~ 7 min

A free Telegram bot answers questions about Tilda and Nolim modifications. Answers are built only from a verified knowledge base: a step-by-step instruction, a link to a ready-made mod and to the author of the material. The bot reads screenshots, keeps AI spend under control, tells the team what it does not know yet, and watches for changes in Tilda code that may affect the mods. It runs in production with 650+ active users and at least 5 annual subscription renewals. The development paid for itself.

Imagine you are building a site on Tilda and a button in a Zero Block disappears on mobile. You search the help center, watch YouTube videos, ask in chats. An hour later you find a piece of code but have no idea where to paste it.

Nolim (nolim.cc) makes ready-made modifications for Tilda. Developers and designers who build sites for their clients use them, and these questions arrive every day. The team wanted an assistant that answers around the clock and is free for the user. It should answer to the point and, instead of code, offer a ready mod that installs without programming.

The project started on 24 March 2026.

- Answers live in four places: Tilda Help, tutorial videos, free courses and mod descriptions on nolim.cc. A newcomer does not know where to look.
- Plain ChatGPT answers “from memory”. It invents settings that do not exist and gives raw code that a beginner cannot paste.
- Every AI answer costs money. A free bot without limits can burn a monthly budget in a week.
- The team cannot see what people ask. No statistics and no list of questions still without an answer.
- Tilda changes without warning. A platform code update can break an installed mod, and the developer hears about it from an unhappy client.

Build an MVP Telegram bot that:
- accepts text and Tilda interface screenshots, and politely rejects PDFs, voice messages and video;
- gives a step-by-step algorithm like “open settings → add a block”;
- never outputs code, but finds a suitable Nolim modification and explains how to set it up;
- links to the author when it quotes a course;
- politely declines off-topic questions (for example, “how to set up ad targeting”);
- limits spend with daily and monthly per-user limits;
- hands unanswered questions to the team and shows metrics;
- replies in 10–15 seconds, even when a screenshot has to be analysed;
- tracks changes in Tilda code and warns if they may affect the mods.

We built the bot with RAG. In plain words: before answering, the bot searches the knowledge base for the most similar fragments of instructions and writes the answer only from them. That is how an experienced support agent works: they check the regulations first. So the bot does not make things up, and if the base has no answer, it says so honestly.

All logic lives in n8n, the knowledge base sits in the Qdrant vector database, and statistics and the admin panel live in a plain SeaTable table. No separate admin panel had to be developed.

Telegram Bot API — the interface: users talk to the bot in a familiar messenger and install nothing.

n8n — the brain of the system: receives messages, checks format and limits, searches the base, calls AI and saves logs.

Qdrant — the vector knowledge base: finds instruction fragments close in meaning to the question even when the words differ.

FireCrawl — collecting materials from websites for the base.

OpenAI — screenshot analysis (GPT-4o) and turning texts into vectors for search.

DeepSeek — generating the final answer at a low cost.

Redis — fast limit counting and memory of the latest messages in a dialogue.

SeaTable — logs, alerts, metrics and the admin panel without separate development.

Dokploy — deploying all services on the client’s own server.

- AI and Telegram keys are stored in the protected n8n vault, and the database is closed to outside access.
- The bot is protected against prompt injection, that is, attempts to talk it into breaking the rules or revealing internal data.
- The mods’ source code is unavailable to the bot under any request.
- At peak load requests queue up instead of getting lost.

The bot runs in production and became for Nolim not just support but part of the loyalty system. Customers get a free assistant and an early warning about risks to their sites, so they stay in the Nolim ecosystem and renew subscriptions.

- 650+ active users.
- At least 5 annual subscription renewals thanks to the bot.
- The development paid for itself.
- 18–20 days of development.
- A reply in 10–15 seconds, including screenshot analysis.

If you have many materials (help center, courses, product descriptions) and support drowns in repeated questions, a bot like this can be built for your content. It will answer only from your base, reveal nothing extra and show what it still lacks. Describe the task in the brief and we will estimate the timeline and a deployment option on your own server.