Building a Telegram-Based AI Assistant for Home Assistant - V1

Smart home systems like Home Assistant are extremely powerful but often come with a learning curve for non-technical users. Triggering automations, querying historical states, or interacting with sensors typically requires navigating a web interface or configuring YAML scripts. This can be unintuitive and limits broader adoption.

But here’s the thing — while Home Assistant is amazing, configuring automations or querying sensor data can quickly become technical, especially for non-developers. YAML files, nested configurations, long entity IDs — it can all feel overwhelming.

Sometimes, all you really wish for is a friendly, magical assistant that you can simply chat with, and it just gets things done. That’s the experience we’re aiming to create — a natural, frictionless, chat-based interface that turns complex automation into a conversation.

Imagine typing:

• "Turn off the garden lights" — and your outdoor lights go off instantly.

• "What was the temperature in the bedroom yesterday at 3 PM?" — and you get back, "It was 23.4°C around that time."

• "Lock the front door" — and a secure confirmation message comes right back.

No dashboards, no digging through YAML, no confusion. Just a simple chat, and your home responds.

It’s this level of simplicity that makes automation truly accessible.

The problem we're solving:

• Enable natural language interaction with Home Assistant.

• Provide a mobile-friendly interface - for this we leverage using Telegram bots.

• Integrate AI capabilities for interpreting user intent.

• Enable insightful retrospection through historical sensor data querying — the cream of this solution — allowing users to ask natural questions like "How long was the door open yesterday?"

  "What was the temperature trend last week?"

  "When did the heater last turn on?"

  "How many hours has the 3D printer run this week?"

  "Where is John right now?"

  "Is my phone charging?"

  in addition to executing services to turn on/off entities

⚡ Existing Solutions: Pros & Cons

🧐 My Approach: Conversational AI Assistant

I introduce a Telegram-based AI Assistant that acts as a natural language bridge between the user and their Home Assistant system. The assistant uses:

• OpenAI for natural language understanding and routing.

• Telegram as a messaging interface.

• Node-RED for flow orchestration.

• Home Assistant as data/action backends.

• MariaDB as backend for custom applications (or perhaps HomeAssistant history integration)

NOTE -While this article focuses primarily on Home Assistant, it's worth noting that I’ve also integrated some custom solutions for expense and food tracking using MariaDB. These capabilities are included within the Telegram assistant for now, but the vision is to eventually decouple and expand them into standalone modules—perhaps a topic for a future article!?

---

🤖 Creating Your Telegram Bot

I will be very brief here. There are tons of articles and youtube videos which will explain on how to create a telegram bot, but overall its the below steps

1. Open Telegram and search for @BotFather.

2. Start a chat and use the /newbot command.

3. Follow the prompts to set a name and username for your bot.

4. Once created, you’ll receive a Bot Token — copy this securely.

5. Plug this token into your Node-RED telegram receiver and telegram sender nodes to enable communication.

That's it! Your bot is now ready to receive messages and interact with your flow.

---

Telegram Message Entry Point & Domain Evaluation

Flows Covered:

• Evaluate Domain

• Get HA Entities Info

• Database Router

Key Nodes:

• Telegram receiver (polling): Captures user messages.

• Setup: Prepares metadata and initial message structuring.

• Open AI - evaluate domain of question: Sends user message to OpenAI to classify the domain (HomeAssistant, Finances, Meals, Chat).

• switch: Routes the message to the correct processing area.

Optional Branches:

• If domain = Finances, routes to MariaDB (expenses2).

• If domain = Meals, routes to healthify_db.

---

Home Assistant Logic & Optional Database Interaction

Flows Covered:

• HomeAssistant Group

• MariaDB: EXPENSES

Home Assistant Integration:

• Get HA Entities Info: Fetches all HA entities and filters relevant ones.

• Open AI Tools - Decide next action: Determines whether to execute a service, check history, or get entity status.

• switch: Routes to specific action flows:

  • execute_ha_service

  • find_occurrence=first

  • find_occurrence=last

  • historical_analysis

MariaDB Integration (Optional):

• SHOW TABLES & DESCRIBE: Constructs a database schema for OpenAI.

• Open AI Tools: Interprets the user's financial query.

• expenses2: Executes SQL query and returns results.

Example Queries:

• "How much did I spend on groceries last month?"

• "Show me all food logs from last week."

---

Core Action Flows

Flows Covered:

• inline query response

• execute_ha_service

• find_first_occurrence / find_last_occurrence

• historical_analysis

• respond

Key Actions:

• callback_query: Handles button-based follow-ups.

• Prepare msg.startdate and msg.enddate: Formats dates for fixed-time queries.

• get history: relative/fixed: Pulls historical data from HA.

• Without Lodash: Optimizes and formats data.

Example Questions This Handles:

1. "What was the temperature in the living room yesterday?"

2. "When was the last time the garage door opened?"

3. "How long was the heater on last week?"

4. "Was the kitchen light turned on today?"

---

📷 Telegram Response & Error Handling

Flows Covered:

• Telegram Response

• Error Handler

Telegram Messaging:

• send modified message: Returns the assistant's response.

• Update telegram-response-msg-id: Tracks message threading.

Error Handling:

• catch: all: Catches errors from any node.

• Error Response: Formats and sends a human-readable error message back to Telegram.

---

⚖️ Bringing It All Together

flowchart TD
    A["Telegram Message"] --> B["Setup Metadata"]
    B --> C["OpenAI: Classify Domain"] & E["Catch Errors"]
    C -- HomeAssistant --> D["Decide Action (AI)"]
    C -- Finances/Meals --> G["Build Schema -> SQL Query"]
    D -- Execute --> D1["Call HA Service"]
    D -- First Occurrence --> D2["Query History First"]
    D -- Last Occurrence --> D3["Query History Last"]
    D -- Historical --> D4["Analyze HA History"]
    G --> H["Run SQL -> Format Response"]
    D1 --> Z["Telegram Response"]
    D2 --> Z
    D3 --> Z
    D4 --> Z
    H --> Z
    Z --> Y["Update Msg ID / Debug"]
    E --> F["Send Error via Telegram"]

---

✅ The Results? You decide!

💪 Capabilities Summary

---

✨ Future Improvements

• Memory + Context Tracking: Multi-turn conversations.

• Ability to track cost per message/evaluation from calls to OpenAI

• Ability to retry, in case the answer is not to your satisfaction or incorrect.

• More robustness, reduce OpenAI calls

• Split into different bots like. - HomeAssistant, Finances, Health serving individual purposes (better to maintain)

• Visualization: Possibly support automated genertaion of graphs using AI.

• Webhooks & Smart Alerts from Home Assistant.

---

🎉 Conclusion

This version of our Telegram-based AI Assistant brings a human-friendly interface to the powerful world of smart home automation. With AI-powered interpretation, natural language support, and rich data handling (via Home Assistant and MariaDB), we’ve built a modular, extensible system ready to empower users to interact with their homes like never before.

Whether it’s checking how long the AC was running last week, finding out when the garage door last opened, or tracking your grocery spending — this assistant turns everyday questions into real-time, useful insights.

Disclaimer: This assistant uses OpenAI's tool_calls mechanism to dynamically determine which Node-RED flow to trigger based on user queries. While powerful, this AI routing mechanism is still evolving. The current implementation can be inconsistent due to the probabilistic nature of AI responses. It's not uncommon for the same question to yield slightly different results across attempts, and occasional hallucinations or misinterpretations may occur. Strengthening the reliability of this tool-call chaining is a key focus area for upcoming versions.

Let’s call this Version 1, but it’s just the beginning. Stay tuned for further updates.
Connect with me if you are interested further in this nodered flow.