Autoreply and AI

Auto-reply and AI

The bot supports two auto-reply mechanisms: keyword-based canned responses (built-in) and AI-powered replies via LLM (optional). Both can be used simultaneously — keyword rules are checked first, and the LLM handles messages that don't match any rule. On top of that, the LLM can triage tickets, remember conversations and translate staff replies.

---

Keyword-Based Auto-reply Rules

Define question-answer pairs in config.yaml. When a user message contains the question text as a substring, the answer is sent automatically. First matching rule wins.

Configuration

autoreply:
  - question: "install"
    answer: "You can install using our [Getting Started guide](https://github.com/bostrot/telegram-support-bot/wiki/Getting-started)"
  - question: "refund policy"
    answer: "We offer full refunds within 30 days of purchase."
  - question: "hours"
    answer: "Our support team is available Monday-Friday, 9am-5pm CET."

How Matching Works

User Message	Matches Rule	Reason
"How do I install?"	question: "install"	Substring match
"What's your refund policy?"	question: "refund policy"	Substring match
"Do you offer refunds?"	no match	"refund policy" does not appear as a whole
"Installation help please"	question: "install"	Matches substring "install"

Matching is a plain includes() — write the question in the casing users are likely to type, or add several rules.

Auto-reply Behavior in Staff Chat

When an auto-reply is sent, the following occurs based on your settings:

• If show_auto_replied: true, the ticket is still forwarded to the staff chat, marked with language.automatedReplySent
• The user sees the reply with a signature from language.automatedReplyAuthor (default: "BottyBot")
• Users can provide feedback via the doesntHelp ("This does not help") button on auto-replies

Markdown in Answers

Auto-reply answers support Markdown formatting. Use \n for line breaks within YAML strings:

autoreply:
  - question: "pricing"
    answer: "*Plans:*  \n- Basic: $9/mo  \n- Pro: $19/mo  \n- Enterprise: Contact us"

---

LLM-Powered Auto-reply

Enable AI-generated responses using OpenAI or any OpenAI-compatible endpoint (LiteLLM, Ollama, vLLM, …). The bot sends the user's message along with your knowledge base to the LLM, which answers only from that knowledge base.

Configuration

use_llm: true                              # Enable LLM auto-replies
llm_api_key: 'sk-your-api-key-here'        # API key for your provider
llm_base_url: 'https://api.openai.com/v1'  # LiteLLM proxy or direct API endpoint
llm_model: 'gpt-4o-mini'                   # Model identifier
llm_knowledge: |                           # Custom knowledge base (multi-line string)
  Company ABC provides cloud hosting services.
  
  Pricing tiers:
  - Starter: $5/mo, 10GB storage, shared resources
  - Business: $20/mo, 100GB storage, dedicated CPU
  - Enterprise: Custom pricing
  
  Support hours: Mon-Fri 9am-6pm CET.
  
  Refund policy: Full refund within first 14 days. No refunds after that.

Settings Reference

Setting	Type	Default	Description
use_llm	boolean	false	Enable/disable LLM auto-replies. Automatically sets show_auto_replied: true.
llm_api_key	string	—	API key for the LLM provider or proxy
llm_base_url	string	—	Base URL of the OpenAI-compatible endpoint
llm_model	string	—	Model name (e.g., gpt-4o-mini, qwen/qwen3.6-35b-a3b, claude-3-haiku via LiteLLM)
llm_knowledge	string	(empty)	Knowledge base injected as context. Required — with an empty knowledge base the model is told to answer null and never replies (the bot logs a warning at startup).
llm_memory_depth	integer	10	Previous messages of the same ticket sent along as conversation history (0 = disabled)

How It Works

1. User sends a message that doesn't match any keyword autoreply rule
2. Bot builds a prompt: system instructions + llm_knowledge + the last llm_memory_depth messages of the ticket + the new message
3. Sends the request to the configured endpoint
4. If the model answers, the reply is sent to the user attributed as an automated reply; if it answers null (question not covered), the bot logs LLM returned no answer and the message becomes a normal ticket

Fallback Behavior

If the LLM request fails (network error, rate limit, invalid key) the error is logged (Error in LLM response: …) and the message falls through to normal ticket handling: it is forwarded to the staff chat as a regular ticket. The user won't see an error — they'll receive a human response instead.

LiteLLM Proxy Setup

LiteLLM allows you to use any LLM provider through a single OpenAI-compatible endpoint:

1. Install and run LiteLLM proxy
2. Configure it with your preferred providers (OpenAI, Anthropic, Mistral, local models, …)
3. Point llm_base_url to your LiteLLM instance:

use_llm: true
llm_api_key: 'your-litellm-proxy-key'
llm_base_url: 'http://localhost:4000/v1'   # Your LiteLLM proxy URL
llm_model: 'claude-3-haiku'                # Any model your proxy routes to

Knowledge Base Tips

• Keep llm_knowledge concise and well-structured — Q/A pairs or short sections work best
• Include pricing, policies, common procedures, and product details
• The config is read at startup: restart the bot after editing the knowledge base

---

AI Triage

With auto_triage: true every new ticket is classified by the LLM before it reaches staff:

auto_triage: true
sentiment_alert_threshold: 2   # 1 = very angry … 5 = happy

The model returns a category (one of your configured categories, if any), a priority (low/normal/high/urgent), a one-line summary and a sentiment score. Priority and summary are prefixed to the ticket in the staff chat, the priority is stored on the ticket, and when the sentiment score is at or below sentiment_alert_threshold the language.sentimentAlert warning is added.

Translation

translate_enabled: true
translate_target_language: 'en'

When enabled, staff replies are translated to translate_target_language by the LLM before they are sent to the user. Incoming user messages are forwarded unchanged.

Staff assist

staff_assist is present in config-sample.yaml and the code contains a draft generator, but it is not wired to any command yet — the setting currently has no effect.

---

Combining Keyword Rules and LLM

When both are configured, the bot checks in this order:

1. Keyword autoreply rules → if matched, send canned answer immediately
2. LLM auto-reply → if no keyword match and use_llm: true, ask the model; it answers only when the knowledge base covers the question
3. Normal ticket → if neither applies (or the LLM fails), forward to the staff group as a regular ticket

This means you can use keyword rules for exact, policy-critical answers (refund terms, legal disclaimers) while letting the LLM handle open-ended questions.