Most AI sales bots are a chat widget that hallucinates prices. VBWD's is a first-class part of a self-hosted commerce platform: it lives in your own messenger or Telegram, grounds answers in your own documents via RAG, pulls exact prices from the live catalogue so it never invents one, and replies in rich clickable cards. A practical walkthrough of meinchat, the bot plugins, and configuring a consultant.
A chat widget guesses. A bot wired into the platform knows — and answers in buttons you can tap.
Most "AI sales bot" products are a chat widget bolted onto your website that hallucinates prices and forgets it's talking to a customer. VBWD takes a different route: the bot is a first-class part of a self-hosted commerce platform, it can live inside your own messenger or Telegram, it quotes real prices from your live catalogue, and it talks in rich, tappable cards instead of a wall of text. Here's how the pieces fit — and how you'd actually configure a sales consultant with it.
The building blocks
The bot stack in VBWD is deliberately layered, and understanding the layers is the key to using it well.
bot-base is the transport-neutral core. It defines the neutral message types every bot speaks — text, links, and choice buttons — plus a registry of messenger providers and a command dispatcher. Critically, it isn't a bot itself; it's the framework the bots plug into. The nicest design decision: there's no registration ceremony. Being an enabled plugin that implements the seam is the registration, and a disabled plugin contributes nothing automatically.
meinchat is VBWD's own messenger — real 1-on-1 and group chat, self-hosted, running inside your platform. bot-meinchat is the bridge that lets a bot live inside it, in-process, with no webhook. The bot is provisioned as a real user with a bot role and a nickname, so any customer can find it in search and start a conversation as if messaging a person.
bot-telegram is the same bot logic on a different transport. Webhooks are the production path, validated with a per-bot secret token; long-polling exists for development when you have no public HTTPS. The point of the layering: you write the bot once, and it works in both your own messenger and Telegram, because both are just messenger providers behind the same neutral interface.
The two bots that matter for sales
bot-search adds a /search command over your catalogue. It reads a core search registry that shop, booking, and subscription plugins each register into — so it can surface products, bookable resources, and plans, and tap-through to a detail card. And it comes with a safety property worth knowing: the core registry hard-blocks users and invoices from ever being searchable, by refusing the registration outright. A bot cannot be pointed at your customer list, even by mistake.
bot-meinchat-llm is the real sales consultant — a RAG-grounded LLM bot. It answers via an explicit /consultant command or ambiently: a guest in the widget can just type a question, no command needed. And it has the three properties that separate a real sales bot from a demo.
Why it doesn't hallucinate prices
This is the feature that matters most and is most often faked elsewhere. The consultant bot's product knowledge is grounded in retrieval over your own documents — markdown and PDFs you drop into a corpus directory, indexed with Postgres full-text search. No external vector database to run. It answers from your content, not the model's imagination.
And prices don't come from the language model at all. Recommendations and prices are pulled from the live catalogue through the platform's pricing engine, so — in the project's own words — the bot never invents a price. That's the difference between a bot that says "this plan is around forty dollars, I think" and one that quotes the exact current price because it read it from the same system your checkout uses.
There's a second corpus, too, and it's a clever touch: separate from the product knowledge, a training directory of "how to sell" lessons — your sales method, example dialogues, objection handling — is always injected into the bot's context. So you're not just teaching it what you sell; you're teaching it how your best salesperson sells.
Rich content: buttons and choices, not walls of text
Here's where the conversation stops feeling like a chatbot and starts feeling like an app. Because bot-base defines rich message blocks — text, links, and choice buttons — the bot replies with tappable options instead of asking the customer to type.
In practice that means a product recommendation arrives as a card with a clear action: a tap to view details, a tap to open the full page, a tap to choose between options. bot-search's result cards carry an action scheme under the hood — view a product, open its page — and the "open page" choice can carry a real URL, so on a rich client the customer navigates straight to it instead of the bot dispatching another round of text. The experience is: the bot shows you three plans as three buttons, you tap one, it shows you that plan's details with a "start checkout" action. No typing, no misunderstanding, no dead ends.
That rich-content layer is also what makes the bot work identically across meinchat and Telegram — both render the same neutral choice blocks in their own native button UI.
How you'd configure the sales consultant
The setup is genuinely mostly configuration, not code:
- Enable the plugins — bot-base, then bot-meinchat and/or bot-telegram for your channels, then bot-meinchat-llm for the consultant.
- Point it at a model — the bot holds no API key of its own. It resolves a central LLM connection you configure once in the admin, so you choose the provider (or a model you host) in one place.
- Fill the two corpora — drop your product and pricing knowledge into the RAG directory, and your sales method and example dialogues into the training directory. Reindex from the admin.
- Tune the prompts — the system and user prompt templates are editable files, seeded with defaults, so you shape the bot's voice without touching code.
- Decide the reward behaviour — on a buy intent the consultant can mint a referral coupon and a checkout deep link, so a sale it closes is attributable, and there's a guest token economy (one token per word) with a switch to not charge guests for the bot's own answers — on the reasoning that a sales pitch shouldn't drain a prospect's balance.
Explore the plugin catalogue, the architecture, and the developer docs for the specifics.
The honest limits
meinchat is text and images — there's no voice or video. The consultant's retrieval is full-text, not semantic embeddings: cheaper and simpler to run, but not the same as a vector search, so the quality of your corpus matters. It's self-hosted, so you run the server and configure the LLM connection. And a sales bot is only as good as the content and sales lessons you feed it — the platform gives you a grounded, rich-content, price-honest consultant; the persuasion still comes from what you teach it.
Why it's different
The thread through all of this is that the bot isn't a bolt-on — it's part of a commerce platform that already knows your catalogue, your prices, your customers, and your channels. That's why it can quote a real price, why it physically can't be made to search your customer list, why "start checkout" is a real button and not a suggestion, and why writing it once gets you a consultant in both your own messenger and Telegram. A chat widget guesses. A bot wired into the platform knows — and answers in buttons you can tap.
Learn more about VBWD
VBWD is a self-hosted, source-available platform for building subscription products, marketplaces, and AI-powered apps. Explore it further:
- 🌐 Website and documentation: vbwd.cc — see the plugins, architecture, and developer docs.
- 💻 Source code and plugins on GitHub: github.com/VBWD-platform
- 🎥 Watch VBWD in action: demo video 1 and demo video 2
- 💼 Follow the project on LinkedIn: linkedin.com/company/vbwd https://redrobot.online/2026/07/18/how-to-build-a-sales-bot-on-vbwd-that-quotes-real-prices-and-talks-in-tappable-buttons/
No comments:
Post a Comment