+    const changes = []
+    for (const [name, t] of ta) if (!tb.has(name)) changes.push({ kind: 'table-added', table: name, detail: t.comment || '' })
+    for (const [name] of tb) if (!ta.has(name)) changes.push({ kind: 'table-removed', table: name })
+    for (const [name, t] of ta) {
+      const old = tb.get(name)
+      if (!old) continue
+      const oc = new Map(old.columns.map(c => [c.name, c]))
+      const nc = new Map(t.columns.map(c => [c.name, c]))
+      for (const [cn, c] of nc) if (!oc.has(cn)) changes.push({ kind: 'column-added', table: name, column: cn, detail: c.columnType })
+      for (const [cn, c] of oc) if (!nc.has(cn)) changes.push({ kind: 'column-removed', table: name, column: cn, detail: c.columnType })
+      for (const [cn, c] of nc) {
+        const o = oc.get(cn)
+        if (o && o.columnType !== c.columnType) changes.push({ kind: 'column-type', table: name, column: cn, detail: `${o.columnType} → ${c.columnType}` })
+      }
+    }
+    return changes
+  },
+
+  /** Things that reference removed/changed objects: widgets, examples, metrics, notes, alerts, policy. */
+  schemaImpacts(ds, changes) {
+    const removedTables = new Set(changes.filter(c => c.kind === 'table-removed').map(c => c.table.toLowerCase()))
+    const removedCols = changes.filter(c => c.kind === 'column-removed' || c.kind === 'column-type')
+    const hits = []
+    const touches = (sql) => {
+      if (!sql) return []
+      const low = String(sql).toLowerCase()
+      const out = []
+      for (const t of removedTables) if (new RegExp(`(^|[^\\w])${t.split('.').pop().replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}([^\\w]|$)`).test(low)) out.push(`表 ${t}`)
+      for (const c of removedCols) if (new RegExp(`(^|[^\\w])${c.column.toLowerCase().replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}([^\\w]|$)`).test(low) && low.includes(c.table.split('.').pop().toLowerCase())) out.push(`${c.table}.${c.column}`)
+      return out
+    }
+    for (const d of this.state.dashboards) for (const w of d.widgets) if (w.datasourceId === ds.id) { const t = touches(w.sql); if (t.length) hits.push({ kind: 'widget', id: w.id, name: `${d.name} / ${w.title}`, refs: t }) }
+    const k = this.knowledgeFor(ds.id)
+    for (const e of k.examples) { const t = touches(e.sql); if (t.length) hits.push({ kind: 'example', id: e.id, name: e.question, refs: t }) }
+    for (const m of k.metrics) { const t = touches(`${m.expression} ${m.table} ${m.filter} ${m.timeColumn}`); if (t.length) hits.push({ kind: 'metric', id: m.id, name: m.name, refs: t }) }
+    for (const key of Object.keys(k.notes)) { const [tb, col] = key.split('.'); if (removedTables.has(tb.toLowerCase()) || removedCols.some(c => c.table === tb && c.column === col)) hits.push({ kind: 'note', id: key, name: key, refs: [key] }) }
+    for (const a of this.state.alerts ?? []) if (a.datasourceId === ds.id) { const t = touches(a.sql); if (t.length) hits.push({ kind: 'alert', id: a.id, name: a.name, refs: t }) }
+    return hits
+  },
+
+  recordSchemaChange(ds, before, after) {
+    const changes = this.diffSchemas(before, after)
+    if (!changes.length) return undefined
+    const entry = { id: newId('sc'), at: Date.now(), datasourceId: ds.id, datasourceName: ds.name, changes: changes.slice(0, 500), impacts: this.schemaImpacts(ds, changes), status: 'open' }
+    this.records.put('schemaChange', entry)
+    try { this.store?.putSnapshot({ id: newId('ss'), kind: 'schema', ref: ds.id, createdAt: before.analyzedAt ?? Date.now(), body: { analyzedAt: before.analyzedAt, tables: before.tables.map(t => ({ name: t.name, columns: t.columns.map(c => ({ name: c.name, columnType: c.columnType })) })) } }) } catch {}
+    this.onSchemaChange?.(entry)
+    if (this.state.settings.schemaNotifyChannel) {
+      const lines = entry.changes.slice(0, 15).map(c => `- ${{ 'table-added': '新增表', 'table-removed': '删除表', 'column-added': '新增字段', 'column-removed': '删除字段', 'column-type': '类型变更' }[c.kind]}：${c.table}${c.column ? `.${c.column}` : ''}${c.detail ? `（${c.detail}）` : ''}`)
+      const imp = entry.impacts.length ? `\n\n**受影响（${entry.impacts.length}）**\n${entry.impacts.slice(0, 10).map(i => `- ${i.name}`).join('\n')}` : ''
+      this.sendToChannel({ channelId: this.state.settings.schemaNotifyChannel, title: `表结构变更：${ds.name}`, text: `${lines.join('\n')}${imp}`, source: 'schema' }).catch(() => {})
+    }
+    return entry
+  },
+
+  listSchemaChanges({ datasourceId, status } = {}) {
+    return this.records.list('schemaChange', { ref: datasourceId, status, limit: 200 })
+  },
+
+  async ackSchemaChange({ id, status = 'acked' }) {
+    const c = this.records.get('schemaChange', id)
+    if (!c) throw new UserError('nl2sql/invalid', '变更记录不存在')
+    c.status = status === 'open' ? 'open' : 'acked'
+    c.ackedAt = Date.now()
+    this.records.put('schemaChange', c)
+    await this.save()
+    return c
+  },
+
+  /** Re-analyze every data source (scheduled schema watch). */
+  async checkAllSchemas() {
+    const out = []
+    for (const ds of this.state.datasources) {
+      const before = this.records.count('schemaChange', { ref: ds.id })
+      try {
+        await this.analyzeSchema(ds.id, { refresh: true })
+        out.push({ datasourceId: ds.id, name: ds.name, ok: true, changes: this.records.count('schemaChange', { ref: ds.id }) - before })
+      } catch (error) {
+        out.push({ datasourceId: ds.id, name: ds.name, ok: false, error: error.message })
+      }
+    }
+    return out
+  },
+
+  // --------------------------------------------------- knowledge import/export
+
+  exportKnowledge({ datasourceId, kinds } = {}) {
+    const ids = datasourceId ? [this.requireDatasource(datasourceId).id] : this.state.datasources.map(d => d.id)
+    const want = new Set(kinds?.length ? kinds : ['glossary', 'metric', 'note', 'example'])
+    return {
+      format: 'dsh-nl2sql-knowledge',
+      version: 1,
+      exportedAt: new Date().toISOString(),
+      datasources: ids.map(id => {
+        const ds = this.findDatasource(id)
+        const k = this.knowledgeFor(id)
+        return {
+          name: ds.name, type: ds.type ?? 'mysql', database: ds.database,
+          ...(want.has('glossary') ? { glossary: k.glossary.map(({ id: _i, ...g }) => g) } : {}),
+          ...(want.has('metric') ? { metrics: k.metrics.map(({ id: _i, ...m }) => m) } : {}),
+          ...(want.has('note') ? { notes: k.notes } : {}),
+          ...(want.has('example') ? { examples: k.examples.map(({ id: _i, ...e }) => e) } : {}),
+        }
+      }),
+    }
+  },
+
+  /**
+   * Import knowledge exported by exportKnowledge. Each source block maps onto
+   * `targetDatasourceId` (single block) or the data source with the same name.
+   * mode: merge (default; same term / metric name / question replaced) or replace.
+   */
+  async importKnowledge({ data, targetDatasourceId, mode = 'merge', dryRun = false }) {
+    const doc = typeof data === 'string' ? JSON.parse(data) : data
+    if (doc?.format !== 'dsh-nl2sql-knowledge' || !Array.isArray(doc.datasources)) throw new UserError('nl2sql/invalid', '不是有效的业务知识导出文件（format 应为 dsh-nl2sql-knowledge）')
+    const report = []
+    for (const block of doc.datasources) {
+      const ds = targetDatasourceId && doc.datasources.length === 1 ? this.requireDatasource(targetDatasourceId) : this.findDatasource(block.name)
+      if (!ds) { report.push({ source: block.name, skipped: '找不到同名数据源，请指定目标数据源' }); continue }
+      const k = mode === 'replace' ? { glossary: [], metrics: [], notes: {}, examples: [] } : this.knowledgeFor(ds.id)
+      const count = { glossary: 0, metrics: 0, notes: 0, examples: 0, invalid: 0 }
+      for (const g of block.glossary ?? []) {
+        if (!g?.term || !g?.definition) { count.invalid++; continue }
+        const prev = k.glossary.find(x => x.term === g.term)
+        k.glossary = upsert(k.glossary, { id: prev?.id ?? newId('g'), term: String(g.term).slice(0, 60), synonyms: (g.synonyms ?? []).map(String).slice(0, 10), definition: String(g.definition).slice(0, 500), sql: String(g.sql ?? '').slice(0, 500) })
+        count.glossary++
+      }
+      for (const m of block.metrics ?? []) {
+        if (!m?.name || !m?.expression) { count.invalid++; continue }
+        const prev = k.metrics.find(x => x.name === m.name)
+        k.metrics = upsert(k.metrics, { id: prev?.id ?? newId('m'), name: String(m.name).slice(0, 60), expression: String(m.expression).slice(0, 500), table: String(m.table ?? '').slice(0, 64), filter: String(m.filter ?? '').slice(0, 500), timeColumn: String(m.timeColumn ?? '').slice(0, 64), unit: String(m.unit ?? '').slice(0, 20), description: String(m.description ?? '').slice(0, 500), dimensions: Array.isArray(m.dimensions) ? m.dimensions.map(String).slice(0, 20) : [] })
+        count.metrics++
+      }
+      for (const [key, text] of Object.entries(block.notes ?? {})) { if (typeof text === 'string' && text.trim()) { k.notes[String(key).slice(0, 200)] = text.slice(0, 1000); count.notes++ } }
+      for (const e of block.examples ?? []) {
+        try { admitSql(e?.sql ?? '', { dialect: ds.type ?? 'mysql' }) } catch { count.invalid++; continue }
+        if (!e?.question) { count.invalid++; continue }
+        const prev = k.examples.find(x => x.question === e.question)
+        k.examples = upsert(k.examples, { id: prev?.id ?? newId('e'), question: String(e.question).slice(0, 500), sql: e.sql, note: String(e.note ?? '').slice(0, 500), createdAt: e.createdAt ?? Date.now() }).slice(-500)
+        count.examples++
+      }
+      if (!dryRun) this.state.knowledge[ds.id] = k
+      report.push({ source: block.name, target: ds.name, ...count })
+    }
+    if (!dryRun) await this.save()
+    return { dryRun, mode, report }
+  },
+
+  // ------------------------------------------------------- semantic layer
+
+  knowledgeFor(id) {
+    const k = this.state.knowledge[id] ?? {}
+    return { glossary: k.glossary ?? [], metrics: k.metrics ?? [], notes: k.notes ?? {}, examples: k.examples ?? [] }
+  },
+
+  getKnowledge(datasourceId) {
+    this.requireDatasource(datasourceId)
+    return this.knowledgeFor(datasourceId)
+  },
+
+  async saveKnowledgeItem({ datasourceId, kind, item }) {
+    const ds = this.requireDatasource(datasourceId)
+    const k = this.knowledgeFor(ds.id)
+    if (kind === 'note') {
+      const key = requireText(item?.key, 'key', 200)
+      const text = String(item?.text ?? '').trim().slice(0, 1000)
+      if (text) k.notes[key] = text
+      else delete k.notes[key]
+    } else if (kind === 'glossary') {
+      const entry = {
+        id: item?.id || newId('g'),
+        term: requireText(item?.term, 'term', 60),
+        synonyms: (Array.isArray(item?.synonyms) ? item.synonyms : String(item?.synonyms ?? '').split(/[,，、]/)).map(x => String(x).trim()).filter(Boolean).slice(0, 10),
+        definition: requireText(item?.definition, 'definition', 500),
+        sql: String(item?.sql ?? '').trim().slice(0, 500),
+      }
+      k.glossary = upsert(k.glossary, entry)
+    } else if (kind === 'metric') {
+      const entry = {
+        id: item?.id || newId('m'),
+        name: requireText(item?.name, 'name', 60),
+        expression: requireText(item?.expression, 'expression', 500),
+        table: String(item?.table ?? '').trim().slice(0, 64),
+        filter: String(item?.filter ?? '').trim().slice(0, 500),
+        timeColumn: String(item?.timeColumn ?? '').trim().slice(0, 64),
+        unit: String(item?.unit ?? '').trim().slice(0, 20),
+        description: String(item?.description ?? '').trim().slice(0, 500),
+        dimensions: (Array.isArray(item?.dimensions) ? item.dimensions : String(item?.dimensions ?? '').split(/[,，、\s]+/)).map(x => String(x).trim()).filter(Boolean).slice(0, 20),
+        synonyms: (Array.isArray(item?.synonyms) ? item.synonyms : String(item?.synonyms ?? '').split(/[,，、]/)).map(x => String(x).trim()).filter(Boolean).slice(0, 10),
+      }
+      k.metrics = upsert(k.metrics, entry)
+    } else if (kind === 'example') {
+      const sql = requireText(item?.sql, 'sql', 20000)
+      admitSql(sql, { dialect: ds.type ?? 'mysql' })
+      const entry = { id: item?.id || newId('e'), question: requireText(item?.question, 'question', 500), sql, note: String(item?.note ?? '').slice(0, 500), createdAt: item?.createdAt ?? Date.now() }
+      k.examples = upsert(k.examples, entry).slice(-500)
+    } else {
+      throw new UserError('nl2sql/invalid', `unknown knowledge kind ${kind}`)
+    }
+    this.state.knowledge[ds.id] = k
+    await this.save()
+    return k
+  },
+
+  async deleteKnowledgeItem({ datasourceId, kind, id }) {
+    const ds = this.requireDatasource(datasourceId)
+    const k = this.knowledgeFor(ds.id)
+    if (kind === 'note') delete k.notes[id]
+    else if (kind === 'glossary') k.glossary = k.glossary.filter(x => x.id !== id)
+    else if (kind === 'metric') k.metrics = k.metrics.filter(x => x.id !== id)
+    else if (kind === 'example') k.examples = k.examples.filter(x => x.id !== id)
+    this.state.knowledge[ds.id] = k
+    await this.save()
+    return k
+  },
+
+  /** Rank stored examples and glossary/metric entries against a question (character bigram overlap, works for Chinese). */
+  searchKnowledge(datasourceId, question, limit = 5) {
+    const k = this.knowledgeFor(datasourceId)
+    const grams = textGrams
+    const q = grams(question)
+    const score = (text) => {
+      const g = grams(text)
+      let hit = 0
+      for (const x of g) if (q.has(x)) hit++
+      return g.size === 0 ? 0 : hit / Math.sqrt(g.size * Math.max(q.size, 1))
+    }
+    const rank = (list, textOf) => list.map(x => ({ x, s: score(textOf(x)) })).filter(r => r.s > 0.08).sort((a, b) => b.s - a.s).slice(0, limit).map(r => ({ ...r.x, score: Number(r.s.toFixed(3)) }))
+    return {
+      examples: rank(k.examples, e => e.question),
+      glossary: rank(k.glossary, g => `${g.term} ${(g.synonyms ?? []).join(' ')} ${g.definition}`),
+      metrics: rank(k.metrics, m => `${m.name} ${m.description}`),
+    }
+  },
+
+  /** Draft table/column notes from comments and enum samples, for the user to review. */
+  async draftNotes(datasourceId) {
+    const ds = this.requireDatasource(datasourceId)
+    const schema = await this.visibleSchema(ds)
+    const k = this.knowledgeFor(ds.id)
+    const drafts = []
+    for (const t of schema.tables) {
+      for (const c of t.columns) {
+        const key = `${t.name}.${c.name}`
+        if (k.notes[key] || !c.values?.length || c.values.length > 12) continue
+        if (!/(status|state|type|kind|level|状态|类型)/i.test(c.name + c.comment)) continue
+        drafts.push({ key, text: `取值：${c.values.map(v => `${v.value}`).join(' / ')}${c.comment ? `（${c.comment}）` : ''}，请补充每个取值的业务含义` })
+      }
+    }
+    return drafts.slice(0, 60)
+  },
+}