Prevent Hallucinated Bot Commands: Build an LLM Tool Loop for Telegram in Node.js

When constructing a Telegram bot powered by a Large Language Model, developers initially feed the user's message directly to an inference API and relay the raw output back to the bot using sendMessage(). While this approach functions well for generic queries, it falters when the bot is tasked with executing concrete actions like checking order status or booking a table.

Relying solely on prompt engineering—asking the LLM to return plain text or JSON formatted data—inevitably leads to failure in production. The model tends to hallucinate missing parameters, insert conversational text around the JSON block, format output with unparseable Markdown fences, or fabricate database records when critical information is absent.

Moreover, feeding sensitive system context or credentials into the LLM system prompt can compromise the Telegram bot's token and internal database structure. To create a dependable transactional bot, developers must implement strict function calling, or tool use, in Node.js. This approach ensures the LLM does not directly execute actions or access API keys.

Instead, the model generates structured intent JSON, the Node.js application executes local code (such as querying a PostgreSQL database or making a REST service call), and the results are returned to the model, which then crafts a final human-readable response for the Telegram user.