managarten/services/matrix-ollama-bot/src/bot/matrix.service.ts
Till-JS dc0d425f61 🐛 fix(matrix-web): handle Matrix SSO loginToken callback
Add loginWithLoginToken function to exchange Matrix SSO loginToken for credentials.
The app layout now detects the loginToken URL parameter and completes the SSO flow.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-02-01 15:02:47 +01:00

725 lines
22 KiB
TypeScript

import { Injectable } from '@nestjs/common';
import { ConfigService } from '@nestjs/config';
import {
BaseMatrixService,
MatrixBotConfig,
MatrixRoomEvent,
KeywordCommandDetector,
COMMON_KEYWORDS,
} from '@manacore/matrix-bot-common';
import { TranscriptionService, SessionService, CreditService } from '@manacore/bot-services';
import { OllamaService } from '../ollama/ollama.service';
import { SYSTEM_PROMPTS } from '../config/configuration';
// Ollama is local, so credits are minimal
const OLLAMA_CHAT_CREDITS = 0.1;
interface UserSession {
systemPrompt: string;
model: string;
history: { role: 'user' | 'assistant'; content: string }[];
pendingImage?: { url: string; mimeType: string };
}
// Models excluded from !all comparison (specialized, not for general chat)
const NON_CHAT_MODELS = ['deepseek-r1:1.5b'];
// Models that support vision/image input
const VISION_MODELS = ['llava', 'llava:7b', 'llava:13b', 'bakllava', 'moondream'];
// Natural language keyword detector
const keywordDetector = new KeywordCommandDetector([
...COMMON_KEYWORDS,
{ keywords: ['modelle', 'models', 'welche modelle', 'liste modelle'], command: 'models' },
{ keywords: ['verbindung', 'connection', 'online'], command: 'status' },
{ keywords: ['lösche verlauf', 'clear', 'neustart', 'reset', 'vergiss alles'], command: 'clear' },
]);
@Injectable()
export class MatrixService extends BaseMatrixService {
private sessions: Map<string, UserSession> = new Map();
constructor(
configService: ConfigService,
private readonly transcriptionService: TranscriptionService,
private ollamaService: OllamaService,
private sessionService: SessionService,
private creditService: CreditService
) {
super(configService);
}
protected override async handleAudioMessage(
roomId: string,
event: MatrixRoomEvent,
sender: string
): Promise<void> {
try {
const mxcUrl = event.content.url;
if (!mxcUrl) return;
const audioBuffer = await this.downloadMedia(mxcUrl);
const text = await this.transcriptionService.transcribe(audioBuffer);
if (!text) {
await this.sendMessage(roomId, '❌ Sprachnachricht konnte nicht erkannt werden.');
return;
}
await this.sendMessage(roomId, `🎤 *"${text}"*`);
await this.handleTextMessage(roomId, event, text, sender);
} catch (error) {
this.logger.error(`Audio transcription error: ${error}`);
await this.sendMessage(roomId, '❌ Fehler bei der Spracherkennung.');
}
}
protected getConfig(): MatrixBotConfig {
return {
homeserverUrl:
this.configService.get<string>('matrix.homeserverUrl') || 'http://localhost:8008',
accessToken: this.configService.get<string>('matrix.accessToken') || '',
storagePath:
this.configService.get<string>('matrix.storagePath') || './data/bot-storage.json',
allowedRooms: this.configService.get<string[]>('matrix.allowedRooms') || [],
};
}
protected getIntroductionMessage(): string | null {
return `**Hallo! Ich bin Manai, eure lokale KI-Assistentin.**
Alle Daten bleiben auf diesem Server - 100% DSGVO-konform!
**Quick Start:**
- Schreibt einfach eine Nachricht
- Sagt "hilfe" für alle Befehle
- Sagt "modelle" um KI-Modelle zu sehen`;
}
async onModuleInit() {
await super.onModuleInit();
if (!this.client) return;
this.botUserId = await this.client.getUserId();
this.logger.log(`Bot user ID: ${this.botUserId}`);
// Setup room join handler for welcome message
this.client.on('room.join', this.handleRoomJoin.bind(this));
// Handle image messages
this.client.on('room.message', async (roomId: string, event: any) => {
if (event.sender === this.botUserId) return;
const content = event.content as {
msgtype?: string;
body?: string;
url?: string;
info?: { mimetype?: string };
};
// Handle image messages - store for later use with !vision
if (content.msgtype === 'm.image' && content.url) {
const session = this.getSession(event.sender);
session.pendingImage = {
url: content.url,
mimeType: content.info?.mimetype || 'image/png',
};
this.logger.log(`Image received from ${event.sender}, stored for !vision command`);
await this.sendMessage(
roomId,
`Bild empfangen! Nutze jetzt:\n- \`!vision [Frage zum Bild]\` - Bild mit einem Modell analysieren\n- \`!vision:all [Frage]\` - Bild mit allen Vision-Modellen vergleichen`
);
}
});
}
private async handleRoomJoin(roomId: string, event: any) {
// Only send welcome when someone else joins (not the bot itself)
if (event.state_key === this.botUserId) return;
if (!this.isRoomAllowed(roomId)) return;
this.logger.log(`User ${event.state_key} joined room ${roomId}`);
// Send welcome message
await this.sendWelcomeMessage(roomId, event.state_key);
}
private async sendWelcomeMessage(roomId: string, userId: string) {
const welcomeText = `**Willkommen im Mana Chat, ${this.extractUsername(userId)}!**
Ich bin **Manai**, deine lokale KI-Assistentin (100% DSGVO-konform).
**So nutzt du mich:**
- Schreib einfach eine Nachricht - ich antworte!
- Sag "hilfe" oder "modelle" für mehr Infos
- Oder nutze Befehle wie \`!help\`
**Quick Start:**
- "Was ist TypeScript?" -> Ich erkläre es dir
- "modelle" -> Zeigt verfügbare KI-Modelle
- \`!all Erkläre Recursion\` -> Vergleicht alle Modelle
Viel Spass!`;
await this.sendMessage(roomId, welcomeText);
}
private extractUsername(userId: string): string {
// Extract username from @user:server.com format
const match = userId.match(/@([^:]+)/);
return match ? match[1] : userId;
}
private getSession(senderId: string): UserSession {
if (!this.sessions.has(senderId)) {
this.sessions.set(senderId, {
systemPrompt: SYSTEM_PROMPTS.default,
model: this.ollamaService.getDefaultModel(),
history: [],
});
}
return this.sessions.get(senderId)!;
}
protected async handleTextMessage(
roomId: string,
_event: MatrixRoomEvent,
message: string,
sender: string
): Promise<void> {
// Handle commands with ! prefix
if (message.startsWith('!')) {
await this.handleCommand(roomId, sender, message);
return;
}
// Check for natural language keywords
const detectedCommand = keywordDetector.detect(message);
if (detectedCommand) {
this.logger.log(`Detected keyword command: ${detectedCommand}`);
await this.handleCommand(roomId, sender, `!${detectedCommand}`);
return;
}
// Regular chat message
await this.handleChat(roomId, sender, message);
}
private async handleCommand(roomId: string, sender: string, body: string) {
const [command, ...args] = body.slice(1).split(' ');
const argString = args.join(' ');
switch (command.toLowerCase()) {
case 'help':
case 'start':
await this.sendHelp(roomId);
break;
case 'models':
await this.sendModels(roomId, sender);
break;
case 'model':
await this.setModel(roomId, sender, argString);
break;
case 'mode':
await this.setMode(roomId, sender, argString);
break;
case 'clear':
await this.clearHistory(roomId, sender);
break;
case 'status':
await this.sendStatus(roomId, sender);
break;
case 'all':
await this.handleAllModels(roomId, sender, argString);
break;
case 'vision':
await this.handleVision(roomId, sender, argString);
break;
case 'vision:all':
await this.handleVisionAll(roomId, sender, argString);
break;
case 'pin':
await this.pinHelpMessage(roomId);
await this.sendMessage(roomId, 'Hilfe wurde angepinnt!');
break;
default:
await this.sendMessage(
roomId,
`Unbekannter Befehl: !${command}\n\nVerwende !help für eine Liste der Befehle.`
);
}
}
private async sendHelp(roomId: string) {
const helpText = `**Manai - Lokale KI (100% DSGVO-konform)**
**Einfache Befehle** (sag einfach):
- "hilfe" - Diese Hilfe
- "modelle" - Verfügbare KI-Modelle
- "status" - Verbindungsstatus
- "lösche verlauf" - Chat zurücksetzen
**Power-User Befehle** (mit !):
- \`!model [name]\` - Modell wechseln
- \`!all [frage]\` - Alle Modelle vergleichen
- \`!mode [modus]\` - Modus ändern (default/code/translate/summarize)
**Bild-Analyse:**
1. Sende ein Bild
2. Dann: \`!vision [frage]\` oder \`!vision:all [frage]\`
**Verwendung:**
Schreibe einfach eine Nachricht und ich antworte!
**Beispiele:**
- "Was ist Kubernetes?" -> Direkte Antwort
- "modelle" -> Zeigt alle Modelle
- \`!all Erkläre Docker\` -> Vergleicht alle Modelle
**Aktuelles Modell:** \`${this.ollamaService.getDefaultModel()}\``;
await this.sendMessage(roomId, helpText);
}
private async sendModels(roomId: string, sender: string) {
const models = await this.ollamaService.listModels();
if (models.length === 0) {
await this.sendMessage(roomId, 'Keine Modelle gefunden. Ist Ollama gestartet?');
return;
}
const session = this.getSession(sender);
const modelList = models
.map((m) => {
const sizeMB = (m.size / 1024 / 1024).toFixed(0);
const active = m.name === session.model ? ' ✓' : '';
return `- \`${m.name}\` (${sizeMB} MB)${active}`;
})
.join('\n');
await this.sendMessage(
roomId,
`**Verfügbare Modelle:**\n\n${modelList}\n\nWechseln mit: \`!model [name]\``
);
}
private async setModel(roomId: string, sender: string, modelName: string) {
if (!modelName) {
const session = this.getSession(sender);
await this.sendMessage(
roomId,
`Aktuelles Modell: \`${session.model}\`\n\nVerwendung: \`!model gemma3:4b\``
);
return;
}
const models = await this.ollamaService.listModels();
const exists = models.some((m) => m.name === modelName);
if (!exists) {
const available = models.map((m) => m.name).join(', ');
await this.sendMessage(
roomId,
`Modell "${modelName}" nicht gefunden.\n\nVerfügbar: ${available}`
);
return;
}
const session = this.getSession(sender);
session.model = modelName;
session.history = [];
this.logger.log(`User ${sender} switched to model ${modelName}`);
await this.sendMessage(roomId, `Modell gewechselt zu: \`${modelName}\``);
}
private async setMode(roomId: string, sender: string, mode: string) {
const availableModes = Object.keys(SYSTEM_PROMPTS);
if (!mode) {
const session = this.getSession(sender);
const currentMode =
Object.entries(SYSTEM_PROMPTS).find(([_, v]) => v === session.systemPrompt)?.[0] ||
'custom';
await this.sendMessage(
roomId,
`Aktueller Modus: \`${currentMode}\`\n\nVerfügbar: ${availableModes.join(', ')}`
);
return;
}
const normalizedMode = mode.toLowerCase();
if (!SYSTEM_PROMPTS[normalizedMode]) {
await this.sendMessage(
roomId,
`Unbekannter Modus: ${mode}\n\nVerfügbar: ${availableModes.join(', ')}`
);
return;
}
const session = this.getSession(sender);
session.systemPrompt = SYSTEM_PROMPTS[normalizedMode];
session.history = [];
this.logger.log(`User ${sender} switched to mode ${normalizedMode}`);
await this.sendMessage(roomId, `Modus gewechselt zu: \`${normalizedMode}\``);
}
private async clearHistory(roomId: string, sender: string) {
const session = this.getSession(sender);
session.history = [];
this.logger.log(`User ${sender} cleared history`);
await this.sendMessage(roomId, 'Chat-Verlauf gelöscht.');
}
private async sendStatus(roomId: string, sender: string) {
const connected = await this.ollamaService.checkConnection();
const models = await this.ollamaService.listModels();
const session = this.getSession(sender);
const loggedIn = this.sessionService.isLoggedIn(sender);
const authSession = this.sessionService.getSession(sender);
const token = this.sessionService.getToken(sender);
let statusText = `**Ollama Status**\n\n`;
statusText += `**Verbindung:** ${connected ? 'Online' : 'Offline'}\n`;
statusText += `**Modelle:** ${models.length}\n`;
statusText += `**Dein Modell:** \`${session.model}\`\n`;
statusText += `**Chat-Verlauf:** ${session.history.length} Nachrichten\n`;
if (loggedIn && authSession && token) {
const balance = await this.creditService.getBalance(token);
statusText += `**👤 Angemeldet als:** ${authSession.email}\n`;
statusText += `**⚡ Credits:** ${balance.balance.toFixed(2)}\n`;
}
statusText += `**DSGVO:** Alle Daten lokal`;
await this.sendMessage(roomId, statusText);
}
private async handleAllModels(roomId: string, sender: string, message: string) {
if (!message.trim()) {
await this.sendMessage(
roomId,
`**Verwendung:** \`!all [Deine Frage]\`\n\nBeispiel: \`!all Was ist 2+2?\`\n\nDie Frage wird an alle Chat-Modelle gesendet und du siehst die Antworten zum Vergleich.`
);
return;
}
const allModels = await this.ollamaService.listModels();
// Filter out non-chat models (OCR, specialized models)
const models = allModels.filter((m) => !NON_CHAT_MODELS.includes(m.name));
if (models.length === 0) {
await this.sendMessage(roomId, 'Keine Chat-Modelle gefunden. Ist Ollama gestartet?');
return;
}
const skipped = allModels.length - models.length;
const skippedNote = skipped > 0 ? ` (${skipped} spezialisierte Modelle übersprungen)` : '';
await this.sendMessage(
roomId,
`**Vergleiche ${models.length} Chat-Modelle...**${skippedNote}\n\nFrage: "${message}"`
);
// Send typing indicator
await this.client.setTyping(roomId, true, 300000);
const session = this.getSession(sender);
const messages: { role: 'user' | 'assistant' | 'system'; content: string }[] = [
{ role: 'system', content: session.systemPrompt },
{ role: 'user', content: message },
];
const results: { model: string; response: string; duration: number; error?: string }[] = [];
for (const model of models) {
const startTime = Date.now();
try {
this.logger.log(`Querying model ${model.name}...`);
const response = await this.ollamaService.chat(messages, model.name);
const duration = Date.now() - startTime;
results.push({ model: model.name, response, duration });
} catch (error) {
const duration = Date.now() - startTime;
const errorMessage = error instanceof Error ? error.message : 'Unbekannter Fehler';
results.push({ model: model.name, response: '', duration, error: errorMessage });
}
}
await this.client.setTyping(roomId, false);
// Format results
let resultText = `**Modellvergleich**\n\n**Frage:** "${message}"\n\n---\n\n`;
for (const result of results) {
const durationSec = (result.duration / 1000).toFixed(1);
if (result.error) {
resultText += `**${result.model}** ${durationSec}s\nFehler: ${result.error}\n\n---\n\n`;
} else {
// Truncate long responses for readability
const truncatedResponse =
result.response.length > 500
? result.response.substring(0, 500) + '...'
: result.response;
resultText += `**${result.model}** ${durationSec}s\n${truncatedResponse}\n\n---\n\n`;
}
}
await this.sendMessage(roomId, resultText);
}
private async handleChat(roomId: string, sender: string, message: string) {
const session = this.getSession(sender);
// Send typing indicator
await this.client.setTyping(roomId, true, 30000);
try {
// Add user message to history
session.history.push({ role: 'user', content: message });
// Keep only last 10 messages
if (session.history.length > 10) {
session.history = session.history.slice(-10);
}
// Build messages with system prompt
const messages: { role: 'user' | 'assistant' | 'system'; content: string }[] = [
{ role: 'system', content: session.systemPrompt },
...session.history,
];
const response = await this.ollamaService.chat(messages, session.model);
// Add assistant response to history
session.history.push({ role: 'assistant', content: response });
// Stop typing indicator
await this.client.setTyping(roomId, false);
// Send response (Matrix has higher message limits than Telegram)
await this.sendMessage(roomId, response);
} catch (error) {
await this.client.setTyping(roomId, false);
this.logger.error(`Error processing message:`, error);
const errorMessage = error instanceof Error ? error.message : 'Unbekannter Fehler';
await this.sendMessage(roomId, `Fehler: ${errorMessage}`);
}
}
private async handleVision(roomId: string, sender: string, prompt: string) {
const session = this.getSession(sender);
if (!session.pendingImage) {
await this.sendMessage(
roomId,
`Kein Bild vorhanden!\n\nSende zuerst ein Bild, dann nutze \`!vision [Frage zum Bild]\``
);
return;
}
if (!prompt.trim()) {
await this.sendMessage(
roomId,
`**Verwendung:** \`!vision [Deine Frage zum Bild]\`\n\nBeispiel: \`!vision Was siehst du auf diesem Bild?\``
);
return;
}
// Find available vision models
const allModels = await this.ollamaService.listModels();
const visionModels = allModels.filter((m) => VISION_MODELS.some((v) => m.name.includes(v)));
if (visionModels.length === 0) {
await this.sendMessage(
roomId,
`Keine Vision-Modelle gefunden!\n\nInstalliere ein Vision-Modell mit:\n\`ollama pull llava\``
);
return;
}
const model = visionModels[0].name;
await this.sendMessage(roomId, `Analysiere Bild mit \`${model}\`...`);
await this.client.setTyping(roomId, true, 120000);
try {
// Download image from Matrix
const imageData = await this.downloadMatrixImage(session.pendingImage.url);
const response = await this.ollamaService.chatWithImage(prompt, imageData, model);
await this.client.setTyping(roomId, false);
await this.sendMessage(roomId, `**${model}:**\n\n${response}`);
} catch (error) {
await this.client.setTyping(roomId, false);
const errorMsg = error instanceof Error ? error.message : 'Unbekannter Fehler';
await this.sendMessage(roomId, `Fehler bei der Bildanalyse: ${errorMsg}`);
}
}
private async handleVisionAll(roomId: string, sender: string, prompt: string) {
const session = this.getSession(sender);
if (!session.pendingImage) {
await this.sendMessage(
roomId,
`Kein Bild vorhanden!\n\nSende zuerst ein Bild, dann nutze \`!vision:all [Frage zum Bild]\``
);
return;
}
if (!prompt.trim()) {
await this.sendMessage(
roomId,
`**Verwendung:** \`!vision:all [Deine Frage zum Bild]\`\n\nBeispiel: \`!vision:all Beschreibe was du siehst\``
);
return;
}
// Find available vision models
const allModels = await this.ollamaService.listModels();
const visionModels = allModels.filter((m) => VISION_MODELS.some((v) => m.name.includes(v)));
if (visionModels.length === 0) {
await this.sendMessage(
roomId,
`Keine Vision-Modelle gefunden!\n\nInstalliere Vision-Modelle mit:\n\`ollama pull llava\`\n\`ollama pull moondream\``
);
return;
}
await this.sendMessage(
roomId,
`**Vergleiche ${visionModels.length} Vision-Modelle...**\n\nFrage: "${prompt}"`
);
await this.client.setTyping(roomId, true, 300000);
try {
// Download image from Matrix once
const imageData = await this.downloadMatrixImage(session.pendingImage.url);
const results: { model: string; response: string; duration: number; error?: string }[] = [];
for (const model of visionModels) {
const startTime = Date.now();
try {
this.logger.log(`Querying vision model ${model.name}...`);
const response = await this.ollamaService.chatWithImage(prompt, imageData, model.name);
const duration = Date.now() - startTime;
results.push({ model: model.name, response, duration });
} catch (error) {
const duration = Date.now() - startTime;
const errorMessage = error instanceof Error ? error.message : 'Unbekannter Fehler';
results.push({ model: model.name, response: '', duration, error: errorMessage });
}
}
await this.client.setTyping(roomId, false);
// Format results
let resultText = `**Vision-Modellvergleich**\n\n**Frage:** "${prompt}"\n\n---\n\n`;
for (const result of results) {
const durationSec = (result.duration / 1000).toFixed(1);
if (result.error) {
resultText += `**${result.model}** ${durationSec}s\nFehler: ${result.error}\n\n---\n\n`;
} else {
const truncatedResponse =
result.response.length > 500
? result.response.substring(0, 500) + '...'
: result.response;
resultText += `**${result.model}** ${durationSec}s\n${truncatedResponse}\n\n---\n\n`;
}
}
await this.sendMessage(roomId, resultText);
} catch (error) {
await this.client.setTyping(roomId, false);
const errorMsg = error instanceof Error ? error.message : 'Unbekannter Fehler';
await this.sendMessage(roomId, `Fehler: ${errorMsg}`);
}
}
private async downloadMatrixImage(mxcUrl: string): Promise<string> {
// Convert mxc:// URL to HTTP URL and download
const httpUrl = this.client.mxcToHttp(mxcUrl);
this.logger.log(`Downloading image from ${httpUrl}`);
const response = await fetch(httpUrl);
if (!response.ok) {
throw new Error(`Failed to download image: ${response.status}`);
}
const buffer = await response.arrayBuffer();
const base64 = Buffer.from(buffer).toString('base64');
return base64;
}
private async pinHelpMessage(roomId: string) {
try {
const helpContent = this.getHelpContent();
const htmlBody = this.markdownToHtmlLocal(helpContent);
const eventId = await this.client.sendMessage(roomId, {
msgtype: 'm.text',
body: helpContent,
format: 'org.matrix.custom.html',
formatted_body: htmlBody,
});
await this.client.sendStateEvent(roomId, 'm.room.pinned_events', '', {
pinned: [eventId],
});
this.logger.log(`Pinned help message in room ${roomId}`);
} catch (error) {
this.logger.error(`Failed to pin help message in ${roomId}:`, error);
}
}
private getHelpContent(): string {
return `**Manai - Befehls-Übersicht**
**Einfach sagen:**
- "hilfe" - Diese Übersicht
- "modelle" - Verfügbare KI-Modelle
- "status" - Bot-Status
- "lösche verlauf" - Chat zurücksetzen
**Power-User (mit !):**
- \`!model [name]\` - Modell wechseln
- \`!all [frage]\` - Alle Modelle vergleichen
- \`!vision [frage]\` - Bild analysieren
**Nutzung:** Einfach schreiben und ich antworte!`;
}
private markdownToHtmlLocal(markdown: string): string {
return (
markdown
// Code blocks
.replace(/```(\w+)?\n([\s\S]*?)```/g, '<pre><code>$2</code></pre>')
// Inline code
.replace(/`([^`]+)`/g, '<code>$1</code>')
// Bold
.replace(/\*\*([^*]+)\*\*/g, '<strong>$1</strong>')
// Italic
.replace(/\*([^*]+)\*/g, '<em>$1</em>')
// Line breaks
.replace(/\n/g, '<br/>')
);
}
}