SupportAgent Overview

SupportAgent is a hybrid AI assistant that combines knowledge base search (RAG) with tool calling. It operates in the Telegram, VKontakte, and Max support chats.

Architecture#

Components#

The system consists of three layers:

  1. SupportMessageDispatcher - accepts incoming messages from all channels, deduplicates requests (one user - one active processing), resolves userId from socialId.

  2. SupportRagService - the main request-processing service. Detects intent (question vs. command), searches the knowledge base, manages conversation history.

  3. SupportToolDispatcher - the tool registry. Accepts a tool name and parameters, calls the required ISupportTool.

Processing flow#

User sends a message
  → SupportMessageDispatcher
    → Resolve userId from socialId
    → SupportRagService.AskAsync()
      → LM Studio: detect intent (question or command?)
      ├─ Question → search the knowledge base → build an answer via LLM
      └─ Command → SupportToolDispatcher.RunAsync()
          → LM Studio + tool_calling → select tool
          → ISupportTool.ExecuteAsync()
          → result → LLM builds the final answer
  → send the response to the user

Intent detection#

Before processing, the agent analyzes the message: is it a question or a request to perform an action?

  • Question: "How do I create a giveaway?", "What are sparks?" - the agent looks for the answer in the knowledge base via semantic search.
  • Command: "Show my activities", "Check my balance" - the agent calls the corresponding tool.

If the language model is unavailable or unsure about the intent, the agent falls back to knowledge base mode.

Knowledge base (RAG)#

SupportAgent uses Retrieval-Augmented Generation:

  1. The user's query is converted into a vector via LM Studio Embeddings
  2. The vector store (km-storage) is searched for the most similar knowledge base document fragments
  3. The retrieved fragments are passed to the language model as context
  4. The LLM builds an answer strictly based on the provided context

If no matching fragments are found, the agent responds with a fallback message and suggests contacting a live specialist.

Conversation history#

The agent maintains conversation context. The last N messages (question + answer) are passed to the LLM on every new request, which allows follow-up questions without repeating context.

History is stored in the database (the SupportChatHistory table) and is tied to the session.

Agent tools#

Tools are the real actions the agent can perform. Each tool implements the ISupportTool interface.

Full list of tools and their description: