Merge branch 'main' into embed_assistant

.env

@@ -132,6 +132,7 @@ PARQUET_EXPORT_SECRET=#DEPRECATED, use ADMIN_API_SECRET instead
 
 RATE_LIMIT= # /!\ Legacy definition of messages per minute. Use USAGE_LIMITS.messagesPerMinute instead
 MESSAGES_BEFORE_LOGIN=# how many messages a user can send in a conversation before having to login. set to 0 to force login right away
+PUBLIC_APP_GUEST_MESSAGE=# a message to the guest user. If not set, a default message will be used
 
 APP_BASE="" # base path of the app, e.g. /chat, left blank as default
 PUBLIC_APP_NAME=ChatUI # name used as title throughout the app

chart/env/prod.yaml

@@ -285,6 +285,7 @@ envVars:
   PUBLIC_APP_COLOR: "yellow"
   PUBLIC_APP_DESCRIPTION: "Making the community's best AI chat models available to everyone."
   PUBLIC_APP_DISCLAIMER_MESSAGE: "Disclaimer: AI is an area of active research with known problems such as biased generation and misinformation. Do not use this application for high-stakes decisions or advice."
+  PUBLIC_APP_GUEST_MESSAGE: "You have reached the guest message limit, Sign In with a free Hugging Face account to continue using HuggingChat."
   PUBLIC_APP_DATA_SHARING: 0
   PUBLIC_APP_DISCLAIMER: 1
   PUBLIC_PLAUSIBLE_SCRIPT_URL: "/js/script.js"

docs/source/configuration/models/providers/google.md

@@ -33,8 +33,9 @@ MODELS=`[
       "type": "vertex",
       "project": "abc-xyz",
       "location": "europe-west3",
-      "model": "gemini-1.5-pro-preview-0409", // model-name
-
+      "extraBody": {
+         "model_version": "gemini-1.5-pro-002",
+      },
       // Optional
       "safetyThreshold": "BLOCK_MEDIUM_AND_ABOVE",
       "apiEndpoint": "", // alternative api endpoint url,

src/lib/components/LoginModal.svelte

@@ -23,9 +23,7 @@
 			{envPublic.PUBLIC_APP_DESCRIPTION}
 		</p>
 		<p class="text-balance rounded-xl border bg-white/80 p-2 text-base text-gray-800">
-			You have reached the guest message limit, <strong class="font-semibold"
-				>Sign In with a free Hugging Face account</strong
-			> to continue using HuggingChat.
+			{envPublic.PUBLIC_APP_GUEST_MESSAGE}
 		</p>
 
 		<form

src/lib/components/ModelCardMetadata.svelte

@@ -2,9 +2,13 @@
 	import CarbonEarth from "~icons/carbon/earth";
 	import CarbonArrowUpRight from "~icons/carbon/arrow-up-right";
 	import BIMeta from "~icons/bi/meta";
+	import CarbonCode from "~icons/carbon/code";
 	import type { Model } from "$lib/types/Model";
 
-	export let model: Pick<Model, "name" | "datasetName" | "websiteUrl" | "modelUrl" | "datasetUrl">;
+	export let model: Pick<
+		Model,
+		"name" | "datasetName" | "websiteUrl" | "modelUrl" | "datasetUrl" | "hasInferenceAPI"
+	>;
 
 	export let variant: "light" | "dark" = "light";
 </script>
@@ -35,6 +39,16 @@
 			<div class="max-sm:hidden">&nbsp;page</div></a
 		>
 	{/if}
+	{#if model.hasInferenceAPI}
+		<a
+			href={"https://huggingface.co/playground?modelId=" + model.name}
+			target="_blank"
+			rel="noreferrer"
+			class="flex items-center hover:underline"
+			><CarbonCode class="mr-1.5 shrink-0 text-xs text-gray-400" />
+			API
+		</a>
+	{/if}
 	{#if model.websiteUrl}
 		<a
 			href={model.websiteUrl}

src/lib/components/ScrollToPreviousBtn.svelte

@@ -0,0 +1,70 @@
+<script lang="ts">
+	import { fade } from "svelte/transition";
+	import { onDestroy } from "svelte";
+	import IconChevron from "./icons/IconChevron.svelte";
+
+	export let scrollNode: HTMLElement;
+	export { className as class };
+
+	let visible = false;
+	let className = "";
+	let observer: ResizeObserver | null = null;
+
+	$: if (scrollNode) {
+		destroy();
+
+		if (window.ResizeObserver) {
+			observer = new ResizeObserver(() => {
+				updateVisibility();
+			});
+			observer.observe(scrollNode);
+		}
+		scrollNode.addEventListener("scroll", updateVisibility);
+	}
+
+	function updateVisibility() {
+		if (!scrollNode) return;
+		visible =
+			Math.ceil(scrollNode.scrollTop) + 200 < scrollNode.scrollHeight - scrollNode.clientHeight &&
+			scrollNode.scrollTop > 200;
+	}
+
+	function scrollToPrevious() {
+		if (!scrollNode) return;
+		const messages = scrollNode.querySelectorAll('[id^="message-"]');
+		const scrollTop = scrollNode.scrollTop;
+		let previousMessage: Element | null = null;
+
+		for (let i = messages.length - 1; i >= 0; i--) {
+			const messageTop =
+				messages[i].getBoundingClientRect().top +
+				scrollTop -
+				scrollNode.getBoundingClientRect().top;
+			if (messageTop < scrollTop - 1) {
+				previousMessage = messages[i];
+				break;
+			}
+		}
+
+		if (previousMessage) {
+			previousMessage.scrollIntoView({ behavior: "smooth", block: "start" });
+		}
+	}
+
+	function destroy() {
+		observer?.disconnect();
+		scrollNode?.removeEventListener("scroll", updateVisibility);
+	}
+
+	onDestroy(destroy);
+</script>
+
+{#if visible}
+	<button
+		transition:fade={{ duration: 150 }}
+		on:click={scrollToPrevious}
+		class="btn absolute flex h-[41px] w-[41px] rounded-full border bg-white shadow-md transition-all hover:bg-gray-100 dark:border-gray-600 dark:bg-gray-700 dark:shadow-gray-950 dark:hover:bg-gray-600 {className}"
+	>
+		<IconChevron classNames="rotate-180 mt-[2px]" />
+	</button>
+{/if}

src/lib/components/chat/ChatInput.svelte

@@ -1,5 +1,5 @@
 <script lang="ts">
-	import { isDesktop } from "$lib/utils/isDesktop";
+	import { browser } from "$app/environment";
 	import { createEventDispatcher, onMount } from "svelte";
 
 	export let value = "";
@@ -13,25 +13,41 @@
 
 	const dispatch = createEventDispatcher<{ submit: void }>();
 
+	function isVirtualKeyboard(): boolean {
+		if (!browser) return false;
+
+		// Check for touch capability
+		if (navigator.maxTouchPoints > 0) return true;
+
+		// Check for touch events
+		if ("ontouchstart" in window) return true;
+
+		// Fallback to user agent string check
+		const userAgent = navigator.userAgent.toLowerCase();
+
+		return /android|webos|iphone|ipad|ipod|blackberry|iemobile|opera mini/i.test(userAgent);
+	}
+
 	$: minHeight = `${1 + minRows * 1.5}em`;
 	$: maxHeight = maxRows ? `${1 + maxRows * 1.5}em` : `auto`;
 
 	function handleKeydown(event: KeyboardEvent) {
-		// submit on enter
 		if (event.key === "Enter" && !event.shiftKey && !isCompositionOn) {
 			event.preventDefault();
-			// blur to close keyboard on mobile
-			textareaElement.blur();
-			// refocus so that user on desktop can start typing without needing to reclick on textarea
-			if (isDesktop(window)) {
-				textareaElement.focus();
+			if (isVirtualKeyboard()) {
+				// Insert a newline at the cursor position
+				const start = textareaElement.selectionStart;
+				const end = textareaElement.selectionEnd;
+				value = value.substring(0, start) + "\n" + value.substring(end);
+				textareaElement.selectionStart = textareaElement.selectionEnd = start + 1;
+			} else {
+				dispatch("submit");
 			}
-			dispatch("submit"); // use a custom event instead of `event.target.form.requestSubmit()` as it does not work on Safari 14
 		}
 	}
 
 	onMount(() => {
-		if (isDesktop(window)) {
+		if (!isVirtualKeyboard()) {
 			textareaElement.focus();
 		}
 	});
@@ -44,7 +60,7 @@
 		style="min-height: {minHeight}; max-height: {maxHeight}">{(value || " ") + "\n"}</pre>
 
 	<textarea
-		enterkeyhint="send"
+		enterkeyhint={!isVirtualKeyboard() ? "enter" : "send"}
 		tabindex="0"
 		rows="1"
 		class="scrollbar-custom absolute top-0 m-0 h-full w-full resize-none scroll-p-3 overflow-x-hidden overflow-y-scroll border-0 bg-transparent p-3 outline-none focus:ring-0 focus-visible:ring-0"

src/lib/components/chat/ChatMessage.svelte

@@ -230,6 +230,7 @@
 {#if message.from === "assistant"}
 	<div
 		class="group relative -mb-4 flex items-start justify-start gap-4 pb-4 leading-relaxed"
+		id="message-assistant-{message.id}"
 		role="presentation"
 		on:click={() => (isTapped = !isTapped)}
 		on:keydown={() => (isTapped = !isTapped)}
@@ -372,6 +373,7 @@
 {#if message.from === "user"}
 	<div
 		class="group relative w-full items-start justify-start gap-4 max-sm:text-sm"
+		id="message-user-{message.id}"
 		role="presentation"
 		on:click={() => (isTapped = !isTapped)}
 		on:keydown={() => (isTapped = !isTapped)}

src/lib/components/chat/ChatWindow.svelte

@@ -27,6 +27,7 @@
 	import AssistantIntroduction from "./AssistantIntroduction.svelte";
 	import ChatMessage from "./ChatMessage.svelte";
 	import ScrollToBottomBtn from "../ScrollToBottomBtn.svelte";
+	import ScrollToPreviousBtn from "../ScrollToPreviousBtn.svelte";
 	import { browser } from "$app/environment";
 	import { snapScrollToBottom } from "$lib/actions/snapScrollToBottom";
 	import SystemPromptModal from "../SystemPromptModal.svelte";
@@ -329,8 +330,14 @@
 				/>
 			{/if}
 		</div>
+
+		<ScrollToPreviousBtn
+			class="fixed right-4 max-md:bottom-[calc(50%+26px)] md:bottom-48 lg:right-10"
+			scrollNode={chatContainer}
+		/>
+
 		<ScrollToBottomBtn
-			class="bottom-36 right-4 max-md:hidden lg:right-10"
+			class="fixed right-4 max-md:bottom-[calc(50%-26px)] md:bottom-36 lg:right-10"
 			scrollNode={chatContainer}
 		/>
 	</div>

src/lib/server/endpoints/cohere/endpointCohere.ts

@@ -62,6 +62,7 @@ export async function endpointCohere(
 				});
 
 				stream = await cohere.chatStream({
+					forceSingleStep: true,
 					message: prompt,
 					rawPrompting: true,
 					model: model.id ?? model.name,
@@ -82,6 +83,7 @@ export async function endpointCohere(
 
 				stream = await cohere
 					.chatStream({
+						forceSingleStep: true,
 						model: model.id ?? model.name,
 						chatHistory: formattedMessages.slice(0, -1),
 						message: formattedMessages[formattedMessages.length - 1].message,

src/lib/server/endpoints/google/endpointVertex.ts

@@ -16,6 +16,7 @@ export const endpointVertexParametersSchema = z.object({
 	model: z.any(), // allow optional and validate against emptiness
 	type: z.literal("vertex"),
 	location: z.string().default("europe-west1"),
+	extraBody: z.object({ model_version: z.string() }).optional(),
 	project: z.string(),
 	apiEndpoint: z.string().optional(),
 	safetyThreshold: z
@@ -49,7 +50,7 @@ export const endpointVertexParametersSchema = z.object({
 });
 
 export function endpointVertex(input: z.input<typeof endpointVertexParametersSchema>): Endpoint {
-	const { project, location, model, apiEndpoint, safetyThreshold, tools, multimodal } =
+	const { project, location, model, apiEndpoint, safetyThreshold, tools, multimodal, extraBody } =
 		endpointVertexParametersSchema.parse(input);
 
 	const vertex_ai = new VertexAI({
@@ -64,7 +65,7 @@ export function endpointVertex(input: z.input<typeof endpointVertexParametersSch
 		const hasFiles = messages.some((message) => message.files && message.files.length > 0);
 
 		const generativeModel = vertex_ai.getGenerativeModel({
-			model: model.id ?? model.name,
+			model: extraBody?.model_version ?? model.id ?? model.name,
 			safetySettings: safetyThreshold
 				? [
 						{

src/lib/server/endpoints/tgi/endpointTgi.ts

@@ -23,8 +23,8 @@ export const endpointTgiParametersSchema = z.object({
 				supportedMimeTypes: ["image/jpeg", "image/webp"],
 				preferredMimeType: "image/webp",
 				maxSizeInMB: 5,
-				maxWidth: 224,
-				maxHeight: 224,
+				maxWidth: 378,
+				maxHeight: 980,
 			}),
 		})
 		.default({}),
@@ -81,22 +81,13 @@ export function endpointTgi(input: z.input<typeof endpointTgiParametersSchema>):
 	};
 }
 
-const whiteImage = {
-	mime: "image/png",
-	image: Buffer.from(
-		"/9j/4AAQSkZJRgABAQAAAQABAAD/2wBDAAEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQH/2wBDAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQEBAQH/wAARCAAQABADAREAAhEBAxEB/8QAHwAAAQUBAQEBAQEAAAAAAAAAAAECAwQFBgcICQoL/8QAtRAAAgEDAwIEAwUFBAQAAAF9AQIDAAQRBRIhMUEGE1FhByJxFDKBkaEII0KxwRVS0fAkM2JyggkKFhcYGRolJicoKSo0NTY3ODk6Q0RFRkdISUpTVFVWV1hZWmNkZWZnaGlqc3R1dnd4eXqDhIWGh4iJipKTlJWWl5iZmqKjpKWmp6ipqrKztLW2t7i5usLDxMXGx8jJytLT1NXW19jZ2uHi4+Tl5ufo6erx8vP09fb3+Pn6/8QAHwEAAwEBAQEBAQEBAQAAAAAAAAECAwQFBgcICQoL/8QAtREAAgECBAQDBAcFBAQAAQJ3AAECAxEEBSExBhJBUQdhcRMiMoEIFEKRobHBCSMzUvAVYnLRChYkNOEl8RcYGRomJygpKjU2Nzg5OkNERUZHSElKU1RVVldYWVpjZGVmZ2hpanN0dXZ3eHl6goOEhYaHiImKkpOUlZaXmJmaoqOkpaanqKmqsrO0tba3uLm6wsPExcbHyMnK0tPU1dbX2Nna4uPk5ebn6Onq8vP09fb3+Pn6/9oADAMBAAIRAxEAPwD+/igAoAKACgD/2Q==",
-		"base64"
-	),
-};
-
 async function prepareMessage(
 	isMultimodal: boolean,
 	message: EndpointMessage,
 	imageProcessor: ImageProcessor
 ): Promise<EndpointMessage> {
 	if (!isMultimodal) return message;
-
-	const files = await Promise.all(message.files?.map(imageProcessor) ?? [whiteImage]);
+	const files = await Promise.all(message.files?.map(imageProcessor) ?? []);
 	const markdowns = files.map(
 		(file) => `![](data:${file.mime};base64,${file.image.toString("base64")})`
 	);

src/lib/server/models.ts

@@ -13,6 +13,7 @@ import JSON5 from "json5";
 import { getTokenizer } from "$lib/utils/getTokenizer";
 import { logger } from "$lib/server/logger";
 import { ToolResultStatus, type ToolInput } from "$lib/types/Tool";
+import { isHuggingChat } from "$lib/utils/isHuggingChat";
 
 type Optional<T, K extends keyof T> = Pick<Partial<T>, K> & Omit<T, K>;
 
@@ -253,10 +254,6 @@ const processModel = async (m: z.infer<typeof modelConfig>) => ({
 	parameters: { ...m.parameters, stop_sequences: m.parameters?.stop },
 });
 
-export type ProcessedModel = Awaited<ReturnType<typeof processModel>> & {
-	getEndpoint: () => Promise<Endpoint>;
-};
-
 const addEndpoint = (m: Awaited<ReturnType<typeof processModel>>) => ({
 	...m,
 	getEndpoint: async (): Promise<Endpoint> => {
@@ -316,10 +313,40 @@ const addEndpoint = (m: Awaited<ReturnType<typeof processModel>>) => ({
 	},
 });
 
-export const models: ProcessedModel[] = await Promise.all(
-	modelsRaw.map((e) => processModel(e).then(addEndpoint))
+const hasInferenceAPI = async (m: Awaited<ReturnType<typeof processModel>>) => {
+	if (!isHuggingChat) {
+		return false;
+	}
+
+	const r = await fetch(`https://huggingface.co/api/models/${m.id}`);
+
+	if (!r.ok) {
+		logger.warn(`Failed to check if ${m.id} has inference API: ${r.statusText}`);
+		return false;
+	}
+
+	const json = await r.json();
+
+	if (json.cardData.inference === false) {
+		return false;
+	}
+
+	return true;
+};
+
+export const models = await Promise.all(
+	modelsRaw.map((e) =>
+		processModel(e)
+			.then(addEndpoint)
+			.then(async (m) => ({
+				...m,
+				hasInferenceAPI: await hasInferenceAPI(m),
+			}))
+	)
 );
 
+export type ProcessedModel = (typeof models)[number];
+
 // super ugly but not sure how to make typescript happier
 export const validModelIdSchema = z.enum(models.map((m) => m.id) as [string, ...string[]]);
 
@@ -357,5 +384,5 @@ export const smallModel = env.TASK_MODEL
 
 export type BackendModel = Optional<
 	typeof defaultModel,
-	"preprompt" | "parameters" | "multimodal" | "unlisted" | "tools"
+	"preprompt" | "parameters" | "multimodal" | "unlisted" | "tools" | "hasInferenceAPI"
 >;