pgkiln 0.69.0 → 0.70.0

package/src/builder/i18n/zh-tw.ts

@@ -1065,6 +1065,40 @@ export const zhTW: BuilderMessages = {
   'fts.gin_yes': "GIN 索引",
   'fts.gin_no': "無 GIN 索引",
   'fts.added': "已為 {table} 新增全文檢索：欄位 {column}，索引 {index}。",
+  // (sprint 136, AP-1) Table Properties › Add vector column
+  'vec.add': "新增向量欄位",
+  'vec.title': "{table} 的向量欄位",
+  'vec.intro': "新增一個嵌入欄位 vector(n)（n：向量提供者的維度數）、一個用於該距離的 HNSW 索引，以及在來源欄位變更時重新嵌入該列的觸發程序。之後，搜尋類型為 vector 或 hybrid 的搜尋組態，或 AI 代理的 Retrieve 工具，就能依含意找到資料列。現有的資料列會立即進入嵌入佇列。",
+  'vec.existing': "目前的向量欄位",
+  'vec.provider': "向量提供者",
+  'vec.provider_help': "此工作區應用程式的共用元件 › 向量提供者。搜尋時使用同一個提供者。",
+  'vec.keep': "保持最新",
+  'vec.keep_help': "函式提供者可以在觸發程序中立即設定嵌入。AI 服務由伺服器呼叫：觸發程序將資料列放入佇列，伺服器在提供者的每分鐘請求數與應用程式的每日 AI 上限內，批次嵌入佇列中的資料列。",
+  'vec.mode_set': "由觸發程序立即設定",
+  'vec.mode_queue': "放入伺服器的佇列",
+  'vec.sources': "來源欄位",
+  'vec.sources_help': "要嵌入的文字：這些欄位，每行一個。其中任一欄位變更時會重新嵌入該列。",
+  'vec.distance': "距離",
+  'vec.distance_help': "大多數嵌入模型使用 cosine；搜尋組態使用相同的距離。",
+  'vec.column': "向量欄位",
+  'vec.no_key': "此表格需要單一欄位的主索引鍵：佇列以此找出資料列。",
+  'vec.pick_provider': "請選擇一個向量提供者。",
+  'vec.no_providers': "此工作區的應用程式還沒有向量提供者：請在共用元件 › 向量提供者中新增。",
+  'vec.missing': "此資料庫未安裝 pgvector 擴充功能，因此暫時無法新增向量欄位。",
+  'vec.missing_create': "資料庫伺服器有 pgvector {version}：請在工作區公用程式 › 向量提供者中建立該擴充功能。",
+  'vec.missing_admin': "資料庫伺服器有 pgvector {version}：由管理員建立該擴充功能（工作區公用程式 › 向量提供者）。",
+  'vec.missing_package': "資料庫伺服器沒有 pgvector 套件：請先安裝（例如 apt install postgresql-17-pgvector，或 Docker 映像檔 pgvector/pgvector），再建立該擴充功能。",
+  'vec.utilities': "工作區公用程式 › 向量提供者",
+  'vec.kept_by': "由 {provider} 依 {columns} 維護",
+  'vec.not_kept': "沒有觸發程序讓它保持最新",
+  'vec.progress': "{column} 已嵌入的資料列",
+  'vec.counts': "{total} 列中已嵌入 {done} 列，{queued} 列在佇列中",
+  'vec.failing': "{n} 列失敗",
+  'vec.reembed_all': "所有資料列，而不只是沒有嵌入的資料列",
+  'vec.backfill': "回填",
+  'vec.refresh': "重新整理",
+  'vec.backfill_queued': "已將 {column} 的 {n} 列放入嵌入佇列。",
+  'vec.added': "已為 {table} 新增向量欄位 {column}；{n} 列已進入嵌入佇列。",
   // ------------------------------------------------------------ Shared Components (sprint 124)
   'comp.nav_entry': '導覽項目',
   'comp.nav_entry.plural': '導覽選單',

package/src/builder/object-actions.ts

@@ -33,6 +33,7 @@ export function objectMenu(schema: string, o: MenuObj): Raw {
       ${table ? item(`${BASE}/sql/objects/security?${new URLSearchParams({ o: o.ref })}`, 'shield', 'Security…') : ''}
       ${table ? item(`${BASE}/sql/load`, 'upload', 'Load data') : ''}
       ${table ? item(`${BASE}/sql/objects/fulltext?${new URLSearchParams({ o: o.ref })}`, 'search', 'Add full-text search…') : ''}
+      ${table ? item(`${BASE}/sql/objects/vector?${new URLSearchParams({ o: o.ref })}`, 'search', 'Add vector column…') : ''}
       ${table ? html`${item(q('index'), 'list', 'Add index')}${item(q('trigger'), 'bolt', 'Add trigger')}${item(q('policy'), 'shield', 'Add policy')}` : ''}
       ${RELATIONS.has(o.type) ? item(`${BASE}/sql/rest/object?${new URLSearchParams({ schema, name: o.name })}`, 'globe', 'REST › Enable / Disable') : ''}
       ${item(q('ddl'), 'code', 'Quick DDL to the worksheet')}

package/src/builder/shared.ts

@@ -76,7 +76,7 @@ export function sharedGroups(appId: number): { key: string; title: string; icon:
     { key: 'data', title: 'Data Sources', icon: 'database', items: [t('rest_source', 'REST Data Sources'), t('web_credential', 'Web Credentials'), t('rest_module', 'RESTful Services'), t('data_load_def', 'Data Load Definitions')] },
     { key: 'workflows', title: 'Workflows and Automations', icon: 'activity', items: [t('workflow_definition', 'Workflows'), t('task_definition', 'Task Definitions'), t('automation', 'Automations')] },
     { key: 'globalization', title: 'Globalization', icon: 'globe', items: [page('globalization', 'Translations and Text Messages'), t('dynamic_translation', 'Dynamic Translations'), settings('globalization', 'Globalization Attributes')] },
-    { key: 'ai', title: 'Generative AI', icon: 'lightbulb', items: [page('ai', 'AI Services'), page('ai-agents', 'AI Agents')] },
+    { key: 'ai', title: 'Generative AI', icon: 'lightbulb', items: [page('ai', 'AI Services'), page('ai-agents', 'AI Agents'), page('vector-providers', 'Vector Providers')] },
   ];
 }
 

package/src/builder/sql.ts

@@ -20,6 +20,7 @@ import { createTemplate, objectBrowserRoutes } from './object-browser.ts';
 import { objectActionRoutes } from './object-actions.ts';
 import { tableEditorRoutes } from './table-editor.ts';
 import { fulltextRoutes } from './fulltext.ts';
+import { vectorColumnRoutes } from './vector-column.ts';
 import { dataSecurityRoutes } from './data-security.ts';
 import { dataModelerRoutes } from './data-modeler.ts';
 import { autoRestBuilderRoutes } from './autorest.ts';
@@ -137,6 +138,8 @@ export async function sqlRoutes(app: FastifyInstance) {
   await app.register(tableEditorRoutes);
   // (sprint 135, AP-2) Table Properties › Add full-text search
   await app.register(fulltextRoutes);
+  // (sprint 136, AP-1) Table Properties › Add vector column
+  await app.register(vectorColumnRoutes);
   await app.register(dataSecurityRoutes);
   // (sprint 98) the Data Modeler: relational diagrams, their DDL and data dictionary
   await app.register(dataModelerRoutes);

package/src/builder/table-editor.ts

@@ -143,6 +143,7 @@ function form(s: Session, def: TableDef, opts: { editing: TableDef | null; schem
       <button class="btn btn-hot" name="mode" value="apply">${icon('check')} ${tx('te.apply')}</button>
       <button class="btn" name="mode" value="worksheet">${icon('code')} ${tx('te.open_ws')}</button>
       ${editing ? html`<a class="btn" href="${BASE}/sql/objects/fulltext?${new URLSearchParams({ o: opts.ref })}">${icon('search')} ${tx('fts.add')}…</a>` : ''}
+      ${editing ? html`<a class="btn" href="${BASE}/sql/objects/vector?${new URLSearchParams({ o: opts.ref })}">${icon('search')} ${tx('vec.add')}…</a>` : ''}
       <a class="btn" href="${editing ? `${BASE}/sql/objects?${new URLSearchParams({ schema: def.schema, k: 'table', o: opts.ref })}` : `${BASE}/sql/objects?${new URLSearchParams({ schema: def.schema })}`}">${tx('common.cancel')}</a>
     </div></form>`;
 }

package/src/builder/vector-column.ts

@@ -0,0 +1,229 @@
+import type { FastifyInstance, FastifyReply } from 'fastify';
+import pg from 'pg';
+import { HNSW_MAX_DIMENSIONS, vectorProviderProblems, type VectorProvider } from '../ai/embeddings.ts';
+import { joinStatements, recordChanges } from '../changelog.ts';
+import { owner } from '../db.ts';
+import { html, type Raw } from '../html.ts';
+import { icon } from '../icons.ts';
+import { clientIp, logActivity } from '../session.ts';
+import { readTable } from '../tabledef.ts';
+import { backfill, DISTANCES, isDistance, kickEmbeddings, pgvectorStatus, queueState, vectorTriggers, type Distance } from '../vector.ts';
+import { clearCompletions } from './code-editor.ts';
+import { tr } from './i18n.ts';
+import { isAdmin } from './locks.ts';
+import { SYSTEM_SCHEMAS } from './object-browser.ts';
+import { back, BASE, csrf, developer, flash, send, shell, workshopTabs, type Body, type Req } from './ui.ts';
+import { appAllowed, currentWorkspace, inWorkspaceSql, usableSchemas, workshopDb } from './workspaces.ts';
+
+// (sprint 136, AP-1) SQL Workshop › Table Properties › Add vector column: an embedding column vector(n) (n: the
+// chosen vector provider's dimensions), an HNSW index for the distance (vector_cosine_ops …), and a trigger that
+// keeps it current when the source columns change:
+//  - a function provider: BEFORE INSERT OR UPDATE OF … → meta.embedding_set (the embedding at once);
+//  - an AI service (or a function, if chosen): AFTER INSERT OR UPDATE OF … → meta.embedding_enqueue (the row's key
+//    queued; the server embeds queued rows in batches, src/vector.ts).
+// Preview DDL shows the statements; Apply runs them in one transaction with the workshop's connection and records
+// them in the Change Log; Open in the worksheet puts them there. The page also lists the table's vector columns
+// with their progress (rows embedded, queued, failing) and a Backfill button that queues the rows without an
+// embedding (or all rows). Names come from the catalog or are checked as plain identifiers before they reach DDL.
+
+type Db = Awaited<ReturnType<typeof workshopDb>>;
+const NAME = /^[a-z_][a-z0-9_]{0,62}$/;
+const ident = (n: string) => (/^[a-z_][a-z0-9_$]*$/.test(n) ? n : pg.escapeIdentifier(n));
+
+interface Col { name: string; type: string; texty: boolean; dims: number | null; vector: boolean }
+
+async function columnsOf(db: Db, oid: number): Promise<Col[]> {
+  return (await db.query<Col>(
+    `select a.attname as name, format_type(a.atttypid, a.atttypmod) as type, t.typcategory = 'S' and a.attgenerated = '' as texty,
+            t.typname = 'vector' as vector, case when t.typname = 'vector' and a.atttypmod > 0 then a.atttypmod end as dims
+       from pg_attribute a join pg_type t on t.oid = a.atttypid
+      where a.attrelid = $1 and a.attnum > 0 and not a.attisdropped order by a.attnum`, [oid])).rows;
+}
+
+/** The single-column primary key of a table (the queue keys rows by it), or null. */
+async function keyOf(db: Db, oid: number) {
+  return (await db.one<{ k: string }>(
+    `select a.attname as k from pg_index i join pg_attribute a on a.attrelid = i.indrelid and a.attnum = i.indkey[0]
+      where i.indrelid = $1 and i.indisprimary and i.indnatts = 1`, [oid]))?.k ?? null;
+}
+
+/** The workspace's applications' vector providers, as options "appId:NAME". */
+async function providerChoices(s: Parameters<typeof currentWorkspace>[0]) {
+  const apps = (await owner.query<{ id: number; alias: string; vector_providers: VectorProvider[] }>(
+    `select id, alias, vector_providers from meta.app a where ${inWorkspaceSql('a', 1)} and jsonb_array_length(vector_providers) > 0 order by alias`, [currentWorkspace(s)])).rows;
+  return apps.flatMap((a) => a.vector_providers.filter((p) => !vectorProviderProblems(p).length).map((p) => ({ key: `${a.id}:${p.name}`, appId: a.id, alias: a.alias, p })));
+}
+
+/** The statements: the column, its HNSW index (up to 2,000 dimensions) and the trigger that keeps it. */
+export function vectorDdl(o: { table: string; vectorType: string; column: string; index: string; trigger: string; dims: number; distance: Distance; mode: 'set' | 'queue'; appId: number; provider: string; key: string; sources: string[] }) {
+  const args = [String(o.appId), o.provider, o.column, o.key, ...o.sources].map((x) => pg.escapeLiteral(x)).join(', ');
+  return [
+    `alter table ${o.table}\n  add column ${ident(o.column)} ${o.vectorType}(${o.dims})`,
+    ...(o.dims <= HNSW_MAX_DIMENSIONS ? [`create index ${ident(o.index)} on ${o.table} using hnsw (${ident(o.column)} ${DISTANCES[o.distance].opclass})`] : []),
+    o.mode === 'set'
+      ? `create trigger ${ident(o.trigger)}\n  before insert or update of ${o.sources.map(ident).join(', ')} on ${o.table}\n  for each row execute function meta.embedding_set(${args})`
+      : `create trigger ${ident(o.trigger)}\n  after insert or update of ${o.sources.map(ident).join(', ')} on ${o.table}\n  for each row execute function meta.embedding_enqueue(${args})`,
+  ];
+}
+
+async function page(req: Req, reply: FastifyReply) {
+  const s = await developer(req, reply);
+  if (!s) return;
+  const tx = tr(s);
+  const db = await workshopDb(s);
+  const isPost = req.method === 'POST';
+  const b: Body = (isPost ? req.body : req.query) ?? {};
+  const ref = String(b.o ?? '');
+  const t = ref ? await readTable(db, ref) : null;
+  const schemas = (await usableSchemas(db)).filter((n) => !SYSTEM_SCHEMAS.includes(n));
+  if (!t || !schemas.includes(t.schema)) return reply.code(404).send('Not found');
+  const table = `${ident(t.schema)}.${ident(t.name)}`;
+  const self = `${BASE}/sql/objects/vector?${new URLSearchParams({ o: table })}`;
+  const oid = (await db.one<{ oid: number }>('select $1::regclass::oid as oid', [table]))!.oid;
+  const st = await pgvectorStatus(db);
+  const cols = await columnsOf(db, oid);
+  const texty = cols.filter((c) => c.texty);
+  const key = await keyOf(db, oid);
+  const choices = await providerChoices(s);
+  const triggers = st.installed ? await vectorTriggers(db, table) : [];
+
+  // Backfill: queue the rows of a vector column that have no embedding (or all of them)
+  if (isPost && b.mode === 'backfill') {
+    const trg = triggers.find((x) => x.vectorColumn === String(b.column ?? ''));
+    if (!trg || !(await appAllowed(s, trg.appId))) return reply.code(404).send('Not found');
+    const n = await backfill(db, trg, b.all === 'on');
+    kickEmbeddings();
+    flash(s, tx('vec.backfill_queued', { n: String(n), column: trg.vectorColumn }));
+    return back(reply, s, self);
+  }
+
+  const pick = choices.find((c) => c.key === String(b.provider ?? '')) ?? (isPost ? undefined : choices[0]);
+  const column = isPost ? String(b.column ?? '').trim() : cols.some((c) => c.name === 'embedding') ? '' : 'embedding';
+  const base = t.name.replace(/[^a-z0-9_]+/gi, '_').toLowerCase().slice(0, 40);
+  const index = isPost ? String(b.index ?? '').trim() : `${base}_${column || 'embedding'}_idx`;
+  const distance: Distance = isDistance(b.distance) ? b.distance : 'cosine';
+  const mode: 'set' | 'queue' = pick?.p.type === 'function' && (isPost ? b.keep === 'set' : true) ? 'set' : 'queue';
+  const picked = new Set<string>();
+  if (isPost) texty.forEach((c, i) => b[`c.${i}.use`] === 'on' && b[`c.${i}.name`] === c.name && picked.add(c.name));
+  else texty.forEach((c) => picked.add(c.name));
+
+  const errors: string[] = [];
+  let ddl: string[] = [];
+  let output: Raw | '' = '';
+  if (isPost && st.installed) {
+    if (!pick) errors.push(tx('vec.pick_provider'));
+    if (!picked.size) errors.push(tx('fts.pick'));
+    if (!key) errors.push(tx('vec.no_key'));
+    for (const n of [column, index]) if (!NAME.test(n)) errors.push(tx('fts.bad_name', { name: n }));
+    if (NAME.test(column) && cols.some((c) => c.name === column)) errors.push(tx('fts.exists', { name: column }));
+    if (NAME.test(index) && (await db.one('select 1 as x from pg_class where relnamespace = $1::regnamespace and relname = $2', [ident(t.schema), index]))) errors.push(tx('fts.exists', { name: index }));
+    const trigger = `${column}_embed`.slice(0, 63);
+    if (!errors.length && pick && key) {
+      ddl = vectorDdl({ table, vectorType: `${ident(st.schema!)}.vector`, column, index, trigger, dims: pick.p.dimensions, distance, mode, appId: pick.appId, provider: pick.p.name, key, sources: [...picked] });
+    }
+    if (ddl.length && b.mode === 'worksheet') {
+      s.state.__SQLDRAFT = `${ddl.join(';\n\n')};`;
+      flash(s, tx('te.review'));
+      return back(reply, s, `${BASE}/sql/worksheet`);
+    }
+    if (ddl.length && b.mode === 'apply') {
+      const c = await db.pool.connect();
+      let failed = '';
+      let at = '';
+      try {
+        await c.query('begin');
+        for (const x of ddl) {
+          at = x;
+          await c.query(x);
+        }
+        await c.query('commit');
+      } catch (e) {
+        failed = (e as Error).message;
+        await c.query('rollback').catch(() => {});
+      } finally {
+        c.release(true);
+      }
+      await logActivity({ username: s.username, event: 'sql_command', ip: clientIp(req), detail: ddl.join(';\n').slice(0, 2000) });
+      clearCompletions();
+      if (!failed) {
+        await recordChanges({ workspaceId: currentWorkspace(s), author: s.username ?? '?', source: 'table', defaultSchema: t.schema, statements: [joinStatements(ddl)] });
+        // the rows there are now: queued at once (the server embeds them; a function provider too)
+        const trg = (await vectorTriggers(db, table)).find((x) => x.vectorColumn === column);
+        const n = trg ? await backfill(db, trg, true) : 0;
+        kickEmbeddings();
+        flash(s, tx('vec.added', { table, column, n: String(n) }));
+        return back(reply, s, self);
+      }
+      output = html`<div class="alert alert-error" role="alert"><p><strong>${tx('te.nothing_changed')}</strong>: ${tx('te.rolled_back')}</p><p>${failed}</p><pre class="source" data-code-view="sql" tabindex="0" aria-label="${tx('te.failed_stmt')}">${at}</pre></div>`;
+    }
+  }
+
+  // the table's vector columns now, with their progress
+  const existing: Raw[] = [];
+  for (const c of cols.filter((x) => x.vector)) {
+    const trg = triggers.find((x) => x.vectorColumn === c.name);
+    const counts = (await db.one<{ total: number; done: number }>(`select count(*)::int as total, count(${ident(c.name)})::int as done from ${table}`))!;
+    const q = await queueState(oid, c.name);
+    const pct = counts.total ? Math.round((counts.done / counts.total) * 100) : 100;
+    existing.push(html`<li class="vec-column"><p><code>${c.name}</code> <code>${c.type}</code>
+        ${trg ? html`· ${tx('vec.kept_by', { provider: trg.provider, columns: trg.sourceColumns.join(', ') })} (${trg.mode === 'set' ? tx('vec.mode_set') : tx('vec.mode_queue')})` : html`· <span class="muted">${tx('vec.not_kept')}</span>`}</p>
+      <p><progress max="100" value="${pct}" aria-label="${tx('vec.progress', { column: c.name })}">${pct}%</progress>
+        ${tx('vec.counts', { done: String(counts.done), total: String(counts.total), queued: String(q.queued) })}${q.failing ? html` · <span class="tag tag-error">${tx('vec.failing', { n: String(q.failing) })}</span>` : ''}</p>
+      ${q.message ? html`<p class="muted">${q.message}</p>` : ''}
+      ${trg ? html`<form method="post" action="${BASE}/sql/objects/vector" class="buttons">${csrf(s)}<input type="hidden" name="o" value="${table}"><input type="hidden" name="column" value="${c.name}">
+        <label class="check"><input type="checkbox" name="all"> ${tx('vec.reembed_all')}</label>
+        <button class="btn" name="mode" value="backfill">${icon('refresh')} ${tx('vec.backfill')}</button>
+        <a class="btn" href="${self}">${tx('vec.refresh')}</a></form>` : ''}</li>`);
+  }
+
+  const title = tx('vec.title', { table });
+  const sel = (a: string, v: string) => (a === v ? ' selected' : '');
+  const admin = await isAdmin(s.username);
+  const form = !st.installed
+    ? html`<div class="alert alert-info" role="status"><p>${tx('vec.missing')}</p>${st.available ? html`<p>${tx(admin && st.canCreate ? 'vec.missing_create' : 'vec.missing_admin', { version: st.availableVersion ?? '' })}</p>` : html`<p>${tx('vec.missing_package')}</p>`}
+        <p><a class="btn" href="${BASE}/vector-providers">${icon('search')} ${tx('vec.utilities')}</a></p></div>`
+    : !choices.length
+      ? html`<p class="alert alert-info">${tx('vec.no_providers')}</p>`
+      : !texty.length
+        ? html`<p class="alert alert-info">${tx('fts.no_text')}</p>`
+        : html`<form method="post" action="${BASE}/sql/objects/vector" class="fts-form">${csrf(s)}
+      <input type="hidden" name="o" value="${table}">
+      <div class="form-grid">
+        <div class="field"><label class="label" for="vec-provider">${tx('vec.provider')}</label><select id="vec-provider" name="provider">${choices.map((c) => html`<option value="${c.key}"${sel(c.key, pick?.key ?? '')}>${c.alias} · ${c.p.name} (${c.p.type === 'ai_service' ? c.p.model : c.p.function}, ${c.p.dimensions})</option>`)}</select>
+          <small class="help">${tx('vec.provider_help')}</small></div>
+        <div class="field"><label class="label" for="vec-keep">${tx('vec.keep')}</label><select id="vec-keep" name="keep"><option value="set"${sel(mode, 'set')}>${tx('vec.mode_set')}</option><option value="queue"${sel(mode, 'queue')}>${tx('vec.mode_queue')}</option></select>
+          <small class="help">${tx('vec.keep_help')}</small></div>
+      </div>
+      <fieldset class="prop-group"><legend>${tx('vec.sources')}</legend>
+        <div class="table-wrap"><table class="report"><thead><tr><th>${tx('fts.include')}</th><th>${tx('te.name')}</th><th>${tx('te.type')}</th></tr></thead><tbody>
+        ${texty.map((c, i) => html`<tr><td class="te-c"><input type="hidden" name="c.${i}.name" value="${c.name}"><input type="checkbox" id="vec-c${i}" name="c.${i}.use"${picked.has(c.name) ? ' checked' : ''}></td>
+          <td><label for="vec-c${i}">${c.name}</label></td><td><code>${c.type}</code></td></tr>`)}
+        </tbody></table></div>
+        <small class="help">${tx('vec.sources_help')}</small></fieldset>
+      <div class="form-grid u-mt1">
+        <div class="field"><label class="label" for="vec-distance">${tx('vec.distance')}</label><select id="vec-distance" name="distance">${(Object.keys(DISTANCES) as Distance[]).map((d) => html`<option${sel(d, distance)}>${d}</option>`)}</select><small class="help">${tx('vec.distance_help')}</small></div>
+        <div class="field"><label class="label" for="vec-column">${tx('vec.column')}</label><input id="vec-column" name="column" value="${column}" required autocomplete="off" spellcheck="false" class="code"></div>
+        <div class="field"><label class="label" for="vec-index">${tx('fts.index')}</label><input id="vec-index" name="index" value="${index}" required autocomplete="off" spellcheck="false" class="code"></div>
+      </div>
+      ${key ? '' : html`<p class="alert alert-error" role="alert">${tx('vec.no_key')}</p>`}
+      ${ddl.length ? html`<h2>DDL</h2><pre class="source te-ddl" data-code-view="sql" tabindex="0" aria-label="DDL">${ddl.map((x) => `${x};`).join('\n\n')}</pre>` : ''}
+      <div class="buttons te-buttons">
+        <button class="btn" name="mode" value="preview">${icon('eye')} ${tx('te.preview')}</button>
+        <button class="btn btn-hot" name="mode" value="apply">${icon('check')} ${tx('te.apply')}</button>
+        <button class="btn" name="mode" value="worksheet">${icon('code')} ${tx('te.open_ws')}</button>
+        <a class="btn" href="${BASE}/sql/objects/table?${new URLSearchParams({ o: table })}">${tx('common.cancel')}</a>
+      </div></form>`;
+  const main = html`<h1 class="u-mb1">${title}</h1>${workshopTabs('objects', s)}
+    <p class="muted">${tx('vec.intro')}</p>
+    ${existing.length ? html`<h2>${tx('vec.existing')}</h2><ul class="vec-columns">${existing}</ul>` : ''}
+    ${errors.length ? html`<div class="alert alert-error" role="alert"><ul>${errors.map((e) => html`<li>${e}</li>`)}</ul></div>` : ''}
+    ${output}
+    <h2>${tx('vec.add')}</h2>
+    ${form}`;
+  return send(reply, s, shell(s, title, [['SQL Workshop', `${BASE}/sql`], [tx('ob.title'), `${BASE}/sql/objects`], [tx('te.title_table', { table }), `${BASE}/sql/objects/table?${new URLSearchParams({ o: table })}`], [tx('vec.add')]], main, 'sql'));
+}
+
+export async function vectorColumnRoutes(app: FastifyInstance) {
+  app.get(`${BASE}/sql/objects/vector`, page);
+  app.post(`${BASE}/sql/objects/vector`, page);
+}

package/src/builder/vector-providers.ts

@@ -0,0 +1,286 @@
+import type { FastifyInstance, FastifyReply } from 'fastify';
+import { embedTexts, vectorProviderProblems, DEFAULT_BATCH, DEFAULT_PER_MINUTE, HNSW_MAX_DIMENSIONS, type VectorProvider } from '../ai/embeddings.ts';
+import { providerLabel } from '../ai/service.ts';
+import { isAiError } from '../ai/types.ts';
+import { owner } from '../db.ts';
+import { html, type Raw } from '../html.ts';
+import { icon } from '../icons.ts';
+import type { Session } from '../session.ts';
+import { clientIp, logActivity } from '../session.ts';
+import { pgvectorStatus, vectorTriggers, type PgvectorStatus } from '../vector.ts';
+import { devDbForApp } from '../workspace-db.ts';
+import { appOr404 } from './forms.ts';
+import { isAdmin } from './locks.ts';
+import { appHeader, back, BASE, csrf, developer, flash, input, region, select, send, shell, type Req } from './ui.ts';
+import { currentWorkspace, inWorkspaceSql } from './workspaces.ts';
+
+// (sprint 136, AP-1; APEX: Workspace Utilities and Shared Components → Vector Providers)
+//
+// Shared Components → Generative AI → Vector providers: the application's providers, a JSON list in meta.app
+// (vector_providers, migration 136) exported with the application. A provider is an AI service with an embedding
+// model (OpenAI text-embedding-3-*, Gemini, any OpenAI-compatible endpoint such as Ollama) or a SQL / PL/pgSQL
+// function returning vector; its dimensions are shown and checked (Test embeds a sample text). The key of an AI
+// service stays with the service (Workspace utilities → AI services).
+//
+// Workspace utilities → Vector providers: whether pgvector is installed (administrators may create it), every
+// provider of the workspace's applications and the vector columns that providers keep current.
+
+type Body = Record<string, string | string[] | undefined>;
+const one = (v: string | string[] | undefined) => (Array.isArray(v) ? v[0] : v) ?? '';
+
+const crumbs = (a: { id: number; name: string }, last?: string): [string, string?][] => [
+  ['App Builder', BASE], [a.name, `${BASE}/apps/${a.id}`], ['Shared Components', `${BASE}/apps/${a.id}/shared`],
+  ...(last ? [['Vector providers', `${BASE}/apps/${a.id}/vector-providers`], [last]] as [string, string?][] : [['Vector providers']] as [string, string?][]),
+];
+
+/** pgvector's state as a short paragraph (with how to get it). */
+export function pgvectorNote(st: PgvectorStatus, admin: boolean, s: Session): Raw {
+  if (st.installed) return html`<p class="muted u-mt0">${icon('check')} pgvector ${st.version} is installed (schema <code>${st.schema}</code>).</p>`;
+  return html`<div class="alert alert-info" role="status"><p><strong>pgvector is not installed in this database.</strong>
+    Vector providers can be defined, but vector columns, vector search and Retrieve tools need the extension.</p>
+    ${st.available
+      ? admin && st.canCreate
+        ? html`<form method="post" action="${BASE}/vector-providers/extension">${csrf(s)}<button class="btn btn-hot">${icon('plus')} Create extension vector ${st.availableVersion}</button></form>`
+        : html`<p>The database server has pgvector ${st.availableVersion}: ${admin ? 'a superuser' : 'an administrator'} runs <code>create extension vector;</code> in this database.</p>`
+      : html`<p>The database server doesn't have the pgvector package: install it (for example <code>apt install postgresql-17-pgvector</code>, the <code>pgvector/pgvector:pg17</code> Docker image, or your cloud provider's <i>vector</i> extension), then create the extension.</p>`}
+  </div>`;
+}
+
+async function servicesOf(appId: number) {
+  return (await owner.query<{ name: string; provider: string; model: string }>(
+    `select s.name, s.provider, s.model from meta.ai_service s join meta.app_ai_service x on x.service_id = s.id and x.app_id = $1
+      where s.enabled and s.provider <> 'anthropic' order by s.name`, [appId])).rows;
+}
+
+async function providerForm(a: any, s: Session, x: Partial<VectorProvider>, original: string, problems: string[]): Promise<Raw> {
+  const services = await servicesOf(a.id);
+  const svc: [string, string][] = [['', '-'], ...services.map((v): [string, string] => [v.name, `${v.name} (${providerLabel(v.provider)})`])];
+  if (x.service && !services.some((v) => v.name === x.service)) svc.push([x.service, `${x.service} (not available to this application)`]);
+  const configs = original
+    ? (await owner.query<{ id: number; name: string; search_type: string }>(`select id, name, search_type from meta.search_config where app_id = $1 and vector_provider = $2 order by name`, [a.id, original])).rows
+    : [];
+  const columns = original ? (await vectorTriggers(await devDbForApp(a.id))).filter((t) => t.appId === a.id && t.provider === original) : [];
+  return html`${problems.length ? html`<div class="alert alert-error" role="alert">${problems.join(' ')}</div>` : ''}
+    <form method="post" action="${BASE}/apps/${a.id}/vector-providers/provider">${csrf(s)}<input type="hidden" name="original" value="${original}">
+      <fieldset class="prop-group"><legend>Vector provider</legend><div class="form-grid">
+        ${input('name', 'Name (static ID)', x.name ?? '', { required: true, placeholder: 'e.g. DOC_EMBEDDINGS', help: 'Upper case letters, digits and _. Search configurations name it; in SQL: meta.embed(text, \'DOC_EMBEDDINGS\') (function providers).' })}
+        ${input('description', 'Description', x.description ?? '')}
+        ${select('type', 'Type', x.type ?? 'ai_service', [['ai_service', 'AI service with an embedding model'], ['function', 'SQL or PL/pgSQL function']], 'Where the embeddings come from.')}
+        ${input('dimensions', 'Dimensions', x.dimensions ?? '', { type: 'number', required: true, placeholder: 'e.g. 1536', help: `The length of each vector: text-embedding-3-small 1536 (or fewer: OpenAI shortens them), text-embedding-3-large 3072, nomic-embed-text 768, gemini-embedding-001 3072. A vector column is vector(n) with this n; an HNSW index takes up to ${HNSW_MAX_DIMENSIONS}.` })}
+      </div></fieldset>
+      <fieldset class="prop-group"><legend>AI service</legend><div class="form-grid">
+        ${services.length ? '' : html`<p class="muted" data-wide>This application may not use an OpenAI, Gemini or OpenAI-compatible AI service yet: an administrator adds one under <a href="${BASE}/ai">Workspace utilities → AI services</a> (Claude has no embeddings) and allows this application to use it.</p>`}
+        ${select('service', 'AI service', x.service ?? '', svc, 'Its key, base URL and daily limits apply; the texts to embed go to its provider.')}
+        ${input('model', 'Embedding model', x.model ?? '', { placeholder: 'e.g. text-embedding-3-small', help: 'Exactly as the provider names it: text-embedding-3-small, gemini-embedding-001, nomic-embed-text, mxbai-embed-large …' })}
+        ${input('batch_size', 'Texts per request', x.batch_size ?? '', { type: 'number', placeholder: String(DEFAULT_BATCH), help: 'When the server embeds queued rows (vector columns).' })}
+        ${input('per_minute', 'Requests per minute', x.per_minute ?? '', { type: 'number', placeholder: String(DEFAULT_PER_MINUTE), help: 'At most this many background requests per minute (per server), under the provider\'s rate limit.' })}
+      </div></fieldset>
+      <fieldset class="prop-group"><legend>Function</legend><div class="form-grid">
+        ${input('function', 'Function', x.function ?? '', { placeholder: 'e.g. sales.embed', help: 'schema.name of a function (text) returning vector, real[] or float8[]. meta.embed calls it as the role that runs the query (its grants apply): grant execute on it to the application\'s role.' })}
+      </div></fieldset>
+      <div class="buttons"><button class="btn btn-hot">${original ? 'Save' : 'Add vector provider'}</button></div>
+    </form>
+    ${original ? html`
+      <form method="post" action="${BASE}/apps/${a.id}/vector-providers/test" class="u-mt1">${csrf(s)}<input type="hidden" name="name" value="${original}">
+        <button class="btn">${icon('play')} Test: embed a sample text</button></form>
+      <h3>Used by</h3>
+      ${configs.length || columns.length ? html`<ul>
+        ${configs.map((c) => html`<li>Search configuration <a href="${BASE}/apps/${a.id}/shared?c=search_config-${c.id}">${c.name}</a> (${c.search_type})</li>`)}
+        ${columns.map((c) => html`<li>Vector column <a href="${BASE}/sql/objects/vector?${new URLSearchParams({ o: c.table })}"><code>${c.table}.${c.vectorColumn}</code></a> (${c.mode === 'set' ? 'set by a trigger' : 'queued for the server'})</li>`)}
+      </ul>` : html`<p class="muted">No search configuration or vector column uses this provider yet.</p>`}
+      <form method="post" action="${BASE}/apps/${a.id}/vector-providers/provider/delete" class="danger-zone">${csrf(s)}<input type="hidden" name="name" value="${original}">
+        <button class="btn btn-danger" data-confirm="Delete vector provider ${original}? Search configurations and vector columns that use it stop working.">Delete vector provider</button></form>` : ''}`;
+}
+
+const intOrUndef = (v: string, what: string, errors: string[]) => {
+  const t = v.trim();
+  if (!t) return undefined;
+  const n = Number(t);
+  if (!Number.isInteger(n)) errors.push(`${what} is a whole number.`);
+  return n;
+};
+
+function providerFrom(b: Body, errors: string[]): VectorProvider {
+  const type = one(b.type) as VectorProvider['type'];
+  const x: VectorProvider = { name: one(b.name).trim().toUpperCase(), type, dimensions: intOrUndef(one(b.dimensions), 'Dimensions', errors) ?? 0 };
+  const description = one(b.description).trim().slice(0, 500);
+  if (description) x.description = description;
+  if (type === 'ai_service') {
+    x.service = one(b.service).trim().toUpperCase();
+    x.model = one(b.model).trim();
+    const batch = intOrUndef(one(b.batch_size), 'Texts per request', errors);
+    if (batch !== undefined) x.batch_size = batch;
+    const pm = intOrUndef(one(b.per_minute), 'Requests per minute', errors);
+    if (pm !== undefined) x.per_minute = pm;
+  }
+  if (type === 'function') x.function = one(b.function).trim();
+  return x;
+}
+
+/** Change the application's list in one transaction (row locked); `change` returns problems to refuse with. */
+async function update(appId: number, change: (list: VectorProvider[]) => string[] | Promise<string[]>) {
+  return owner.tx(async (c) => {
+    const row = (await c.query<{ vector_providers: VectorProvider[] }>('select vector_providers from meta.app where id = $1 for update', [appId])).rows[0];
+    const list = [...(row?.vector_providers ?? [])];
+    const problems = await change(list);
+    if (problems.length) return problems;
+    await c.query('update meta.app set vector_providers = $2, updated_at = now() where id = $1', [appId, JSON.stringify(list)]);
+    return [];
+  });
+}
+
+export async function vectorProviderRoutes(app: FastifyInstance) {
+  const load = async (req: Req, reply: FastifyReply) => {
+    const s = await developer(req, reply);
+    if (!s) return null;
+    const a = await appOr404(req.params.id);
+    if (!a) {
+      reply.code(404).send('Not found');
+      return null;
+    }
+    return { s, a };
+  };
+
+  app.get(`${BASE}/apps/:id(^\\d+$)/vector-providers`, async (req: Req, reply) => {
+    const ok = await load(req, reply);
+    if (!ok) return;
+    const { s, a } = ok;
+    const list = (a.vector_providers ?? []) as VectorProvider[];
+    const st = await pgvectorStatus(await devDbForApp(a.id));
+    const main = html`${appHeader(a, 'shared', s)}
+      <div class="title-row"><h1>Vector providers</h1></div>
+      <p class="muted u-mt0">A vector provider turns text into an embedding: a list of numbers whose distance says how close two texts are in meaning. Vector columns keep the embeddings of a table's rows (SQL Workshop › Table Properties › Add vector column), and search configurations of type vector or hybrid, and the Retrieve tools of AI agents, find rows by meaning. The texts go to the provider's AI service.</p>
+      ${pgvectorNote(st, await isAdmin(s.username), s)}
+      ${region('Providers', html`<div class="table-wrap"><table class="report report-reflow"><thead><tr><th>Name</th><th>Type</th><th>Model or function</th><th class="num">Dimensions</th></tr></thead>
+        <tbody>${list.length ? list.map((p) => html`<tr><td data-label="Name"><a href="${BASE}/apps/${a.id}/vector-providers/provider?name=${encodeURIComponent(p.name)}">${p.name}</a>${p.description ? html` <span class="muted">${p.description}</span>` : ''}</td>
+          <td data-label="Type">${p.type === 'ai_service' ? html`AI service ${p.service}` : 'function'}</td><td data-label="Model or function"><code>${p.type === 'ai_service' ? p.model : p.function}</code></td>
+          <td class="num" data-label="Dimensions">${p.dimensions}</td></tr>`)
+          : html`<tr><td colspan="4" class="empty">No vector providers yet.</td></tr>`}</tbody></table></div>
+        <p><a class="btn" href="${BASE}/apps/${a.id}/vector-providers/provider">${icon('plus')} Add vector provider</a></p>`)}`;
+    return send(reply, s, shell(s, 'Vector providers', crumbs(a), main));
+  });
+
+  app.get(`${BASE}/apps/:id(^\\d+$)/vector-providers/provider`, async (req: Req, reply) => {
+    const ok = await load(req, reply);
+    if (!ok) return;
+    const { s, a } = ok;
+    const name = String((req.query as Record<string, unknown>)?.name ?? '').toUpperCase();
+    const x = name ? ((a.vector_providers ?? []) as VectorProvider[]).find((p) => p.name === name) : {};
+    if (!x) return reply.code(404).send('Not found');
+    const title = name || 'New vector provider';
+    const main = html`${appHeader(a, 'shared', s)}<div class="title-row"><h1 class="ai-name">${title}</h1></div>${region('Vector provider', await providerForm(a, s, x, name, []))}`;
+    return send(reply, s, shell(s, title, crumbs(a, title), main));
+  });
+
+  app.post(`${BASE}/apps/:id(^\\d+$)/vector-providers/provider`, async (req: Req, reply) => {
+    const ok = await load(req, reply);
+    if (!ok) return;
+    const { s, a } = ok;
+    const b = (req.body ?? {}) as Body;
+    const original = one(b.original).toUpperCase();
+    const errors: string[] = [];
+    const x = providerFrom(b, errors);
+    const problems = errors.length ? errors : await update(a.id, (list) => {
+      const out = vectorProviderProblems(x);
+      const at = original ? list.findIndex((p) => p.name === original) : -1;
+      if (original && at < 0) out.push(`Vector provider ${original} no longer exists.`);
+      if (list.some((p, i) => p.name === x.name && i !== at)) out.push(`A vector provider ${x.name} exists already.`);
+      if (out.length) return out;
+      if (at >= 0) list[at] = x;
+      else list.push(x);
+      return [];
+    });
+    if (problems.length) {
+      const main = html`${appHeader(a, 'shared', s)}<div class="title-row"><h1 class="ai-name">${original || 'New vector provider'}</h1></div>${region('Vector provider', await providerForm(a, s, x, original, problems))}`;
+      return send(reply.code(422), s, shell(s, original || 'New vector provider', crumbs(a, original || 'New vector provider'), main));
+    }
+    // a renamed provider keeps its search configurations
+    if (original && original !== x.name) await owner.query('update meta.search_config set vector_provider = $3 where app_id = $1 and vector_provider = $2', [a.id, original, x.name]);
+    flash(s, 'Vector provider saved.');
+    return back(reply, s, `${BASE}/apps/${a.id}/vector-providers/provider?name=${encodeURIComponent(x.name)}`);
+  });
+
+  app.post(`${BASE}/apps/:id(^\\d+$)/vector-providers/provider/delete`, async (req: Req, reply) => {
+    const ok = await load(req, reply);
+    if (!ok) return;
+    const { s, a } = ok;
+    const name = one((req.body as Body)?.name).toUpperCase();
+    await update(a.id, (list) => {
+      const at = list.findIndex((p) => p.name === name);
+      if (at >= 0) list.splice(at, 1);
+      return [];
+    });
+    flash(s, `Vector provider ${name} deleted.`);
+    return back(reply, s, `${BASE}/apps/${a.id}/vector-providers`);
+  });
+
+  // Test: embed a sample text (an AI service: one request, logged under AI usage; a function: as the developer connection)
+  app.post(`${BASE}/apps/:id(^\\d+$)/vector-providers/test`, async (req: Req, reply) => {
+    const ok = await load(req, reply);
+    if (!ok) return;
+    const { s, a } = ok;
+    const name = one((req.body as Body)?.name).toUpperCase();
+    const p = ((a.vector_providers ?? []) as VectorProvider[]).find((x) => x.name === name);
+    if (!p) return reply.code(404).send('Not found');
+    const problems = vectorProviderProblems(p);
+    const to = `${BASE}/apps/${a.id}/vector-providers/provider?name=${encodeURIComponent(name)}`;
+    if (problems.length) {
+      flash(s, problems.join(' '), 'error');
+      return back(reply, s, to);
+    }
+    const t0 = performance.now();
+    try {
+      const [v] = await embedTexts(p, ['The quick brown fox jumps over the lazy dog.'], { appId: a.id, user: s.username, pageNo: null }, await devDbForApp(a.id));
+      flash(s, `${name} answered in ${Math.round(performance.now() - t0)} ms with ${v.length} dimensions, as the provider says: [${v.slice(0, 4).map((n) => n.toFixed(4)).join(', ')}, …].`);
+    } catch (e) {
+      flash(s, `${name}: ${isAiError(e) ? e.message : (e as Error).message}`, 'error');
+    }
+    return back(reply, s, to);
+  });
+
+  // ---------------------------------------------------------------- Workspace utilities → Vector providers
+
+  app.get(`${BASE}/vector-providers`, async (req: Req, reply) => {
+    const s = await developer(req, reply);
+    if (!s) return;
+    const admin = await isAdmin(s.username);
+    const st = await pgvectorStatus(owner);
+    const apps = (await owner.query<{ id: number; name: string; alias: string; vector_providers: VectorProvider[] }>(
+      `select id, name, alias, vector_providers from meta.app a where ${inWorkspaceSql('a', 1)} and jsonb_array_length(vector_providers) > 0 order by name`, [currentWorkspace(s)])).rows;
+    const ids = new Set(apps.map((x) => x.id));
+    const columns = st.installed ? (await vectorTriggers(owner)).filter((t) => ids.has(t.appId)) : [];
+    const rows = apps.flatMap((x) => x.vector_providers.map((p) => ({ app: x, p })));
+    const main = html`<div class="title-row"><h1>Vector providers</h1></div>
+      <p class="muted u-mt0">The vector providers of this workspace's applications (each application defines its own under Shared Components → Vector providers), and the vector columns they keep current.</p>
+      ${pgvectorNote(st, admin, s)}
+      ${region('Providers', html`<div class="table-wrap"><table class="report report-reflow"><thead><tr><th>Application</th><th>Name</th><th>Type</th><th>Model or function</th><th class="num">Dimensions</th></tr></thead>
+        <tbody>${rows.length ? rows.map(({ app: x, p }) => html`<tr><td data-label="Application">${x.name}</td>
+          <td data-label="Name"><a href="${BASE}/apps/${x.id}/vector-providers/provider?name=${encodeURIComponent(p.name)}">${p.name}</a></td>
+          <td data-label="Type">${p.type === 'ai_service' ? html`AI service ${p.service}` : 'function'}</td><td data-label="Model or function"><code>${p.type === 'ai_service' ? p.model : p.function}</code></td>
+          <td class="num" data-label="Dimensions">${p.dimensions}</td></tr>`)
+          : html`<tr><td colspan="5" class="empty">No vector providers yet.</td></tr>`}</tbody></table></div>`)}
+      ${region('Vector columns', html`<div class="table-wrap"><table class="report report-reflow"><thead><tr><th>Column</th><th>Provider</th><th>Source columns</th><th>Kept by</th></tr></thead>
+        <tbody>${columns.length ? columns.map((c) => html`<tr><td data-label="Column"><a href="${BASE}/sql/objects/vector?${new URLSearchParams({ o: c.table })}"><code>${c.table}.${c.vectorColumn}</code></a></td>
+          <td data-label="Provider">${c.provider}</td><td data-label="Source columns">${c.sourceColumns.join(', ')}</td><td data-label="Kept by">${c.mode === 'set' ? 'a trigger' : 'the server\'s queue'}</td></tr>`)
+          : html`<tr><td colspan="4" class="empty">No vector columns yet (SQL Workshop › Table Properties › Add vector column).</td></tr>`}</tbody></table></div>`)}`;
+    return send(reply, s, shell(s, 'Vector providers', [['App Builder', BASE], ['Workspace utilities', `${BASE}/utilities`], ['Vector providers']], main));
+  });
+
+  // administrators: create extension vector (pgvector is not a trusted extension: the owner connection must be a superuser)
+  app.post(`${BASE}/vector-providers/extension`, async (req: Req, reply) => {
+    const s = await developer(req, reply);
+    if (!s) return;
+    if (!(await isAdmin(s.username))) return reply.code(403).send('Only administrators create extensions.');
+    try {
+      await owner.query('create extension if not exists vector');
+      await logActivity({ username: s.username, event: 'sql_command', ip: clientIp(req), detail: 'create extension if not exists vector' });
+      flash(s, 'The pgvector extension is installed.');
+    } catch (e) {
+      flash(s, `create extension vector failed: ${(e as Error).message}`, 'error');
+    }
+    return back(reply, s, `${BASE}/vector-providers`);
+  });
+}
+
+export const vectorProviderOptions = async (appId: number) =>
+  (((await owner.one<{ vector_providers: VectorProvider[] }>('select vector_providers from meta.app where id = $1', [appId]))?.vector_providers ?? []) as VectorProvider[]);