.env.example

@@ -6,7 +6,9 @@ botSource = "https://github.com/ABOCN/TelegramBot"
 botToken = ""
 
 # ai features
+ollamaEnabled = false
 # ollamaApi = "http://ollama:11434"
+# handlerTimeout = "600_000" # set higher if you expect to download larger models
 
 # misc (botAdmins isnt a array here!)
 maxRetries = 9999

README.md

@@ -10,12 +10,6 @@ Kowalski is a a simple Telegram bot made in Node.js.
 
 - You can find Kowalski at [@KowalskiNodeBot](https://t.me/KowalskiNodeBot) on Telegram.
 
-## Translations
-
-<a href="https://weblate.librecloud.cc/engage/kowalski/">
-<img src="https://weblate.librecloud.cc/widget/kowalski/multi-auto.svg" alt="Translation status" />
-</a>
-
 ## Self-host requirements
 
 > [!IMPORTANT]
@@ -26,7 +20,10 @@ Kowalski is a a simple Telegram bot made in Node.js.
 - FFmpeg (only for the `/yt` command)
 - Docker and Docker Compose (only required for Docker setup)
 
-_AI features require a higher-end system with a CPU/GPU_
+### AI Requirements
+
+- High-end CPU *or* GPU (~ 6GB vRAM)
+- If using CPU, enough RAM to load the models (~6GB w/ defaults)
 
 ## Running locally (non-Docker setup)
 
@@ -60,7 +57,7 @@ You can also run Kowalski using Docker, which simplifies the setup process. Make
 1. **Copy compose file**
 
    _Without AI (Ollama)_
-   
+
    ```bash
    mv docker-compose.yml.example docker-compose.yml
    ```
@@ -70,7 +67,7 @@ You can also run Kowalski using Docker, which simplifies the setup process. Make
    ```bash
    mv docker-compose.yml.ai.example docker-compose.yml
    ```
- 
+
 2. **Make sure to setup your `.env` file first!**
 
 3. **Run the container**
@@ -109,7 +106,9 @@ If you prefer to use Docker directly, you can use these instructions instead.
 - **botPrivacy**: Put the link to your bot privacy policy.
 - **maxRetries**: Maximum number of retries for a failing command on Kowalski. Default is 5. If the limit is hit, the bot will crash past this number.
 - **botToken**: Put your bot token that you created at [@BotFather](https://t.me/botfather).
+- **ollamaEnabled** (optional): Enables/disables AI features
 - **ollamaApi** (optional): Ollama API endpoint for various AI features, will be disabled if not set
+- **handlerTimeout** (default: `600_000`): How long handlers will wait before timing out. Set this high if using large AI models.
 - **botAdmins**: Put the ID of the people responsible for managing the bot. They can use some administrative + exclusive commands on any group.
 - **lastKey**: Last.fm API key, for use on `lastfm.js` functions, like see who is listening to what song and etc.
 - **weatherKey**: Weather.com API key, used for the `/weather` command.

src/bot.ts

@@ -4,6 +4,7 @@ import fs from 'fs';
 import { isOnSpamWatch } from './spamwatch/spamwatch';
 import '@dotenvx/dotenvx';
 import './plugins/ytDlpWrapper';
+import { preChecks } from './commands/ai';
 
 // Ensures bot token is set, and not default value
 if (!process.env.botToken || process.env.botToken === 'InsertYourBotTokenHere') {
@@ -11,7 +12,17 @@ if (!process.env.botToken || process.env.botToken === 'InsertYourBotTokenHere')
   process.exit(1)
 }
 
-const bot = new Telegraf(process.env.botToken);
+// Detect AI and run pre-checks
+if (process.env.ollamaEnabled === "true") {
+  if (!(await preChecks())) {
+    process.exit(1)
+  }
+}
+
+const bot = new Telegraf(
+  process.env.botToken,
+  { handlerTimeout: Number(process.env.handlerTimeout) || 600_000 }
+);
 const maxRetries = process.env.maxRetries || 5;
 let restartCount = 0;
 
@@ -21,7 +32,7 @@ const loadCommands = () => {
   try {
     const files = fs.readdirSync(commandsPath)
       .filter(file => file.endsWith('.ts') || file.endsWith('.js'));
-    
+
     files.forEach((file) => {
       try {
         const commandPath = path.join(commandsPath, file);

src/commands/ai.ts

@@ -40,8 +40,8 @@ import { rateLimiter } from "../utils/rate-limiter"
 import { logger } from "../utils/log"
 
 const spamwatchMiddleware = spamwatchMiddlewareModule(isOnSpamWatch)
-//const model = "qwen3:0.6b"
-const model = "deepseek-r1:1.5b"
+export const flash_model = "gemma3:4b"
+export const thinking_model = "deepseek-r1:1.5b"
 
 type TextContext = Context & { message: Message.TextMessage }
 
@@ -54,7 +54,22 @@ export function sanitizeForJson(text: string): string {
     .replace(/\t/g, '\\t')
 }
 
-async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Message) {
+export async function preChecks() {
+  const envs = [
+    "ollamaApi",
+  ]
+
+  for (const env of envs) {
+    if (!process.env[env]) {
+      console.error(`[✨ AI | !] ❌ ${env} not set!`)
+      return false
+    }
+  }
+  console.log("[✨ AI] Pre-checks passed\n")
+  return true
+}
+
+async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Message, model: string) {
   const Strings = getStrings(languageCode(ctx))
 
   if (!ctx.chat) return {
@@ -74,23 +89,50 @@ async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Me
     let fullResponse = ""
     let thoughts = ""
     let lastUpdate = Date.now()
-    
+
     for await (const chunk of aiResponse.data) {
       const lines = chunk.toString().split('\n')
       for (const line of lines) {
         if (!line.trim()) continue
         let ln = JSON.parse(line)
-        
-        if (ln.response.includes("<think>")) { logger.logThinking(true) } else if (ln.response.includes("</think>")) { logger.logThinking(false) }
+
+        if (model === thinking_model && ln.response.includes('<think>')) {
+          const thinkMatch = ln.response.match(/<think>([\s\S]*?)<\/think>/)
+          if (thinkMatch) {
+            const innerContent = thinkMatch[1]
+            if (innerContent.trim().length > 0) {
+              logger.logThinking(ctx.chat.id, replyGenerating.message_id, true)
+            }
+          } else {
+            logger.logThinking(ctx.chat.id, replyGenerating.message_id, true)
+          }
+        } else if (model === thinking_model && ln.response.includes('</think>')) {
+          logger.logThinking(ctx.chat.id, replyGenerating.message_id, false)
+        }
 
         try {
           const now = Date.now()
-          
-          if (ln.response) {
-            const patchedThoughts = ln.response.replace("<think>", "`Thinking...`").replace("</think>", "`Finished thinking`")
-            thoughts += patchedThoughts
-            fullResponse += patchedThoughts
 
+          if (ln.response) {
+            if (model === thinking_model) {
+              let patchedThoughts = ln.response
+              // TODO: hide blank thinking chunks
+              const thinkTagRx = /<think>([\s\S]*?)<\/think>/g
+              patchedThoughts = patchedThoughts.replace(thinkTagRx, (match, p1) => {
+                if (p1.trim().length > 0) {
+                  console.log(p1)
+                  return '`Thinking...`' + p1 + '`Finished thinking`'
+                } else {
+                  return ''
+                }
+              })
+              patchedThoughts = patchedThoughts.replace(/<think>/g, '`Thinking...`')
+              patchedThoughts = patchedThoughts.replace(/<\/think>/g, '`Finished thinking`')
+              thoughts += patchedThoughts
+              fullResponse += patchedThoughts
+            } else {
+              fullResponse += ln.response
+            }
             if (now - lastUpdate >= 1000) {
               await rateLimiter.editMessageWithRetry(
                 ctx,
@@ -103,7 +145,7 @@ async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Me
             }
           }
         } catch (e) {
-          console.error("Error parsing chunk:", e)
+          console.error("[✨ AI | !] Error parsing chunk:", e)
         }
       }
     }
@@ -119,7 +161,7 @@ async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Me
       if (error.response.data.error.includes(`model '${model}' not found`) || error.status === 404) {
         shouldPullModel = true
       } else {
-        console.error("[!] 1", error.response.data.error)
+        console.error("[✨ AI | !] Error zone 1:", error.response.data.error)
         return {
           "success": false,
           "error": error.response.data.error,
@@ -130,23 +172,25 @@ async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Me
     }
 
     if (shouldPullModel) {
-      ctx.telegram.editMessageText(ctx.chat.id, replyGenerating.message_id, undefined, `🔄 Pulling ${model} from ollama...`)
-      console.log(`[i] Pulling ${model} from ollama...`)
+      ctx.telegram.editMessageText(ctx.chat.id, replyGenerating.message_id, undefined, `🔄 Pulling ${model} from ollama...\n\nThis may take a few minutes...`)
+      console.log(`[✨ AI | i] Pulling ${model} from ollama...`)
+      let pullModelStream: any
 
-      const pullModelStream = await axios.post(`${process.env.ollamaApi}/api/pull`, {
-        model: model,
-        stream: false,
-      })
-
-      if (pullModelStream.data.status !== ("success")) {
-        console.error("[!] Something went wrong:", pullModelStream.data)
+      try {
+        pullModelStream = await axios.post(`${process.env.ollamaApi}/api/pull`, {
+          model: model,
+          stream: false,
+          timeout: process.env.ollamaApiTimeout || 10000,
+        })
+      } catch (e: any) {
+        console.error("[✨ AI | !] Something went wrong:", e.response.data.error)
         return {
           "success": false,
           "error": `❌ Something went wrong while pulling ${model}, please try your command again!`,
         }
       }
 
-      console.log("[i] Model pulled successfully")
+      console.log(`[✨ AI | i] ${model} pulled successfully`)
       return {
         "success": true,
         "response": `✅ Pulled ${model} successfully, please retry the command.`,
@@ -154,7 +198,7 @@ async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Me
     }
 
     if (error.response) {
-      console.error("[!] 2", error.response)
+      console.error("[✨ AI | !] Error zone 2:", error.response)
       return {
         "success": false,
         "error": error.response,
@@ -162,7 +206,7 @@ async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Me
     }
 
     if (error.statusText) {
-      console.error("[!] 3", error.statusText)
+      console.error("[✨ AI | !] Error zone 3:", error.statusText)
       return {
         "success": false,
         "error": error.statusText,
@@ -177,15 +221,24 @@ async function getResponse(prompt: string, ctx: TextContext, replyGenerating: Me
 }
 
 export default (bot: Telegraf<Context>) => {
-  bot.command("ask", spamwatchMiddleware, async (ctx) => {
+  const botName = bot.botInfo?.first_name && bot.botInfo?.last_name ? `${bot.botInfo.first_name} ${bot.botInfo.last_name}` : "Kowalski"
+
+  bot.command(["ask", "think"], spamwatchMiddleware, async (ctx) => {
     if (!ctx.message || !('text' in ctx.message)) return;
+    const isAsk = ctx.message.text.startsWith("/ask")
+    const model = isAsk ? flash_model : thinking_model
+    console.log(model)
+    console.log(ctx.message.text)
     const textCtx = ctx as TextContext;
     const reply_to_message_id = replyToMessageId(textCtx)
     const Strings = getStrings(languageCode(textCtx))
     const message = textCtx.message.text
     const author = ("@" + ctx.from?.username) || ctx.from?.first_name
 
-    logger.logCmdStart(author)
+    logger.logCmdStart(
+      author,
+      model === flash_model ? "ask" : "think"
+    )
 
     if (!process.env.ollamaApi) {
       await ctx.reply(Strings.aiDisabled, {
@@ -212,12 +265,14 @@ export default (bot: Telegraf<Context>) => {
     logger.logPrompt(fixedMsg)
 
     const prompt = sanitizeForJson(
-`You are a helpful assistant named Kowalski, who has been given a message from a user.
+`You are a helpful assistant called ${botName}.
+Current Date/Time (UTC): ${new Date().toLocaleString()}
 
-The message is:
+---
 
+Respond to the user's message:
 ${fixedMsg}`)
-    const aiResponse = await getResponse(prompt, textCtx, replyGenerating)
+    const aiResponse = await getResponse(prompt, textCtx, replyGenerating, model)
     if (!aiResponse) return
 
     if (aiResponse.success && aiResponse.response) {
@@ -239,7 +294,6 @@ ${fixedMsg}`)
         error,
         { parse_mode: 'Markdown' }
       )
-      console.error("[!] Error sending response:", aiResponse.error)
     }
   })
 }

src/utils/log.ts

@@ -28,6 +28,8 @@
 //
 // For more information, please refer to <https://unlicense.org/>
 
+import { flash_model, thinking_model } from "../commands/ai"
+
 class Logger {
   private static instance: Logger
   private thinking: boolean = false
@@ -41,42 +43,42 @@ class Logger {
     return Logger.instance
   }
 
-  logCmdStart(user: string): void {
-    console.log(`[START] Received /ask from ${user}`)
+  logCmdStart(user: string, type: "ask" | "think"): void {
+    console.log(`\n[✨ AI | START] Received /${type} for model ${type === "ask" ? flash_model : thinking_model} from ${user}`)
   }
 
-  logThinking(thinking: boolean): void {
+  logThinking(chatId: number, messageId: number, thinking: boolean): void {
     if (thinking) {
-      console.log("[THINKING] Started")
+      console.log(`[✨ AI | THINKING | ${chatId}:${messageId}] Model started thinking`)
     } else {
-      console.log("[THINKING] Ended")
+      console.log(`[✨ AI | THINKING | ${chatId}:${messageId}] Model stopped thinking`)
     }
   }
 
   logChunk(chatId: number, messageId: number, text: string, isOverflow: boolean = false): void {
-    const prefix = isOverflow ? "[OVERFLOW]" : "[CHUNK]"
-    console.log(`${prefix} [${chatId}:${messageId}] ${text.length} chars`)
+    const prefix = isOverflow ? "[✨ AI | OVERFLOW]" : "[✨ AI | CHUNK]"
+    console.log(`${prefix} [${chatId}:${messageId}] ${text.length} chars pushed to Telegram`)
   }
 
   logPrompt(prompt: string): void {
-    console.log(`[PROMPT] ${prompt.length} chars: ${prompt.substring(0, 50)}${prompt.length > 50 ? "..." : ""}`)
+    console.log(`[✨ AI | PROMPT] ${prompt.length} chars: ${prompt.substring(0, 50)}${prompt.length > 50 ? "..." : ""}`)
   }
 
   logError(error: any): void {
     if (error.response?.error_code === 429) {
       const retryAfter = error.response.parameters?.retry_after || 1
-      console.error(`[RATE_LIMIT] Too Many Requests - retry after ${retryAfter}s`)
+      console.error(`[✨ AI | RATE_LIMIT] Too Many Requests - retry after ${retryAfter}s`)
     } else if (error.response?.error_code === 400 && error.response?.description?.includes("can't parse entities")) {
-      console.error("[PARSE_ERROR] Markdown parsing failed, retrying with plain text")
+      console.error("[✨ AI | PARSE_ERROR] Markdown parsing failed, retrying with plain text")
     } else {
       const errorDetails = {
         code: error.response?.error_code,
         description: error.response?.description,
         method: error.on?.method
       }
-      console.error("[ERROR]", JSON.stringify(errorDetails, null, 2))
+      console.error("[✨ AI | ERROR]", JSON.stringify(errorDetails, null, 2))
     }
   }
 }
 
-export const logger = Logger.getInstance() 
+export const logger = Logger.getInstance()

src/utils/rate-limiter.ts

@@ -80,7 +80,7 @@ class RateLimiter {
       const timeout = setTimeout(() => {
         this.processUpdate(ctx, chatId, messageId, options)
       }, this.minInterval - timeSinceLastEdit)
-      
+
       this.updateQueue.set(messageKey, timeout)
       return
     }
@@ -124,7 +124,7 @@ class RateLimiter {
 
         const firstChunk = chunks[0]
         logger.logChunk(chatId, messageId, firstChunk)
-        
+
         try {
           await ctx.telegram.editMessageText(chatId, messageId, undefined, firstChunk, options)
         } catch (error: any) {
@@ -136,11 +136,11 @@ class RateLimiter {
         for (let i = 1; i < chunks.length; i++) {
           const chunk = chunks[i]
           const overflowMessageId = this.overflowMessages.get(messageKey)
-          
+
           if (overflowMessageId) {
-            logger.logChunk(chatId, overflowMessageId, chunk, true)
             try {
               await ctx.telegram.editMessageText(chatId, overflowMessageId, undefined, chunk, options)
+              logger.logChunk(chatId, overflowMessageId, chunk, true)
             } catch (error: any) {
               if (!error.response?.description?.includes("message is not modified")) {
                 throw error
@@ -166,7 +166,7 @@ class RateLimiter {
       } else {
         const sanitizedText = latestText
         logger.logChunk(chatId, messageId, sanitizedText)
-        
+
         try {
           await ctx.telegram.editMessageText(chatId, messageId, undefined, sanitizedText, options)
         } catch (error: any) {
@@ -184,7 +184,7 @@ class RateLimiter {
         const retryAfter = error.response.parameters?.retry_after || 1
         this.isRateLimited = true
         this.rateLimitEndTime = Date.now() + (retryAfter * 1000)
-        
+
         const existingTimeout = this.updateQueue.get(messageKey)
         if (existingTimeout) {
           clearTimeout(existingTimeout)
@@ -193,7 +193,7 @@ class RateLimiter {
         const timeout = setTimeout(() => {
           this.processUpdate(ctx, chatId, messageId, options)
         }, retryAfter * 1000)
-        
+
         this.updateQueue.set(messageKey, timeout)
       } else if (error.response?.error_code === 400) {
         if (error.response?.description?.includes("can't parse entities")) {