use llama-3.2-90b-text-preview and llama-3.2-90b-vision-preview

main.ts

@@ -15,7 +15,11 @@ const system = `\
 全て鳥取弁で語尾が「にゃん」。`;
 
 const n = 10;
-const model = "llama-3.1-70b-versatile";
+const textModel = "llama-3.2-90b-text-preview";
+const visionModel = "llama-3.2-90b-vision-preview";
+const urlPattern = new URLPattern({ pathname: "*.*" });
+const separatorRegex = /[<>]|[^\p{L}\p{N}\p{P}\p{S}]+/u;
+const imageExtensionRegex = /[.](?:gif|png|jpe?g|webp)$/i;
 const usage = `\
 https://dash.deno.com/playground/darazllm
 
@@ -25,7 +29,10 @@ https://dash.deno.com/playground/darazllm
     @darazllm /help     このテキストを表示
 
 使用する会話: 最新${n}件
-モデル: ${model}
+テキスト用モデル: ${textModel}
+画像用モデル: ${visionModel}
+区切り文字: ${separatorRegex}
+画像拡張子: ${imageExtensionRegex}
 システムプロンプト:
 
 ${system}`;
@@ -43,13 +50,36 @@ const kv = await Deno.openKv();
 
 async function chat(messages: Messages): Promise<string | null> {
   const res = await groq.chat.completions.create({
-    model,
+    model: textModel,
     messages: [{ role: "system", content: system }, ...messages],
   });
 
   return res.choices[0].message.content;
 }
 
+function isImageUrl(message: string): boolean {
+  return urlPattern.test(message) && imageExtensionRegex.test(message);
+}
+
+/** llama-3.2-90b-vision-preview はシステムプロンプトに対応してない、かつ、画像は1枚まで */
+async function visionChat(
+  content: Array<Groq.Chat.ChatCompletionContentPart>,
+): Promise<string | null> {
+  const res = await groq.chat.completions.create({
+    model: visionModel,
+    messages: [
+      {
+        role: "user",
+        content: content.map((c) =>
+          c.type === "text" ? { ...c, text: `${system}\n\n${c.text}` } : c
+        ),
+      },
+    ],
+  });
+
+  return res.choices[0].message.content;
+}
+
 const app = new bolt.App({
   token: SLACK_BOT_TOKEN,
   signingSecret: SLACK_SIGNING_SECRET,
@@ -80,8 +110,23 @@ app.message(async (c) => {
   if (prompt) messages.push({ role: "user", content: prompt });
 
   if (isMention || Math.floor(Math.random() * n) === 0) {
+    const content: Array<Groq.Chat.ChatCompletionContentPart> = prompt
+      .split(separatorRegex)
+      .filter(Boolean)
+      .map((text) => {
+        if (isImageUrl(text)) {
+          return { type: "image_url", image_url: { url: text } };
+        } else {
+          return { type: "text", text };
+        }
+      });
+
+    const visionMode = content.some((c) => c.type === "image_url");
+
     try {
-      const res = await chat(messages.slice(-n));
+      const res = visionMode
+        ? await visionChat(content)
+        : await chat(messages.slice(-n));
 
       if (!res) throw new Error("Empty response");