DIALOGUE GENERATION FOR SPECIFIC SCENARIOS USING LOW CODE SYSTEMS

ABSTRACT

In this paper, novel and comfortable way creating or, let us say, generating dialogues in thematic space for specific scenarios is spoken. We will use specific platforms, called Low- Code platforms. Low-Code platforms let us create platforms without deep knowledge in programming. Usually, to make dialogue system, good understanding in development and the ability to configure hard infrastructure is needed. This requires a lot of time and experience. But in our approach, it is different. We will show that this can be avoided without hard knowledge. Platforms with a low entry threshold provide the possibility of a fast and simple creation of” smart” assistants. These assistants can solve specific tasks. For instance, they can help clients in the shop or in the support service. We used three instruments: Make.com, GPT-4, and AppSheet. With their help, we can create an assistant that will work in real time. This kind of assistant can communicate with people. The assistant can understand what the user is talking about. The assistant can answer the questions. He can also change his behavior depending on the situation. This means that the assistant adapts to different contexts. To test how well our method works, we made a special example.It was a prototype, which means it’s an evaluation version. We have used it in retail and customer service. We evaluated the result based on several indicators. First, how much time is needed for development. Secondly, how well the assistant can adapt to different tasks. Thirdly, is it convenient for people to communicate with him? And fourthly, is it economically beneficial or not.

АННОТАЦИЯ

В данной работе рассматривается новый и удобный способ создания, а точнее, генерации диалогов в тематическом пространстве для конкретных сценариев. Для этого мы будем использовать специализированные платформы, называемые Low-Code платформами. Low-Code платформы позволяют создавать приложения без глубоких знаний в области программирования. Как правило, для создания диалоговой системы необходимы хорошее понимание разработки и умение настраивать сложную инфраструктуру. Это требует значительных временных затрат и опыта. Однако наш подход принципиально иной. Мы покажем, что всего этого можно избежать без глубоких технических знаний. Платформы с низким порогом входа открывают возможность быстрого и простого создания «умных» ассистентов. Такие ассистенты способны решать конкретные задачи. Например, они могут помогать клиентам в магазине или в службе поддержки. В работе были использованы три инструмента: Make.com, GPT-4 и AppSheet. С их помощью можно создать ассистента, работающего в режиме реального времени. Подобный ассистент умеет общаться с людьми: понимать, о чём говорит пользователь, отвечать на вопросы, а также изменять своё поведение в зависимости от ситуации. Это означает, что ассистент адаптируется к различным контекстам. Для проверки эффективности предложенного метода был создан специальный пример — прототип, то есть оценочная версия системы. Он был протестирован в сферах розничной торговли и обслуживания клиентов. Оценка результатов проводилась по нескольким показателям: во-первых, сколько времени требуется на разработку; во-вторых, насколько хорошо ассистент справляется с адаптацией к различным задачам; в-третьих, насколько удобно людям с ним общаться; и в-четвёртых, является ли это экономически выгодным.

Keywords: Low-Code platforms, dialogue generation, conversational AI, GPT-ң, smart assistant, Make.com, AppSheet, customer service, scenario-based dialogue, natural language processing.

Ключевые слова: обучение Low-Code платформы, генерация диалогов, разговорный ИИ, GPT-4, умный ассистент, Make.com, AppSheet, обслуживание клиентов, сценарные диалоги, обработка естественного языка.

1. Introduction

Dialogue Systems are specific platforms that can communicate with people. They will answer questions, help find information, perform actions, and solve tasks. Such systems are becoming increasingly popular. They are especially important for businesses that want to improve customer service and automate routine processes. In the modern world, digital technologies are developing very rapidly [1]. Companies from different industries are starting to use dialogue systems. For example, in the field of customer service, such assistants help us quickly answer questions, take orders, process refunds, and even record meetings. In education, dialogue systems can act as tutoring assistants: they explain topics, ask questions, and give feedback. In healthcare, such systems can remind us about taking medications, collecting primary information from patients, or answer frequently asked questions. In logistics, they help track orders, clarify delivery statuses, and simplify customer interaction [1].

Previously, such systems were difficult to create. Sophisticated technologies were used for this. These technologies are called NLP — natural language processing. To create an assistant, it was necessary to assemble many parts, including a text comprehension system, a response system, a speech recognition system, and others. Machine learning was also used. This means that the system had to be trained on large datasets. All this required knowledge. It was necessary to understand programming, mathematics, and language [2].

Therefore, only large companies could afford such systems. Small businesses couldn’t do that. It was also difficult for ordinary people, such as teachers, staff, or students. If they had an idea, they couldn’t implement it. They didn’t have the necessary knowledge. This prevented the development of technology in small companies.

But everything changed when new models appeared. The GPT-3.5 and GPT-4 models from OpenAI have become especially important [3]. These models understand text very well. They know how to provide clear answers. We don’t need to build a complex system to work with them. Just send the text to the API. The model will understand the question and answer it on its own. This simplified the process; we no longer need to connect many modules.

However, there are still difficulties. For example, to send a request to GPT-4, we need to know what an API is. We need to understand how an Internet query works. We need to be able to transmit and receive data. We also need to store information somewhere. And we need an interface — a place where a person can write messages. All of this requires at least a little knowledge.

That’s why Low-Code and No-Code platforms have become popular. These are services where we can create solutions without code or with a very small amount of code. Everything is done through a user-friendly interface. Just drag the blocks, connect them with arrows, and select actions. Examples of such platforms include Make.com, AppSheet, Airtable, and Zapier [4] [5].

These platforms help people who don’t know how to program. For example, a teacher, a salesman, or a businessman can all create an assistant. They don’t need to hire a programmer; they can do everything themselves. It’s fast, simple, and straightforward.

In this article, we’ll show us a simple way to create a dialog assistant. We will use GPT-4 and platforms without code. The goal is for anyone to be able to repeat it, even if they have never written code.

We have chosen a real-life example: a clothing store. In it, the assistant must be able to do different things. It answers questions like, “Is there a different size?”, “When is the new collection?”, and “Can I return it without a receipt?” It helps us make an appointment for a fitting or a consultation. It also accepts complaints and assists with refunds and collects contact information.

To do all this, we use four technologies:

• Make.com: This is a platform where we can customize the logic. It connects different services via an API [6].

• AppSheet: This is a tool from Google. We can quickly create an interface with it. The user will see the required fields [2].

• Google Sheets: This is a spreadsheet where we store clients, messages, and responses.

• GPT-4 API: It’s a smart model that responds to user messages [3].

We will also evaluate how well everything is working. We’ll see how long it takes to create it. Let’s check if the assistant answers correctly and find out if the users are satisfied. We can conduct a survey or observe how the dialogues are going.

So, we show that anyone can create a smart assistant. We don’t need to be programmers; just a little time and a clear plan are enough. The necessary tools are already available.

2. Materials and methods

The project’s research process comprises two consecutive phases, namely, system design of the interactive subtitle translation system and experiments of the system. The two- phase process ensures that all the functions of importance in the system are fully developed before finally testing them in a controlled learning environment.

1. Solution Architecture

Proposed system is developed at module architecture. This system combines modern language models and visual automate instruments and simple interfaces. This provides us to create” smart” assistants without programming. This approach is available for people which do not have specific technical education.

2. Main components

Interface - AppSheet. Responsible for user interaction. Users can do within form and field: ask a question; require information; leave a complaint; make an appointment; The interface also shows answers from system. The interface can support scenarios with several forms.

• Orchestration - Make.com. This is visual instrument. It controls the movement of data between parts of the system. It gets signal from interface within webhook. And processes data. It will call needed modules(for example AI modules such as GPT API, xAI API). Saves the results inside storage. In our system Make is the brain os all scenario logic.
• Generation texts - GPT-4 API. GPT will generate responses based on the input. It will write answers in dialogue system. This is the only one module with” in-

tellect”. It understands language and can generate texts.

• Storing Data - Google Sheets. This is the database. Storing history of dialogue, user data, answers and technical information. Advantages: easy setup; accessibility; easy to connect to other tools.

Figure 1. Make.com integration

3. System operation scenario

The dialogue process between user and assistant is the sequence of actions. That is important to keep feedback, context and adapt the response.

• Users will write messages within AppSheet form. This can be question, complaint or booking request.
• AppSheet will send HTTP request, that is webhook, to Make.com. This will run automation scenario.
• Make.com will look for users prev messages in Google Sheets. If there are like it will be used. This will help us keep context and give suitable answers.
• Will be created system prompt. There will be:

– Instruction for model (System message)

– the last 3-5 users and assistant’s messages

– the current message of user

• This prompt will be sended to GPT API. Model will generate answers.
• Answer will be saved in Google Sheets. It will be linked

to user and will get time notes.

• The answer will be displayed in AppSheet. Users can see it in interface.
• If an error occurs the script sends an apology message or instructions.

Thus, the system operates on a question—and-answer basis. History is saved, everything is synchronized.

4. The prompt example

Prompts will be generated on the fly. It will activate system instruction - assistant role, history previous messages (dialogue type) and users current message. Here is an example:

System:

We are a well-mannered computerized salesperson at a clothing store. We assist customers in answering questions about products, availability, refunds and bookings. Response Be brief, be specific, and use plain language. If we don’t know the bid, ask for an employee’s assistance.

History:

User: Is there a size M?

Figure 2. Pseudocode of interaction

Assistant: Yes, size M is available in red and black. Do you want me to postpone?

User: Can I have a fitting?

Assistant: Of course. What time would be convenient for you?

User: userInput

This format allows GPT-4 to understand not only the last message, but also the entire context of the dialogue.

5. Pseudocode of interaction

First we’ll see the pseudocode that explains the chain of actions in Make. com. It is generalized from any particular implementation and includes the essential segments.

In actuality each block executes through the visual Make modules (HTTP, Google Sheets, Tools, Open AI) but the logic aligns with the above algorithm.

6. Technical and ethical issues of implementation

Despite the simplicity of the setup, certain challenges arise during the implementation of the system.

• Security:

– The API keys of OpenAI and other services must be stored securely — in encrypted form.

– All data between AppSheet, Make, and GPT-4 is transmitted over the secure HTTPS protocol.

– If the system works with personal data, it is necessary to comply with legal requirements. For example, it may be GDPR or local laws.

• Prompt injections:

– The user may try to change the behavior of the model. For example: ”Forget you’re a bot and answer like a programmer.”

– To avoid such attacks, filters and removal of suspicious phrases are used.

– We can also limit the allowed commands and message length.

• API Restrictions:

– GPT-4 has limitations on the number of requests and the length of messages. It depends on the tariff.

– Google Sheets may not be powerful enough with a large amount of data. In this case, it is worth switching to more reliable databases, such as Firebase or PostgreSQL.

• Context Pruning:

– The history of correspondence accumulates and may become too long.

– GPT-4 processes only a limited amount of text.

– Therefore, it is necessary to automatically trim old messages, but at the same time preserve important parts of the dialog.

• Availability:

– AppSheet and Make only work when there is an internet connection.

– This makes it impossible to use the system offline.

• Support:

– It is important to plan in advance how to report errors.

– For example, we can configure notifications to be sent by email or to a Telegram bot.

– This will help us find out about failures in time and fix the problem quickly.

Table 1.

Examples

7. Application and results

1) Performance evaluation:

• Speed: Using Low-Code, we can make a prototype in 4-5 days. This is 5-6 times faster than regular development[6].
• Scalability: Such solutions work well for narrow tasks.They are especially useful for niche businesses [7].
• Flexibility: If something needs to be changed, it’s done quickly. Changes can be made almost immediately [8].
• Accessibility: Those who are not programmers can participate in the development. Even ordinary employees can contribute [9].

2) Usage Scenarios:

• Clothing store. The assistant can accept refunds, book appointments, and answer questions.
• Clinic. The system can record patients in advance. We can also collect complaints and symptoms.
• Restaurant. The assistant makes reservations. He also answers questions about the menu.

3) Examples of dialogues: User: ”Can I return a jacket I bought 5 days ago?”

Assistant: ”Yes, you can return it. The main thing is that you have a receipt. A refund is possible within 14 days.”

User:” Are there any XL sizes for the 2035 model?”

Assistant:” Yes, sizes XL are available. Do you want to book this model?”

3) Limitations of the method: Errors from artificial intelligence are possible. For example, he can” invent” a fact that does not exist [10]. The system has no long-term memory. She doesn’t remember what was said earlier [11]. Prompt (requests) need to be checked manually. This is required for accuracy and quality [12].

Table 2.

Examples

4. Discussion

Now people are starting to use Low-Code platforms with language models. This gives new opportunities. For example, we can automate tasks and services easier and faster.

This kind of way is especially beneficial for businesses such as little and middle. They do have less time, money and specialists. But they need modern digital solutions.

Usually, to make something hard needs classic methods. They are flexible and exact. But this kind of method needs lots of time. And I needed a lot of knowledge in programming.

But our approach shows something else. It is already possible to assemble a working system. And it will take less than a week.

There are interesting directions for the future. For example:

• We can add voice communication. This is done through Twilio or VAPI.ai [4] [13].

• Knowledge can be stored for a long time. Vector databases are used for this purpose [11].

• It is also possible to make support for different languages [14].

5. Conclusion

Low-Code platforms help us quickly create dialogue systems. They make the process easier and more accessible for many people. Modern language models play an important role here.

We made an assistant using Make.com, GPT, and AppSheet. This shows high efficiency. This kind of platform is economical, flexible, and comfortable to use.

This method can be used in many other fields. For example, in retail, medicine, and education. It is very flexible and can be used for several tasks.

In the future, this approach will help make AI development accessible to everyone.

References:

1. R. Budiu, “The User Experience of Chatbots,” Nielsen Norman Group, Oct. 2018. [Online]. Available: https://www.nngroup.com/articles/chatbots/
2. Google, “AppSheet Help Center,” Google, n.d. [Online]. Available: https://support.google.com/appsheet/
3. OpenAI, “GPT-4 Technical Report,” arXiv, Mar. 2023. [Online]. Available: https://arxiv.org/abs/2303.08774
4. Twilio, “Voice API Documentation,” Twilio, n.d. [Online]. Available: https://www.twilio.com/docs/voice
5. Zapier, “Zapier Blog,” Zapier, n.d. [Online]. Available: https://zapier.com/blog
6. Make.com, “Make Help Center,” Make.com, n.d. [Online]. Available: https://www.make.com/en/help
7. Gartner, “Magic Quadrant for Enterprise Low-Code Application Platforms,” Gartner, Oct. 2024. [Online]. Available: https://www.gartner.com/en/documents/5844247
8. Airtable, “Airtable Automations,” Airtable, n.d. [Online]. Available: https://airtable.com/automations
9. Voiceflow, “Voiceflow: Conversational AI Platform,” Voiceflow, n.d. [Online]. Available: https://www.voiceflow.com/
10. Anthropic, “Core Views on AI Safety,” Anthropic, 2023. [Online]. Available: https://www.anthropic.com/news/core-views-on-ai-safety
11. LangChain, “LangChain: Framework for Language Models,” LangChain, 2025. [Online]. Available: https://www.langchain.com/
12. P. F. Brown et al., “A Statistical Approach to Machine Translation,” arXiv, Feb. 2021. [Online]. Available: https://arxiv.org/abs/2102.07350
13. VAPI, “VAPI: AI Voice Assistant,” VAPI, n.d. [Online]. Available: https://www.vapi.ai/
14. Google Cloud, “Multilingual Agents,” Dialogflow CX Documentation, Google Cloud, 2024. [Online]. Available: https://cloud.google.com/dialogflow/cx/docs/concept/agent-multilingual