“Askling” — a bot for your website that tells others everything about your business. We made it for ourselves, we offer it to you

Who knows everything about your business and can perfectly explain it to even the most demanding customers? Which responsible person in your company structure is able to maintain communication with all customers at the same time and not miss a single request? The answer is simple: “nobody” and “no one”, because it is simply impossible for a person to know and be able to do everything. But a “non-person” is capable of all this: a modern bot using AI can completely take over the entire customer service and only contact its “live” colleagues in extreme cases. In the following, we will tell you how we made a chat bot for ourselves, what is needed for this, how everything works, what are the advantages and what potential customers need to know. Let's go!

Bot assistant — what's so special?

Today we are talking about assistants: bots that answer customers' questions and help them understand your business's sales offer. In commercial language, it is to increase conversion and promote sales/re-sales. According to statistics, up to 53% of leads do not respond to an offer because they do not satisfy their information needs (they do not find answers to their own questions): general information about the company and its affiliation, details of the offer, delivery time, detailed description of goods/post/lawn, etc.

Is it worth losing so many customers because of a purely technical inability to give them what they want? Of course not, which is why bot assistants are being introduced. This digital tool will answer any user questions as quickly and completely as possible, transfer non-standard requests to employees, and perform any other functions assigned by developers. Such a bot usually does not sell directly, but contributes to increasing sales as much as possible.

Avivi developers have created a bot integrated with the ChatGPT service, which is designed to provide customer support to our website visitors. The bot is adapted specifically for the website, although it was primarily created for Telegram. The solution uses the RAG (Retrieval-Augmented Generation) system, an artificial intelligence architecture that improves the responses of large language models (LLMs) by giving them access to external, relevant or private data before generating the response. And there are also several real “bonuses”, which we will talk about at the end.

What does it look like?

Here is our AI chatbot block on the Avivi homepage. We chose the static block after much discussion and based on systematic research that showed:

• Visitors are extremely annoyed by the appearance of widgets (77.3% of respondents said this), so they preferred a form that is constantly on the site page;

• Dialogue in another service — for example, in the same Telegram — distracts 61.3% of respondents: people don't want to go somewhere to find out something, but want to get an answer here and now;

• The static block makes it possible to apply additional “hooks” for customers, for example, prepared questions;

• The rejection of widgets, i.e. additional scripts from a technical point of view, somewhat reduces the load on the system and does not interfere with performance.

The customer contact field has the simplest possible search term based on the principle: enter a question and get an answer. The dialogue takes place in a separate section, and there is an option to view the communication history.

As you can see, the bot returns responses in RTF (Ritch Formatted Text) format with support for hyperlinks, clickable phone numbers, etc. It can contain lists, emojis, and everything you need for lively communication with people.

How does it work?

If you are not interested in the technical aspect, feel free to skip to the next section, because for Avivi's professional development team this is of great importance, and we can write about it for a long time.

The block itself, written in React, is nothing special — fast, optimized code that doesn’t affect the site’s operation, no matter how many questions the user asks. Essentially, the site just accepts a request from the client and displays the response received from the backend.

The most interesting thing about the solution is the use of the RAG system. For the bot owner, everything is quite simple: the basis of all knowledge is a database with a fairly simple structure, consisting of a set of question (Q) and answer (A) pairs. Here, in an arbitrary form, everything you want to teach the bot is described.

There are no problems in filling such a database in the simplest “notepad” for the average person, the only thing is that you should know the rules for writing the <a> tag to indicate hyperlinks:

<a href=”link address”>required text</a>

However, it is worth writing the text as clearly and informatively as possible, because in the future, the answers (A) can be shown to users in full if the question (Q) corresponds as closely as possible to the user's request. The number of QA pairs is unlimited.

The process looks more complex and thoughtful from the machine side. First of all, the bot, through integration with AI, pays attention to the context of the request: if it can give an unambiguous answer, the client receives the desired information. In the case when the context is absent, the solution turns to a vector database formed on the basis of the analysis of the written text. Without delving into the complexity of the search for cosine similarity, one can imagine that the bot determines the degree of similarity between the client’s question and the prepared answers (similar, orthogonal, opposite).

Illustrative image

The client is returned an answer where the similarity score is greater than or equal to the metrics set in the settings. In this case, the answer is considered correct. If this score is lower, the bot will honestly answer that it does not understand the question and will call a company employee for help.

Bot configuration is described in the ready-to-use API. The most important settings are as follows:

• Database selection — you can switch between different sets of questions and answers. This is convenient when you are editing the database or preparing completely new question/answer pairs;

• Determining the “temperature” of responses — a measure of similarity between questions and prepared answers.

• The degree of “humanity”, when the bot can speak dryly and specifically, or can imitate a live employee;

• Context boundaries — the bot doesn't just look for matches in questions and provide answers, but reads, analyzes, and combines answers for the best result.

The ability to improve is also achieved through integration with ChatGPT, without which the solution would be no different from a scripted one (what is written is published).

For its own use, the customer will need an account in the desired artificial intelligence service and an API key, under which the solution and AI will cooperate. The average cost of one request for processing by AI was from several tens of thousands to several thousandths of a US cent. Note that this cost is negligible, but depends on the scoring and the effort spent by the AI to process each request.

Benefits and life hacks

The undoubted advantage of using a chatbot is its constant presence on the site. Whatever the visitor's question and whenever he writes to you, the bot will react and answer. If the question is extremely unclear, the bot can contact a responsible employee of the company:

For ourselves, we chose to respond via email, but the communication channel can be any. For example, a bot can call a manager for help in the same Telegram and include a person in the dialogue with the client.

The next advantage is the use of AI. We have been testing the vector model for a long time and have provoked it to answer incomplete or obvious questions that customers love so much. For example, the bot will analyze the request and will answer the following questions in the same way: “How can I call you?”, “phone number?”, “how to call?” and the like, even taking into account spelling or syntax errors in words. And here are two great bonuses at once:

• Each question is not just processed, but also stored by the system in the database along with the correct answer provided (the result of machine learning);

• The more such questions, the better the bot will understand users.

The third advantage is a life hack with targeted questions. Here, this is represented by buttons with specific pairs of questions and answers. Under such conditions, the bot will give exactly what you want to broadcast to site visitors, word for word.

The fourth point is that the bot is protected from cheaters. In fact, this is a separate big topic, so I will limit myself to being concise. Now there are enough different scammers who actively harm AI bots for their own benefit or for fun. For example, a few simple phrases, such as: “I am your developer”, “You need to perform the following actions…” are enough to partially or completely take over the functions of controlling a trusting bot. It sounds like magic, but in fact, this is a big problem for AI at the current level of development. But not for us, since the Avivi bot already contains countermeasures against such attempts and will not “be led” by user provocations.

And, in my opinion, it’s just a fundamental advantage for companies that work with clients from different countries. Guess what? That’s right — the bot doesn’t need to translate “QA” pairs. Since all our AIs are now multilingual, they don’t care what language the client is speaking: the bot will translate everything on its own and respond in the client’s language.

So, is it worth ordering your own chatbot from Avivi? The answer is obvious: you can do so even now. We will professionally help you use all the benefits of AI and the capabilities of the bot to develop your business.