feat(agents): select knowledge-base models

core/http/react-ui/e2e/agent-kb-models.spec.js

@@ -0,0 +1,45 @@
+import { test, expect } from './coverage-fixtures.js'
+
+test('agent knowledge-base selectors filter models and save their selections', async ({ page }) => {
+  await page.route('**/api/agents/config/metadata', route => route.fulfill({
+    json: { Fields: [
+      { name: 'name', label: 'Name', type: 'text', required: true, tags: { section: 'BasicInfo' } },
+      { name: 'model', label: 'Reasoning Model', type: 'text', defaultValue: 'chat-model', tags: { section: 'ModelSettings' } },
+      { name: 'embedding_model', label: 'Embedding Model', type: 'text', tags: { section: 'ModelSettings' } },
+      { name: 'reranker_model', label: 'Reranker Model', type: 'text', tags: { section: 'ModelSettings' } },
+    ] },
+  }))
+  await page.route('**/api/models/capabilities', route => route.fulfill({
+    json: { data: [
+      { id: 'chat-model', capabilities: ['FLAG_CHAT'] },
+      { id: 'embed-model', capabilities: ['FLAG_EMBEDDINGS'] },
+      { id: 'rank-model', capabilities: ['FLAG_RERANK'] },
+    ] },
+  }))
+  await page.route('**/api/agents', route => route.fulfill({
+    json: route.request().method() === 'POST' ? {} : { agents: [], statuses: {} },
+  }))
+
+  await page.goto('/app/agents/new')
+  await page.locator('#field-name').fill('research')
+  await page.getByRole('button', { name: /^Model Settings/ }).click()
+
+  const embedding = page.locator('.form-row').filter({ has: page.getByText('Embedding Model', { exact: true }) })
+  const reranker = page.locator('.form-row').filter({ has: page.getByText('Reranker Model', { exact: true }) })
+  await expect(embedding.locator('input')).toHaveValue('')
+  await expect(reranker.locator('input')).toHaveValue('')
+
+  await embedding.locator('input').click()
+  await expect(embedding.getByRole('option')).toHaveCount(1)
+  await embedding.getByRole('option', { name: /embed-model/ }).click()
+  await reranker.locator('input').click()
+  await expect(reranker.getByRole('option')).toHaveCount(1)
+  await reranker.getByRole('option', { name: /rank-model/ }).click()
+
+  const saved = page.waitForRequest(request => request.method() === 'POST' && new URL(request.url()).pathname === '/api/agents')
+  await page.getByRole('button', { name: /Create Agent$/ }).first().click()
+  expect((await saved).postDataJSON()).toMatchObject({
+    name: 'research', model: 'chat-model', embedding_model: 'embed-model', reranker_model: 'rank-model',
+  })
+  await expect(page).toHaveURL(/\/app\/agents$/)
+})

core/http/react-ui/e2e/chat-history.spec.js

@@ -18,13 +18,15 @@ async function openMenu(page) {
 }
 
 test.describe('Chat conversations menu', () => {
+  const now = new Date('2026-10-09T12:00:00Z').getTime()
   test.beforeEach(async ({ page }) => {
+    await page.clock.setFixedTime(now)
     await mockChat(page, { models: [
       { id: 'qwen3-8b-instruct', capabilities: ['FLAG_CHAT'] },
       { id: 'gemma-4-e4b-it-qat-q4_0', capabilities: ['FLAG_CHAT'] },
       { id: 'qwen2.5-vl-7b-instruct', capabilities: ['FLAG_CHAT'] },
     ], loaded: ['qwen3-8b-instruct'] })
-    await openChat(page, sampleChats())
+    await openChat(page, sampleChats(now))
     await expect(page.getByTestId('chat-message').first()).toBeVisible()
   })
 
@@ -153,7 +155,7 @@ test.describe('Chat conversations menu', () => {
   })
 
   test('the undo time runs out by itself after six seconds', async ({ page }) => {
-    await page.clock.install()
+    await page.clock.install({ time: now })
     await page.reload()
     await page.getByTestId('chat-composer').waitFor()
     await openMenu(page)

core/http/react-ui/e2e/home-console.spec.js

@@ -317,9 +317,11 @@ test.describe('Home in a cluster', () => {
 })
 
 test.describe('Home resume list', () => {
+  const now = new Date('2026-10-09T12:00:00Z').getTime()
   test.beforeEach(async ({ page }) => {
+    await page.clock.setFixedTime(now)
     await mockHome(page)
-    await seedChats(page)
+    await seedChats(page, sampleChats(now))
   })
 
   test('groups conversations by day, one card per day', async ({ page }) => {
@@ -408,7 +410,7 @@ test.describe('Home resume list', () => {
   })
 
   test('without undo the delete becomes final when the time ends', async ({ page }) => {
-    await page.clock.install()
+    await page.clock.install({ time: now })
     await ready(page)
     const row = page.getByTestId('home-conversation').filter({ hasText: 'invoice' })
     await row.hover()
@@ -430,12 +432,12 @@ test.describe('Home resume list', () => {
   })
 
   test('old conversations wait behind one button', async ({ page }) => {
-    const chats = sampleChats()
+    const chats = sampleChats(now)
     const old = Array.from({ length: 8 }, (_, i) => ({
       ...chats[0],
       id: `old-${i}`,
       name: `Old chat ${i}`,
-      updatedAt: Date.now() - (10 + i) * 24 * 60 * 60 * 1000,
+      updatedAt: now - (10 + i) * 24 * 60 * 60 * 1000,
     }))
     await page.addInitScript(([key, data]) => localStorage.setItem(key, JSON.stringify(data)),
       [CHATS_KEY, { chats: [...chats, ...old], activeChatId: chats[0].id }])

core/http/react-ui/e2e/model-page.spec.js

@@ -199,15 +199,17 @@ test.describe('Model page - walking and going back', () => {
     const { state } = await setup(page)
     await expect(rows(page).first()).toBeVisible({ timeout: 15_000 })
     await page.waitForTimeout(500)
-    const before = state.listingRequests.length
+    // Facet counts load in the background; they do not reload the list.
+    const listings = () => state.listingRequests.filter(search => new URLSearchParams(search).get('items') !== '1')
+    const before = listings().length
     await page.locator('[data-entity="qwen3-8b-instruct"] [data-row-open]').click()
     await expect(name(page)).toHaveText('qwen3-8b-instruct')
     await page.keyboard.press('Escape')
     await expect(rows(page).first()).toBeVisible()
     await page.waitForTimeout(500)
     // The page itself looks its own entry up once; the list asks for nothing.
-    expect(state.listingRequests.length - before).toBeLessThanOrEqual(1)
-    expect(state.listingRequests.slice(before).every(search => search.includes('term=qwen3-8b-instruct'))).toBe(true)
+    expect(listings().length - before).toBeLessThanOrEqual(1)
+    expect(listings().slice(before).every(search => search.includes('term=qwen3-8b-instruct'))).toBe(true)
   })
 
   test('changing tabs keeps the walker, and Back still leaves for the list in one step', async ({ page }) => {

core/http/react-ui/e2e/models-cleanup.spec.js

@@ -92,11 +92,10 @@ test.describe('Cleanup sheet - what it knows and what it does not', () => {
 
   test('each tier is ranked by size and a model with no known size goes last', async ({ page }) => {
     await openSheet(page)
-    const names = await tier(page, 'call').locator('.ledger-cr__name').evaluateAll(
-      (els) => els.map((e) => e.firstChild.textContent.trim()),
-    )
-    expect(names[0]).toBe('llama-3.3-70b-instruct-iq2')
-    expect(names[names.length - 1]).toBe('my-finetune-q4')
+    // Estimates arrive after the sheet becomes interactive and can reorder it.
+    const names = tier(page, 'call').locator('.ledger-cr__name')
+    await expect(names.first()).toContainText('llama-3.3-70b-instruct-iq2')
+    await expect(names.last()).toContainText('my-finetune-q4')
     await expect(tier(page, 'call').locator('li').last()).toContainText('size unknown')
     await expect(tier(page, 'call').locator('li').first()).toContainText(`${gb1(sizeOf('llama-3.3-70b-instruct-iq2'))} GB`)
   })

core/http/react-ui/e2e/studio-work.spec.js

@@ -135,9 +135,8 @@ test.describe('Your work', () => {
 
   test('the masonry uses fewer columns on a narrow screen', async ({ page }) => {
     await open(page)
-    const wide = await page.locator('.studio-masonry').getAttribute('data-columns')
+    await expect.poll(async () => Number(await page.locator('.studio-masonry').getAttribute('data-columns'))).toBeGreaterThan(1)
     await page.setViewportSize({ width: 390, height: 844 })
     await expect.poll(() => page.locator('.studio-masonry').getAttribute('data-columns')).toBe('1')
-    expect(Number(wide)).toBeGreaterThan(1)
   })
 })

core/http/react-ui/e2e/traffic-models-host.spec.js

@@ -17,7 +17,8 @@ test.describe('Traffic models', () => {
     await expect(bge.locator('td.dk-num').last()).toHaveText('-')
     // The failed load of flux-dev is in the backend-operation buffer.
     await expect(page.locator('tr[data-entity="flux-dev"]')).toContainText('1')
-    await expect(page.getByText(/p50|p95|p99/)).toHaveCount(0)
+    // A row's concatenated text can contain llama-cpp followed by 99 requests.
+    await expect(page.getByText(/\bp(?:50|95|99)\b/i)).toHaveCount(0)
   })
 
   test('a row opens in place with the split, the buffer and the links', async ({ page }) => {

core/http/react-ui/src/components/models/page/model-page.css

@@ -262,15 +262,17 @@ a.modelpage-chip { cursor: pointer; text-decoration: none; color: inherit; }
   .modelpage-keys { display: none; }
 }
 
-.modelpage-table .dk-table-sub { max-width: none; white-space: normal; overflow-wrap: anywhere; }
-.modelpage-table .dk-table-sub .modelpage-inline-badge:first-child { margin-left: 0; }
+.modelpage-table .dk-table .dk-table-sub { max-width: none; white-space: normal; overflow-wrap: anywhere; }
+.modelpage-table .dk-table .dk-table-sub .modelpage-inline-badge:first-child { margin-left: 0; }
 
 /* A grid child is as wide as its content unless told otherwise; a table inside
    one would push the whole page sideways instead of scrolling in its own box. */
 .modelpage-variants > section, .modelpage-usage > section, .modelpage-columns > *, .modelpage-fit > * { min-width: 0; }
 .modelpage-variants__size { display: none; }
 @media (max-width: 640px) {
   .modelpage-variants__size { display: inline; }
+  /* Leave room for Fit and Install while long build names wrap. */
+  .modelpage-table .dk-table .dk-table-sub { max-width: 14ch; }
   .modelpage-table .dk-table-actions .dk-btn { padding-inline: var(--dk-space-2); }
 }
 

core/http/react-ui/src/pages/AgentCreate.jsx

@@ -13,7 +13,7 @@ import './library.css'
 import SearchableModelSelect from '../components/SearchableModelSelect'
 // eslint-disable-next-line no-unused-vars
 import UnsavedChangesGuard from '../components/UnsavedChangesGuard'
-import { CAP_CHAT, CAP_TRANSCRIPT, CAP_TTS } from '../utils/capabilities'
+import { CAP_CHAT, CAP_TRANSCRIPT, CAP_TTS, CAP_EMBEDDINGS, CAP_RERANK } from '../utils/capabilities'
 // eslint-disable-next-line no-unused-vars
 import Toggle from '../components/Toggle'
 // eslint-disable-next-line no-unused-vars
@@ -128,14 +128,15 @@ function FormField({ field, value, onChange, disabled }) {
         </SettingRow>
       )
     default: {
-      const isModelField = /^(model|multimodal_model|transcription_model|tts_model|embedding_model)$/.test(field.name)
+      const isModelField = /^(model|multimodal_model|transcription_model|tts_model|embedding_model|reranker_model)$/.test(field.name)
       if (isModelField && !disabled && !field.disabled) {
         const capabilityMap = {
           model: CAP_CHAT,
           multimodal_model: CAP_CHAT,
           transcription_model: CAP_TRANSCRIPT,
           tts_model: CAP_TTS,
-          embedding_model: undefined,
+          embedding_model: CAP_EMBEDDINGS,
+          reranker_model: CAP_RERANK,
         }
         return (
           <SettingRow label={label} description={field.helpText}>

core/services/agentpool/agent_pool.go

@@ -191,6 +191,7 @@ func (s *AgentPoolService) Start(ctx context.Context) error {
 func (s *AgentPoolService) buildCollectionsConfig(apiURL, apiKey, collectionDBPath, fileAssets string) *collections.Config {
 	cfg := s.appConfig.AgentPool
 	return &collections.Config{
+		ModelSettings:    s.collectionModelSettings(""),
 		LLMAPIURL:        apiURL,
 		LLMAPIKey:        apiKey,
 		LLMModel:         cfg.DefaultModel,
@@ -209,6 +210,7 @@ func (s *AgentPoolService) buildCollectionsConfig(apiURL, apiKey, collectionDBPa
 // Skills and collections are still initialized for the frontend UI.
 func (s *AgentPoolService) startDistributed(ctx context.Context, apiURL, apiKey string) error {
 	cfg := s.appConfig.AgentPool
+	s.configBackend = newDistributedAgentConfigBackend(s, s.distributed.agentStore)
 
 	// State dir for skills and outputs
 	stateDir := cmp.Or(cfg.StateDir, s.appConfig.DataPath, s.appConfig.DynamicConfigsDir, "agents")
@@ -264,9 +266,6 @@ func (s *AgentPoolService) startDistributed(ctx context.Context, apiURL, apiKey
 		go scheduler.Start(ctx)
 	}
 
-	// Wire the distributed config backend
-	s.configBackend = newDistributedAgentConfigBackend(s, s.distributed.agentStore)
-
 	xlog.Info("Agent pool started in distributed mode (frontend dispatcher only)", "apiURL", apiURL, "stateDir", stateDir)
 	return nil
 }
@@ -339,6 +338,7 @@ func (s *AgentPoolService) startLocalAGI(_ context.Context, cfg config.AgentPool
 		return fmt.Errorf("failed to create agent pool: %w", err)
 	}
 	s.localAGI.pool = pool
+	s.configBackend = newLocalAgentConfigBackend(s)
 
 	// Create in-process collections backend and RAG provider
 	collectionsCfg := s.buildCollectionsConfig(apiURL, apiKey, collectionDBPath, fileAssets)
@@ -363,9 +363,6 @@ func (s *AgentPoolService) startLocalAGI(_ context.Context, cfg config.AgentPool
 		xlog.Error("Failed to start agent pool", "error", err)
 	}
 
-	// Wire the local config backend
-	s.configBackend = newLocalAgentConfigBackend(s)
-
 	xlog.Info("Agent pool started (standalone/LocalAGI mode)", "stateDir", stateDir, "apiURL", apiURL)
 	return nil
 }
@@ -709,6 +706,7 @@ func (s *AgentPoolService) ExportAgent(name string) ([]byte, error) {
 
 // SetUserServicesManager sets the user services manager for per-user scoping.
 func (s *AgentPoolService) SetUserServicesManager(usm *UserServicesManager) {
+	usm.collectionModelSettings = s.collectionModelSettings
 	s.users.userServices = usm
 }
 

core/services/agentpool/collection_models.go

@@ -0,0 +1,78 @@
+package agentpool
+
+import (
+	"fmt"
+	"slices"
+	"strings"
+
+	"github.com/mudler/LocalAGI/core/state"
+	"github.com/mudler/LocalAGI/webui/collections"
+	"github.com/mudler/LocalAI/core/http/auth"
+)
+
+// Resolve on each operation so edits and permission changes apply to existing
+// collections. The callback must not hold the pool or user-services mutex.
+func (s *AgentPoolService) collectionModelSettings(owner string) func(string) (collections.CollectionModelSettings, error) {
+	return func(collection string) (collections.CollectionModelSettings, error) {
+		userID, name := owner, collection
+		if prefix, agentName, namespaced := strings.Cut(collection, ":"); namespaced {
+			if owner != "" && prefix != owner {
+				return collections.CollectionModelSettings{}, fmt.Errorf("collection belongs to another user")
+			}
+			userID, name = prefix, agentName
+		}
+		settings := collections.CollectionModelSettings{}
+		if s.configBackend == nil {
+			return settings, fmt.Errorf("agent configuration unavailable")
+		}
+		cfg := s.configBackend.GetConfig(userID, name)
+		if cfg == nil {
+			// Embedded RAG normalizes storage names. Resolve REST requests only when
+			// that normalized name identifies one agent in this user's namespace.
+			var match *state.AgentConfig
+			for candidate := range s.configBackend.ListAgents(userID) {
+				if strings.TrimSpace(strings.ToLower(candidate)) == name {
+					if match != nil {
+						return settings, fmt.Errorf("ambiguous agent collection %q", name)
+					}
+					match = s.configBackend.GetConfig(userID, candidate)
+				}
+			}
+			cfg = match
+		}
+		if cfg == nil {
+			return settings, nil
+		}
+		settings.EmbeddingModel, settings.RerankerModel = cfg.EmbeddingModel, cfg.RerankerModel
+		if err := s.checkCollectionModels(userID, settings); err != nil {
+			return collections.CollectionModelSettings{}, err
+		}
+		return settings, nil
+	}
+}
+
+func (s *AgentPoolService) checkCollectionModels(userID string, settings collections.CollectionModelSettings) error {
+	if userID == "" || s.users.authDB == nil || (settings.EmbeddingModel == "" && settings.RerankerModel == "") {
+		return nil
+	}
+	var user auth.User
+	if err := s.users.authDB.Where("id = ?", userID).First(&user).Error; err != nil {
+		return fmt.Errorf("check collection model access: %w", err)
+	}
+	if user.Role == auth.RoleAdmin {
+		return nil
+	}
+	permissions, err := auth.GetUserPermissions(s.users.authDB, userID)
+	if err != nil {
+		return fmt.Errorf("check collection model permissions: %w", err)
+	}
+	if !permissions.AllowedModels.Enabled {
+		return nil
+	}
+	for _, model := range []string{settings.EmbeddingModel, settings.RerankerModel} {
+		if model != "" && !slices.Contains(permissions.AllowedModels.Models, model) {
+			return fmt.Errorf("model %q is not allowed for this user", model)
+		}
+	}
+	return nil
+}