Merge pull request #1630 from cheahjs/feat/split-large-chunks

feat: split large openai responses into smaller chunks

src/lib/apis/streaming/index.ts

@@ -0,0 +1,70 @@
+type TextStreamUpdate = {
+	done: boolean;
+	value: string;
+};
+
+// createOpenAITextStream takes a ReadableStreamDefaultReader from an SSE response,
+// and returns an async generator that emits delta updates with large deltas chunked into random sized chunks
+export async function createOpenAITextStream(
+	messageStream: ReadableStreamDefaultReader,
+	splitLargeDeltas: boolean
+): Promise<AsyncGenerator<TextStreamUpdate>> {
+	let iterator = openAIStreamToIterator(messageStream);
+	if (splitLargeDeltas) {
+		iterator = streamLargeDeltasAsRandomChunks(iterator);
+	}
+	return iterator;
+}
+
+async function* openAIStreamToIterator(
+	reader: ReadableStreamDefaultReader
+): AsyncGenerator<TextStreamUpdate> {
+	while (true) {
+		const { value, done } = await reader.read();
+		if (done) {
+			yield { done: true, value: '' };
+			break;
+		}
+		const lines = value.split('\n');
+		for (const line of lines) {
+			if (line !== '') {
+				console.log(line);
+				if (line === 'data: [DONE]') {
+					yield { done: true, value: '' };
+				} else {
+					const data = JSON.parse(line.replace(/^data: /, ''));
+					console.log(data);
+
+					yield { done: false, value: data.choices[0].delta.content ?? '' };
+				}
+			}
+		}
+	}
+}
+
+// streamLargeDeltasAsRandomChunks will chunk large deltas (length > 5) into random sized chunks between 1-3 characters
+// This is to simulate a more fluid streaming, even though some providers may send large chunks of text at once
+async function* streamLargeDeltasAsRandomChunks(
+	iterator: AsyncGenerator<TextStreamUpdate>
+): AsyncGenerator<TextStreamUpdate> {
+	for await (const textStreamUpdate of iterator) {
+		if (textStreamUpdate.done) {
+			yield textStreamUpdate;
+			return;
+		}
+		let content = textStreamUpdate.value;
+		if (content.length < 5) {
+			yield { done: false, value: content };
+			continue;
+		}
+		while (content != '') {
+			const chunkSize = Math.min(Math.floor(Math.random() * 3) + 1, content.length);
+			const chunk = content.slice(0, chunkSize);
+			yield { done: false, value: chunk };
+			await sleep(5);
+			content = content.slice(chunkSize);
+		}
+	}
+}
+
+const sleep = (ms: number) => new Promise((resolve) => setTimeout(resolve, ms));

src/lib/components/chat/Settings/Interface.svelte

@@ -17,11 +17,17 @@
 	let titleAutoGenerateModelExternal = '';
 	let fullScreenMode = false;
 	let titleGenerationPrompt = '';
+	let splitLargeChunks = false;
 
 	// Interface
 	let promptSuggestions = [];
 	let showUsername = false;
 
+	const toggleSplitLargeChunks = async () => {
+		splitLargeChunks = !splitLargeChunks;
+		saveSettings({ splitLargeChunks: splitLargeChunks });
+	};
+
 	const toggleFullScreenMode = async () => {
 		fullScreenMode = !fullScreenMode;
 		saveSettings({ fullScreenMode: fullScreenMode });
@@ -197,6 +203,28 @@
 					</button>
 				</div>
 			</div>
+
+			<div>
+				<div class=" py-0.5 flex w-full justify-between">
+					<div class=" self-center text-xs font-medium">
+						{$i18n.t('Fluidly stream large external response chunks')}
+					</div>
+
+					<button
+						class="p-1 px-3 text-xs flex rounded transition"
+						on:click={() => {
+							toggleSplitLargeChunks();
+						}}
+						type="button"
+					>
+						{#if splitLargeChunks === true}
+							<span class="ml-2 self-center">{$i18n.t('On')}</span>
+						{:else}
+							<span class="ml-2 self-center">{$i18n.t('Off')}</span>
+						{/if}
+					</button>
+				</div>
+			</div>
 		</div>
 
 		<hr class=" dark:border-gray-700" />

src/lib/i18n/locales/en-US/translation.json

@@ -152,6 +152,7 @@
 	"File Mode": "",
 	"File not found.": "",
 	"Fingerprint spoofing detected: Unable to use initials as avatar. Defaulting to default profile image.": "",
+	"Fluidly stream large external response chunks": "",
 	"Focus chat input": "",
 	"Format your variables using square brackets like this:": "",
 	"From (Base Model)": "",

src/routes/(app)/+page.svelte

@@ -39,6 +39,7 @@
 	import { RAGTemplate } from '$lib/utils/rag';
 	import { LITELLM_API_BASE_URL, OLLAMA_API_BASE_URL, OPENAI_API_BASE_URL } from '$lib/constants';
 	import { WEBUI_BASE_URL } from '$lib/constants';
+	import { createOpenAITextStream } from '$lib/apis/streaming';
 
 	const i18n = getContext('i18n');
 
@@ -599,38 +600,22 @@
 				.pipeThrough(splitStream('\n'))
 				.getReader();
 
-			while (true) {
+			const textStream = await createOpenAITextStream(reader, $settings.splitLargeChunks);
-				const { value, done } = await reader.read();
+			console.log(textStream);
+
+			for await (const update of textStream) {
+				const { value, done } = update;
 				if (done || stopResponseFlag || _chatId !== $chatId) {
 					responseMessage.done = true;
 					messages = messages;
 					break;
 				}
 
-				try {
+				if (responseMessage.content == '' && value == '\n') {
-					let lines = value.split('\n');
+					continue;
-
+				} else {
-					for (const line of lines) {
+					responseMessage.content += value;
-						if (line !== '') {
+					messages = messages;
-							console.log(line);
-							if (line === 'data: [DONE]') {
-								responseMessage.done = true;
-								messages = messages;
-							} else {
-								let data = JSON.parse(line.replace(/^data: /, ''));
-								console.log(data);
-
-								if (responseMessage.content == '' && data.choices[0].delta.content == '\n') {
-									continue;
-								} else {
-									responseMessage.content += data.choices[0].delta.content ?? '';
-									messages = messages;
-								}
-							}
-						}
-					}
-				} catch (error) {
-					console.log(error);
 				}
 
 				if ($settings.notificationEnabled && !document.hasFocus()) {

src/routes/(app)/c/[id]/+page.svelte

@@ -42,6 +42,7 @@
 		OLLAMA_API_BASE_URL,
 		WEBUI_BASE_URL
 	} from '$lib/constants';
+	import { createOpenAITextStream } from '$lib/apis/streaming';
 
 	const i18n = getContext('i18n');
 
@@ -611,38 +612,22 @@
 				.pipeThrough(splitStream('\n'))
 				.getReader();
 
-			while (true) {
+			const textStream = await createOpenAITextStream(reader, $settings.splitLargeChunks);
-				const { value, done } = await reader.read();
+			console.log(textStream);
+
+			for await (const update of textStream) {
+				const { value, done } = update;
 				if (done || stopResponseFlag || _chatId !== $chatId) {
 					responseMessage.done = true;
 					messages = messages;
 					break;
 				}
 
-				try {
+				if (responseMessage.content == '' && value == '\n') {
-					let lines = value.split('\n');
+					continue;
-
+				} else {
-					for (const line of lines) {
+					responseMessage.content += value;
-						if (line !== '') {
+					messages = messages;
-							console.log(line);
-							if (line === 'data: [DONE]') {
-								responseMessage.done = true;
-								messages = messages;
-							} else {
-								let data = JSON.parse(line.replace(/^data: /, ''));
-								console.log(data);
-
-								if (responseMessage.content == '' && data.choices[0].delta.content == '\n') {
-									continue;
-								} else {
-									responseMessage.content += data.choices[0].delta.content ?? '';
-									messages = messages;
-								}
-							}
-						}
-					}
-				} catch (error) {
-					console.log(error);
 				}
 
 				if ($settings.notificationEnabled && !document.hasFocus()) {