The rise of Large Language Models (LLMs), such as ChatGPT, has revolutionized the way humans interact with machines. This advancement has enabled the development of applications that were previously considered unattainable, particularly in domains requiring a deep and contextualized understanding of natural language.
Our study investigates the feasibility of an artificial intelligence assistant capable of providing personalized user support by tracking preferences, personality, and short-, medium-, and long-term goals. The project aims to create a system that acts as a personal advisor, combining the contextual reasoning capabilities of LLMs with information storage and decision-making techniques.
One of the challenges of the study lies in integrating traditional business logic with the flexibility of LLMs. Unlike conventional technical problems that can be solved with well-defined algorithms, human-related issues involve subjectivity, emotions, and an infinite range of contexts. As a result, we had to explore new architectural paradigms and methodologies to make this solution viable.
Natural Language Processing (NLP) research has been ongoing for decades, but significant advancements have only emerged in recent years with the introduction of transformer-based architectures, as described in the paper “Attention is All You Need” (2017). This approach allowed language models to retain complex contextual information, making them more capable of understanding and generating coherent text.
Since then, various studies have refined techniques for interacting with LLMs, including prompt engineering, the use of vector databases for context retention, and integration with external systems to enhance responses. However, a major challenge remains: LLMs, on their own, lack persistent memory and are prone to generating inconsistent or hallucinated responses.
Given this scenario, interesting questions arose during this study:
- How can an artificial “memory” be created so that the assistant remembers past interactions?
- How can the AI ensure that its responses align with the user’s personality?
- How can misinterpretations and inadequate responses be mitigated?
- What strategies can be used to efficiently integrate LLMs with traditional systems?
Our study seeks to answer these questions by developing an architectural model for AI-based user-assistant interactions. Several technologies and frameworks were explored, with the solution built on a modular architecture, integrating:
- Advanced LLMs: We evaluated models such as GPT-4 and LLaMA, considering cost, latency, and accuracy for different tasks.
- Vector Databases (Faiss): Used to store interactions and retrieve relevant context in real time, enabling a pseudo-memory system.
- Backend in Python: Manages communication between the frontend, AI models, and storage modules.
- Frontend in React.js: Provides the user interface for system interaction.
- Prompt Engineering Frameworks: Various approaches were tested to structure messages sent to LLMs, optimizing response quality.
- Integration with External Systems: The system can access and process data from emails, calendars, and other sources to offer more personalized support.
Our study combines these technologies to create a digital assistant that offers an interactive and humane experience.
Study Details
To bring the project to life, we explored various architectural and methodological approaches aimed at integrating LLMs into a system capable of continuous and personalized user interaction. This section details the challenges faced, the solutions implemented, and the key results achieved. The study was structured around three main areas:
Memory and Context Persistence
Since LLMs lack intrinsic memory, we developed a persistence mechanism based on vector databases and relational storage. Each user interaction is stored, categorized, and retrieved on demand to ensure that the assistant retains relevant information over time.
The system successfully retrieved past interactions and used them to personalize responses, ensuring a continuous and coherent dialogue over time.
Integration with Traditional Systems
We implemented a backend capable of interacting with external APIs, such as emails, calendars, and databases, allowing the assistant to provide personalized suggestions and reminders.
The system was able to access and process user calendar events, suggest reminders, and proactively organize appointments.
Prompt Engineering Techniques
Multiple prompt engineering techniques were experimented with, for example zero-shot, one-shot and few-shot prompting, used according to the context, required “prompt injection” protection, and with related costs and response delays in mind.
Accuracy and Response Formatting
With the implementation of validation rules, we significantly reduced the rate of incorrect or poorly formatted responses, ensuring greater predictability in user communication. We also added multiple techniques to avoid what we call “prompt injections”, where the user asks from the system improper or “out-of context” requests.
This study demonstrated that it is possible to integrate LLMs with traditional systems to create intelligent and personalized assistants. The combination of vector databases, prompt engineering, and API integration helps enhance language models, paving the way for new applications.
However, challenges remain, such as improving structured reasoning capabilities, reducing computational costs, and expanding to voice interaction.
We believe that with the advancement of AI technologies, more natural and efficient interactions between humans and machines will become commonplace.
