feat: sanitize titles by stripping <think> markers across multiple components and endpoints

Author: gary149

scira-chat

@@ -0,0 +1 @@
+Subproject commit 892ec4a60d23cd1c19b1a079c799d24a8a79a262

src/lib/components/chat/OpenReasoningResults.svelte

@@ -18,7 +18,7 @@
 
 <details
 	bind:open={isOpen}
-	class="group flex w-fit max-w-full flex-col rounded-xl border border-gray-200 bg-white shadow-sm dark:border-gray-800 dark:bg-gray-900"
+	class="group mb-4 flex w-fit max-w-full flex-col rounded-xl border border-gray-200 bg-white shadow-sm dark:border-gray-800 dark:bg-gray-900"
 >
 	<summary
 		class="
@@ -69,7 +69,7 @@
 	</summary>
 
 	<div
-		class="space-y-4 border-t border-gray-200 px-5 pb-2 pt-2 text-sm text-gray-600 dark:border-gray-800 dark:text-gray-400"
+		class="space-y-4 border-t border-gray-200 p-3 text-sm text-gray-600 dark:border-gray-800 dark:text-gray-400"
 	>
 		{#key content}
 			<MarkdownRenderer {content} />

src/lib/server/api/routes/groups/conversations.ts

@@ -172,9 +172,12 @@ export const conversationGroup = new Elysia().use(authPlugin).group("/conversati
 								}
 							}
 
-							// Only include defined values in the update
+
+							// Only include defined values in the update (sanitize title)
 							const updateValues = {
-								...(body.title !== undefined && { title: body.title }),
+								...(body.title !== undefined && {
+									title: body.title.replace(/<\/?think>/gi, "").trim(),
+								}),
 								...(body.model !== undefined && { model: body.model }),
 							};
 

src/lib/server/generateFromDefaultEndpoint.ts

@@ -1,25 +1,30 @@
-import { taskModel } from "$lib/server/models";
+import { taskModel, models } from "$lib/server/models";
 import { MessageUpdateType, type MessageUpdate } from "$lib/types/MessageUpdate";
 import type { EndpointMessage } from "./endpoints/endpoints";
 
 export async function* generateFromDefaultEndpoint({
 	messages,
 	preprompt,
 	generateSettings,
+	modelId,
 }: {
 	messages: EndpointMessage[];
 	preprompt?: string;
 	generateSettings?: Record<string, unknown>;
+	/** Optional: use this model instead of the default task model */
+	modelId?: string;
 }): AsyncGenerator<MessageUpdate, string, undefined> {
 	try {
-		const endpoint = await taskModel.getEndpoint();
+		// Choose endpoint based on provided modelId, else fall back to taskModel
+		const model = modelId ? (models.find((m) => m.id === modelId) ?? taskModel) : taskModel;
+		const endpoint = await model.getEndpoint();
 		const tokenStream = await endpoint({ messages, preprompt, generateSettings });
 
 		for await (const output of tokenStream) {
 			// if not generated_text is here it means the generation is not done
 			if (output.generated_text) {
 				let generated_text = output.generated_text;
-				for (const stop of [...(taskModel.parameters?.stop ?? []), "<|endoftext|>"]) {
+				for (const stop of [...(model.parameters?.stop ?? []), "<|endoftext|>"]) {
 					if (generated_text.endsWith(stop)) {
 						generated_text = generated_text.slice(0, -stop.length).trimEnd();
 					}

src/lib/server/textGeneration/generate.ts

@@ -99,6 +99,7 @@ Do not use prefixes such as Response: or Answer: when answering to the user.`,
 						generateSettings: {
 							max_new_tokens: 1024,
 						},
+						modelId: model.id,
 					});
 					finalAnswer = summary;
 					yield {
@@ -210,7 +211,7 @@ Do not use prefixes such as Response: or Answer: when answering to the user.`,
 			) {
 				lastReasoningUpdate = new Date();
 				try {
-					generateSummaryOfReasoning(reasoningBuffer).then((summary) => {
+					generateSummaryOfReasoning(reasoningBuffer, model.id).then((summary) => {
 						status = summary;
 					});
 				} catch (e) {

src/lib/server/textGeneration/reasoning.ts

@@ -1,10 +1,9 @@
 import { generateFromDefaultEndpoint } from "../generateFromDefaultEndpoint";
-import { taskModel } from "../models";
 import { getReturnFromGenerator } from "$lib/utils/getReturnFromGenerator";
 import { logger } from "../logger";
 
-export async function generateSummaryOfReasoning(buffer: string): Promise<string> {
-	let summary: string | undefined;
+export async function generateSummaryOfReasoning(buffer: string, modelId?: string): Promise<string> {
+    let summary: string | undefined;
 
 	const messages = [
 		{
@@ -18,24 +17,25 @@ The text might be incomplete, try your best to summarize it in one very short se
 
 	// Tools removed: no tool-based summarization path
 
-	if (!summary) {
-		summary = await getReturnFromGenerator(
-			generateFromDefaultEndpoint({
-				messages: [
-					{
-						from: "user",
-						content: buffer.slice(-300),
-					},
-				],
-				preprompt: `You are tasked with summarizing the latest reasoning steps. Never describe results of the reasoning, only the process. Remain vague in your summary.
+    if (!summary) {
+        summary = await getReturnFromGenerator(
+            generateFromDefaultEndpoint({
+                messages: [
+                    {
+                        from: "user",
+                        content: buffer.slice(-300),
+                    },
+                ],
+                preprompt: `You are tasked with summarizing the latest reasoning steps. Never describe results of the reasoning, only the process. Remain vague in your summary.
             The text might be incomplete, try your best to summarize it in one very short sentence, starting with a gerund and ending with three points. 
             Example: "Thinking about life...", "Summarizing the results...", "Processing the input..."`,
-				generateSettings: {
-					max_new_tokens: 50,
-				},
-			})
-		);
-	}
+                generateSettings: {
+                    max_new_tokens: 50,
+                },
+                modelId,
+            })
+        );
+    }
 
 	if (!summary) {
 		return "Reasoning...";

src/lib/server/textGeneration/title.ts

@@ -4,7 +4,7 @@ import { logger } from "$lib/server/logger";
 import { MessageUpdateType, type MessageUpdate } from "$lib/types/MessageUpdate";
 import type { Conversation } from "$lib/types/Conversation";
 import { getReturnFromGenerator } from "$lib/utils/getReturnFromGenerator";
-import { taskModel } from "../models";
+// taskModel no longer used directly here; we pass the current model instead
 
 export async function* generateTitleForConversation(
 	conv: Conversation
@@ -15,7 +15,7 @@ export async function* generateTitleForConversation(
 		if (conv.title !== "New Chat" || !userMessage) return;
 
 		const prompt = userMessage.content;
-		const title = (await generateTitle(prompt)) ?? "New Chat";
+		const title = (await generateTitle(prompt, conv.model)) ?? "New Chat";
 
 		yield {
 			type: MessageUpdateType.Title,
@@ -26,32 +26,40 @@ export async function* generateTitleForConversation(
 	}
 }
 
-export async function generateTitle(prompt: string) {
-	if (config.LLM_SUMMARIZATION !== "true") {
-		return prompt.split(/\s+/g).slice(0, 5).join(" ");
-	}
+export async function generateTitle(prompt: string, modelId?: string) {
+    if (config.LLM_SUMMARIZATION !== "true") {
+        return prompt.split(/\s+/g).slice(0, 5).join(" ");
+    }
 
 	// Tools removed: no tool-based title path
 
-	return await getReturnFromGenerator(
-		generateFromDefaultEndpoint({
-			messages: [{ from: "user", content: prompt }],
-			preprompt:
-				"You are a summarization AI. Summarize the user's request into a single short sentence of four words or less. Do not try to answer it, only summarize the user's query. Always start your answer with an emoji relevant to the summary",
-			generateSettings: {
-				max_new_tokens: 30,
-			},
-		})
-	)
+    return await getReturnFromGenerator(
+        generateFromDefaultEndpoint({
+            messages: [{ from: "user", content: prompt }],
+            preprompt:
+                "You are a summarization AI. Summarize the user's request into a single short sentence of four words or less. Do not try to answer it, only summarize the user's query. Always start your answer with an emoji relevant to the summary",
+            generateSettings: {
+                max_new_tokens: 30,
+            },
+            modelId,
+        })
+    )
 		.then((summary) => {
-			// add an emoji if none is found in the first three characters
-			if (!/\p{Emoji}/u.test(summary.slice(0, 3))) {
-				return "💬 " + summary;
+			const firstFive = prompt.split(/\s+/g).slice(0, 5).join(" ");
+			const trimmed = summary.trim();
+			// Fallback: if empty, use emoji + first five words
+			if (!trimmed) {
+				return "💬 " + firstFive;
+			}
+			// Ensure emoji prefix if missing
+			if (!/\p{Emoji}/u.test(trimmed.slice(0, 3))) {
+				return "💬 " + trimmed;
 			}
-			return summary;
+			return trimmed;
 		})
 		.catch((e) => {
 			logger.error(e);
-			return null;
+			const firstFive = prompt.split(/\s+/g).slice(0, 5).join(" ");
+			return "💬 " + firstFive;
 		});
 }

src/routes/+layout.svelte

@@ -78,19 +78,21 @@
 			});
 	}
 
-	async function editConversationTitle(id: string, title: string) {
-		client
-			.conversations({ id })
-			.patch({ title })
-			.then(handleResponse)
-			.then(async () => {
-				conversations = conversations.map((conv) => (conv.id === id ? { ...conv, title } : conv));
-			})
-			.catch((err) => {
-				console.error(err);
-				$error = String(err);
-			});
-	}
+    async function editConversationTitle(id: string, title: string) {
+        // Always strip <think> markers before saving and updating UI
+        const cleanTitle = title.replace(/<\/?think>/gi, "").trim();
+        client
+            .conversations({ id })
+            .patch({ title: cleanTitle })
+            .then(handleResponse)
+            .then(async () => {
+                conversations = conversations.map((conv) => (conv.id === id ? { ...conv, title: cleanTitle } : conv));
+            })
+            .catch((err) => {
+                console.error(err);
+                $error = String(err);
+            });
+    }
 
 	onDestroy(() => {
 		clearTimeout(errorToastTimeout);

src/routes/+layout.ts

@@ -27,6 +27,9 @@ export const load = async ({ depends, fetch, url }) => {
 			conv.title = conv.title.replace(/\p{Emoji}/gu, "");
 		}
 
+		// Always strip <think> markers from titles for sidebar display
+		conv.title = conv.title.replace(/<\/?think>/gi, "");
+
 		// remove invalid unicode and trim whitespaces
 		conv.title = conv.title.replace(/\uFFFD/gu, "").trimStart();
 

src/routes/conversation/+server.ts

@@ -63,7 +63,8 @@ export const POST: RequestHandler = async ({ locals, request }) => {
 			error(404, "Conversation not found");
 		}
 
-		title = conversation.title;
+		// Strip <think> markers from imported titles
+		title = conversation.title.replace(/<\/?think>/gi, "").trim();
 		messages = conversation.messages;
 		rootMessageId = conversation.rootMessageId ?? rootMessageId;
 		values.model = conversation.model;
@@ -83,7 +84,8 @@ export const POST: RequestHandler = async ({ locals, request }) => {
 
 	const res = await collections.conversations.insertOne({
 		_id: new ObjectId(),
-		title: title || "New Chat",
+		// Always store sanitized titles
+		title: (title || "New Chat").replace(/<\/?think>/gi, "").trim(),
 		rootMessageId,
 		messages,
 		model: values.model,