feat: add ability to retry completions

README.md

@@ -7,10 +7,11 @@ A minimal web-UI for talking to [Ollama](https://github.com/jmorganca/ollama/) s
 - Large prompt fields with code editor features
 - Markdown parsing w/syntax highlighting
 - Easily copy markdown as raw text
-- _"Knowledge"_ lets you add context to sessions
+- Customizable system prompts
 - Saves all changes on your browser's `localStorage`
-- Responsive layout
-- Dark mode
+- Desktop & mobile friendly layout
+- Light & dark themes
+- Retryable completions
 - Streams completions
 
 ### Live demo

src/lib/ollama.ts

@@ -1,11 +1,11 @@
 import { get } from "svelte/store";
-import type { Session } from "$lib/sessions";
 import { settingsStore } from "$lib/store";
 
-type OllamaCompletionRequest = {
-	context: number[];
+export type OllamaCompletionRequest = {
 	prompt: string;
 	model: string;
+	context?: number[];
+	system?: string;
 }
 
 export type OllamaCompletionResponse = {
@@ -42,31 +42,10 @@ export type OllamaTagResponse = {
 	models: OllamaModel[];
 };
 
-export async function ollamaGenerate(session: Session, abortSignal: AbortSignal) {
+export async function ollamaGenerate(payload: OllamaCompletionRequest, abortSignal: AbortSignal) {
 	const settings = get(settingsStore);
 	if (!settings) throw new Error('No Ollama server specified');
 
-	let payload: OllamaCompletionRequest = {
-		model: session.model,
-		context: session.context,
-		prompt: session.messages[session.messages.length - 1].content
-	};
-
-	const firstMessage = session.messages[0]
-	if (firstMessage.knowledge) {
-		payload.prompt = `
-			<CONTEXT
-				name="${firstMessage.knowledge.name}"
-				id="${firstMessage.knowledge.id}"
-				updatedAt="${firstMessage.knowledge.updatedAt}"
-			>
-				${firstMessage.knowledge.content}
-			</CONTEXT>
-
-			${payload.prompt}
-		`;
-	}
-
 	return await fetch(`${settings.ollamaServer}/api/generate`, {
 		method: 'POST',
 		headers: { 'Content-Type': 'text/event-stream' },

src/lib/sessions.ts

@@ -7,13 +7,13 @@ export interface Message {
 	role: 'user' | 'ai' | 'system';
 	content: string;
 	knowledge?: Knowledge;
+	context?: number[];
 }
 
 export interface Session {
 	id: string;
 	model: string;
 	messages: Message[];
-	context: number[];
 	updatedAt?: string;
 	knowledge?: Knowledge;
 }
@@ -35,7 +35,7 @@ export const loadSession = (id: string): Session => {
 		const model = get(settingsStore)?.ollamaModel || '';
 
 		// Create a new session
-		session = { id, model, messages: [], context: [], updatedAt: new Date().toISOString() };
+		session = { id, model, messages: [], updatedAt: new Date().toISOString() };
 	}
 
 	return session;

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

@@ -5,7 +5,11 @@
 
 	import { loadKnowledge, type Knowledge } from '$lib/knowledge';
 	import { settingsStore, knowledgeStore } from '$lib/store';
-	import { ollamaGenerate, type OllamaCompletionResponse } from '$lib/ollama';
+	import {
+		ollamaGenerate,
+		type OllamaCompletionRequest,
+		type OllamaCompletionResponse
+	} from '$lib/ollama';
 	import {
 		saveSession,
 		type Message,
@@ -42,6 +46,7 @@
 	let prompt: string;
 	let promptCached: string;
 	let promptTextarea: HTMLTextAreaElement;
+	let tokenizedContext: number[];
 	let isPromptFullscreen = false;
 	let shouldFocusTextarea = false;
 
@@ -55,52 +60,6 @@
 	$: knowledge = knowledgeId ? loadKnowledge(knowledgeId) : null;
 	$: shouldFocusTextarea = !isPromptFullscreen;
 
-	afterUpdate(() => {
-		if (shouldFocusTextarea && promptTextarea) {
-			promptTextarea.focus();
-			shouldFocusTextarea = false;
-		}
-	});
-
-	async function scrollToBottom() {
-		if (!messageWindow) return;
-		await tick();
-		messageWindow.scrollTop = messageWindow.scrollHeight;
-	}
-
-	function handleError(error: Error) {
-		resetPrompt();
-
-		let content: string;
-		if (error.message === 'Failed to fetch') {
-			content = `Couldn't connect to Ollama. Is the [server running](/settings)?`;
-		} else {
-			content = `Sorry, something went wrong.\n\`\`\`\n${error}\n\`\`\``;
-		}
-
-		const message: Message = { role: 'system', content };
-		session.messages = [...session.messages, message];
-	}
-
-	async function handleCompletionDone(completion: string, context: number[]) {
-		const message: Message = { role: 'ai', content: completion };
-		session.messages = [...session.messages, message];
-		session.updatedAt = new Date().toISOString();
-
-		if (knowledge) {
-			session.knowledge = knowledge;
-
-			// Now that we used the knowledge, we no longer need an `id`
-			// This will prevent `knowledge` from being used again
-			knowledgeId = '';
-		}
-
-		completion = '';
-		promptCached = '';
-		shouldFocusTextarea = true;
-		saveSession({ ...session, context });
-	}
-
 	async function handleSubmit() {
 		if (!prompt) return;
 
@@ -117,16 +76,48 @@
 		}
 
 		const message: Message = { role: 'user', content: prompt };
-		abortController = new AbortController();
 		promptCached = prompt;
 		prompt = '';
-		completion = '';
 		session.messages = knowledgeContext
 			? [knowledgeContext, ...session.messages, message]
 			: [...session.messages, message];
 
+		const previousAiResponse = session.messages[session.messages.length - 2];
+		let payload = {
+			model: session.model,
+			context: previousAiResponse?.context,
+			prompt: session.messages[session.messages.length - 1].content,
+			system: previousAiResponse?.knowledge?.content
+		};
+
+		await handleCompletion(payload);
+	}
+
+	async function handleRetry(index: number) {
+		// Remove all the messages after the index
+		session.messages = session.messages.slice(0, index);
+
+		const mostRecentUserMessage = session.messages.filter((m) => m.role === 'user').at(-1);
+		const mostRecentSystemMessage = session.messages.filter((m) => m.role === 'system').at(-1);
+		if (!mostRecentUserMessage) throw new Error('No user message to retry');
+
+		let payload = {
+			model: session.model,
+			context: session.messages[index - 2]?.context, // Last AI response
+			prompt: mostRecentUserMessage.content,
+			system: mostRecentSystemMessage?.knowledge?.content
+		};
+
+		await handleCompletion(payload);
+	}
+
+	async function handleCompletion(payload: OllamaCompletionRequest) {
+		abortController = new AbortController();
+		completion = '';
+		tokenizedContext = [];
+
 		try {
-			const ollama = await ollamaGenerate(session, abortController.signal);
+			const ollama = await ollamaGenerate(payload, abortController.signal);
 
 			if (ollama && ollama.body) {
 				const reader = ollama.body.pipeThrough(new TextDecoderStream()).getReader();
@@ -137,7 +128,8 @@
 					if (!ollama.ok && value) throw new Error(JSON.parse(value).error);
 
 					if (done) {
-						handleCompletionDone(completion, session.context);
+						if (!tokenizedContext) throw new Error('Ollama response is missing context');
+						handleCompletionDone(completion, tokenizedContext);
 						break;
 					}
 
@@ -147,7 +139,7 @@
 					for (const line of jsonLines) {
 						const { response, context } = JSON.parse(line) as OllamaCompletionResponse;
 						completion += response;
-						session.context = context;
+						tokenizedContext = context;
 					}
 				}
 			}
@@ -157,6 +149,27 @@
 		}
 	}
 
+	async function handleCompletionDone(completion: string, context: number[]) {
+		abortController = new AbortController();
+
+		const message: Message = { role: 'ai', content: completion, context };
+		session.messages = [...session.messages, message];
+		session.updatedAt = new Date().toISOString();
+
+		if (knowledge) {
+			session.knowledge = knowledge;
+
+			// Now that we used the knowledge, we no longer need an `id`
+			// This will prevent `knowledge` from being used again
+			knowledgeId = '';
+		}
+
+		completion = '';
+		promptCached = '';
+		shouldFocusTextarea = true;
+		saveSession({ ...session });
+	}
+
 	function resetPrompt() {
 		// Reset the prompt to the last sent message
 		prompt = promptCached;
@@ -171,6 +184,33 @@
 		event.preventDefault();
 		handleSubmit();
 	}
+
+	async function scrollToBottom() {
+		if (!messageWindow) return;
+		await tick();
+		messageWindow.scrollTop = messageWindow.scrollHeight;
+	}
+
+	function handleError(error: Error) {
+		resetPrompt();
+
+		let content: string;
+		if (error.message === 'Failed to fetch') {
+			content = `Couldn't connect to Ollama. Is the [server running](/settings)?`;
+		} else {
+			content = `Sorry, something went wrong.\n\`\`\`\n${error}\n\`\`\``;
+		}
+
+		const message: Message = { role: 'system', content };
+		session.messages = [...session.messages, message];
+	}
+
+	afterUpdate(() => {
+		if (shouldFocusTextarea && promptTextarea) {
+			promptTextarea.focus();
+			shouldFocusTextarea = false;
+		}
+	});
 </script>
 
 <div class="session">
@@ -200,7 +240,7 @@
 
 				{#each session.messages as message, i (session.id + i)}
 					{#key message.role}
-						<Article {message} />
+						<Article {message} retryIndex={message.role === 'ai' ? i : undefined} {handleRetry} />
 					{/key}
 				{/each}
 
@@ -224,7 +264,7 @@
 								<FieldSelectModel />
 								<div class="prompt-editor__knowledge">
 									<FieldSelect
-										label="Knowledge"
+										label="System prompt"
 										name="knowledge"
 										disabled={!$knowledgeStore}
 										options={$knowledgeStore?.map((k) => ({ value: k.id, option: k.name }))}