feat: add dedicated UI for reasoning model responses (#261)

DeepSeek R1 returns it's reasoning process wrapped in `<think></think>`
tags.
We parse those while the completion is underway and move it's contents
to a collapsible UI component.

README.md

@@ -7,6 +7,7 @@ A minimal web-UI for talking to [Ollama](https://github.com/jmorganca/ollama/) s
 - Support for **Ollama** & **OpenAI** models
 - Multi-server support
 - Large prompt fields
+- Support for reasoning models
 - Markdown rendering with syntax highlighting
 - Code editor features
 - Customizable system prompts & advanced Ollama parameters

package.json

@@ -19,7 +19,7 @@
 	},
 	"devDependencies": {
 		"@ianvs/prettier-plugin-sort-imports": "^4.3.1",
-		"@playwright/test": "^1.43.0",
+		"@playwright/test": "^1.50.0",
 		"@sveltejs/adapter-auto": "^3.3.1",
 		"@sveltejs/adapter-cloudflare": "^4.7.4",
 		"@sveltejs/adapter-node": "^5.2.9",

src/i18n/en/index.ts

@@ -113,6 +113,7 @@ const en = {
 	pullModelPlaceholder: 'Model tag (e.g. llama3.1)',
 	pullingModel: 'Pulling model',
 	random: 'Random',
+	reasoning: 'Reasoning',
 	refreshToUpdate: 'Refresh to update',
 	releaseHistory: 'Release history',
 	repeatLastN: 'Repeat last N',

src/i18n/i18n-types.ts

@@ -462,6 +462,10 @@ type RootTranslation = {
 	 * R​a​n​d​o​m
 	 */
 	random: string
+	/**
+	 * R​e​a​s​o​n​i​n​g
+	 */
+	reasoning: string
 	/**
 	 * R​e​f​r​e​s​h​ ​t​o​ ​u​p​d​a​t​e
 	 */
@@ -1072,6 +1076,10 @@ The completion in progress will stop
 	 * Random
 	 */
 	random: () => LocalizedString
+	/**
+	 * Reasoning
+	 */
+	reasoning: () => LocalizedString
 	/**
 	 * Refresh to update
 	 */

src/lib/chat/openai.ts

@@ -1,9 +1,10 @@
 import OpenAI from 'openai';
 import type { ChatCompletionMessageParam } from 'openai/resources/index.mjs';
 
-import type { Server } from '$lib/servers';
+import type { Server } from '$lib/connections';
+import type { Model } from '$lib/settings';
 
-import type { ChatRequest, ChatStrategy, Model } from './index';
+import type { ChatRequest, ChatStrategy } from './index';
 
 export class OpenAIStrategy implements ChatStrategy {
 	private openai: OpenAI;

src/lib/sessions.ts

@@ -11,6 +11,7 @@ import { formatTimestampToNow } from './utils';
 export interface Message extends ChatMessage {
 	knowledge?: Knowledge;
 	context?: number[];
+	reasoning?: string;
 }
 
 export interface Session {
@@ -31,6 +32,7 @@ export interface Editor {
 	isNewSession: boolean;
 	shouldFocusTextarea: boolean;
 	completion?: string;
+	reasoning?: string;
 	promptTextarea?: HTMLTextAreaElement;
 	abortController?: AbortController;
 }

src/routes/motd/motd.md

@@ -1,18 +1,19 @@
-`2024-11-25`
+`2025-1-26`
 
 ### Message of the day
 
 # Welcome to Hollama: a simple web interface for [Ollama](https://ollama.ai)
 
 #### What's new?
 
+- **Reasoning responses** (i.e. [`deepseek-r1`](https://ollama.com/library/deepseek-r1)) are now displayed in a dedicated UI component.
 - **Multiple-server support** allows you to connect to one or more Ollama (and/or OpenAI) servers at the same time.
-- **Models list can be filtered** by keyword for each server.
-- **Servers can be labeled** to help you identify them in the models list.
-- **Hallo Welt!** UI is now available in German.
 
 #### Previously, in Hollama
 
+- **Models list can be filtered** by keyword for each server.
+- **Servers can be labeled** to help you identify them in the models list.
+- **Hallo Welt!** UI is now available in German.
 - **OpenAI models** are now _(optionally)_ available in Sessions. Set your own API key in [Settings](/settings)
 - **[Knowledge](/knowledge)** can now be used as context at any point in a Session.
 - **Model** and **advanced Ollama settings** can be changed at any time on an existing session

src/routes/sessions/[id]/+page.svelte

@@ -30,6 +30,9 @@
 	import Messages from './Messages.svelte';
 	import Prompt from './Prompt.svelte';
 
+	const THINK_TAG = '<think>';
+	const END_THINK_TAG = '</think>';
+
 	interface Props {
 		data: PageData;
 	}
@@ -135,8 +138,9 @@
 	async function handleCompletion(messages: Message[]) {
 		editor.abortController = new AbortController();
 		editor.isCompletionInProgress = true;
-		editor.prompt = ''; // Reset the prompt form field
+		editor.prompt = '';
 		editor.completion = '';
+		editor.reasoning = '';
 
 		const server = $serversStore.find((s) => s.id === session.model?.serverId);
 		if (!server) throw new Error('Server not found');
@@ -161,19 +165,49 @@
 			}
 
 			if (!strategy) throw new Error('Invalid strategy');
+
+			let isInThinkTag = false;
 			await strategy.chat(chatRequest, editor.abortController.signal, async (chunk) => {
-				editor.completion += chunk;
+				// This is required primarily for testing, because both the reasoning
+				// and the completion are returned in a single chunk.
+				if (chunk.includes(THINK_TAG) && chunk.includes(END_THINK_TAG)) {
+					const start = chunk.indexOf(THINK_TAG) + THINK_TAG.length;
+					const end = chunk.indexOf(END_THINK_TAG);
+					editor.reasoning += chunk.slice(start, end);
+					chunk = chunk.slice(end);
+				}
+
+				if (chunk.includes(THINK_TAG)) {
+					isInThinkTag = true;
+					chunk = chunk.replace(THINK_TAG, '');
+				}
+
+				if (chunk.includes(END_THINK_TAG)) {
+					isInThinkTag = false;
+					chunk = chunk.replace(END_THINK_TAG, '');
+				}
+
+				if (isInThinkTag) {
+					editor.reasoning += chunk;
+				} else {
+					editor.completion += chunk;
+				}
+
 				await scrollToBottom();
 			});
 
-			// After the completion save the session
-			const message: Message = { role: 'assistant', content: editor.completion };
+			const message: Message = {
+				role: 'assistant',
+				content: editor.completion,
+				reasoning: editor.reasoning
+			};
+
 			session.messages = [...session.messages, message];
 			session.updatedAt = new Date().toISOString();
 			saveSession(session);
 
-			// Final housekeeping
 			editor.completion = '';
+			editor.reasoning = '';
 			editor.shouldFocusTextarea = true;
 			editor.isCompletionInProgress = false;
 			await scrollToBottom();

src/routes/sessions/[id]/Article.svelte

@@ -1,5 +1,7 @@
 <script lang="ts">
-	import { BrainIcon, Pencil, RefreshCw, Trash2 } from 'lucide-svelte';
+	import { BrainIcon, ChevronDown, ChevronUp, Pencil, RefreshCw, Trash2 } from 'lucide-svelte';
+	import { quadInOut } from 'svelte/easing';
+	import { slide } from 'svelte/transition';
 
 	import LL from '$i18n/i18n-svelte';
 	import Badge from '$lib/components/Badge.svelte';
@@ -18,6 +20,7 @@
 
 	let isKnowledgeAttachment: boolean | undefined;
 	let isUserRole: boolean | undefined;
+	let isReasoningVisible: boolean = false;
 
 	$: if (message) {
 		isKnowledgeAttachment = message.knowledge?.name !== undefined;
@@ -84,11 +87,32 @@
 			</div>
 		</nav>
 
-		<div class="markdown">
-			{#if message.content}
-				<Markdown markdown={message.content} />
-			{/if}
-		</div>
+		{#if message.reasoning}
+			<div class="reasoning" transition:slide={{ easing: quadInOut, duration: 200 }}>
+				<button
+					class="reasoning__button"
+					on:click={() => (isReasoningVisible = !isReasoningVisible)}
+				>
+					{$LL.reasoning()}
+					{#if isReasoningVisible}
+						<ChevronUp class="base-icon" />
+					{:else}
+						<ChevronDown class="base-icon" />
+					{/if}
+				</button>
+				{#if isReasoningVisible}
+					<article
+						class="article article--reasoning"
+						transition:slide={{ easing: quadInOut, duration: 200 }}
+					>
+						<Markdown markdown={message.reasoning} />
+					</article>
+				{/if}
+			</div>
+		{/if}
+		{#if message.content}
+			<Markdown markdown={message.content} />
+		{/if}
 	</article>
 {/if}
 
@@ -104,6 +128,10 @@
 		@apply border-transparent bg-shade-0;
 	}
 
+	.article--reasoning {
+		@apply max-w-full border-b-0 border-l-0 border-r-0;
+	}
+
 	.article__interactive,
 	.attachment__interactive {
 		@apply -mr-2 opacity-100;
@@ -148,4 +176,12 @@
 	.attachment__content {
 		@apply flex items-center gap-2;
 	}
+
+	.reasoning {
+		@apply rounded bg-shade-1 text-xs;
+	}
+
+	.reasoning__button {
+		@apply flex w-full items-center justify-between gap-2 p-2;
+	}
 </style>

src/routes/sessions/[id]/Messages.svelte

@@ -43,5 +43,11 @@
 {/each}
 
 {#if editor.isCompletionInProgress}
-	<Article message={{ role: 'assistant', content: editor.completion || '...' }} />
+	<Article
+		message={{
+			role: 'assistant',
+			content: editor.completion || '...',
+			reasoning: editor.reasoning
+		}}
+	/>
 {/if}

tests/controls.test.ts

@@ -207,6 +207,7 @@ test('can set ollama model and runtime options', async ({ page }) => {
 			},
 			{
 				role: 'assistant',
+				reasoning: '',
 				content: MOCK_SESSION_1_RESPONSE_1.message.content
 			},
 			{
@@ -296,6 +297,7 @@ test('can set ollama model and runtime options', async ({ page }) => {
 			},
 			{
 				role: 'assistant',
+				reasoning: '',
 				content: MOCK_SESSION_1_RESPONSE_1.message.content
 			},
 			{
@@ -304,6 +306,7 @@ test('can set ollama model and runtime options', async ({ page }) => {
 			},
 			{
 				role: 'assistant',
+				reasoning: '',
 				content: MOCK_SESSION_1_RESPONSE_2.message.content
 			},
 			{

tests/knowledge.test.ts

@@ -270,7 +270,11 @@ test('can use knowledge as system prompt in the session', async ({ page }) => {
 			messages: [
 				{ role: 'system', content: MOCK_KNOWLEDGE[0].content, knowledge: MOCK_KNOWLEDGE[0] },
 				{ role: 'user', content: 'What is this about?' },
-				{ role: 'assistant', content: MOCK_SESSION_WITH_KNOWLEDGE_RESPONSE_1.message.content },
+				{
+					role: 'assistant',
+					content: MOCK_SESSION_WITH_KNOWLEDGE_RESPONSE_1.message.content,
+					reasoning: ''
+				},
 				{ role: 'user', content: 'Gotcha, thanks for the clarification' }
 			]
 		})