Development of a Web Service for Creating Digital Human Replicas

Task

The main task set by the client was very ambitious and extended beyond standard chatbot development. The goal was to create a “digital personality imprint” — a replica of a specific person in the form of a chatbot that would not only answer questions but also perfectly replicate this person’s communication style, use facts from their biography, and demonstrate their key personality traits.

The ultimate goal of the system, as envisioned by the client, was to enable relatives and loved ones to continue “communicating” with the person after their passing, preserving their digital memory.

To put this concept in terms of technical and business tasks, the project needed to solve the following key objectives:

1. Provide deeply personalized AI conversations: Ensure that the AI goes beyond general responses, delivering conversations that are consistent, natural, and based on a unique personality profile.
2. Implement seamless onboarding (Wizard): Develop an intuitive multi-step survey form (wizard) to collect specific data about the user. This data — ranging from favorite books and biographical facts to political views and communication style — would serve as the foundation for the AI's “personality.”
3. Create a reliable contextual retrieval mechanism (RAG): Build an architecture that allows the AI to dynamically extract relevant facts from the created profile in real time, generating appropriate responses during conversations.

Initially, the target audience appeared narrow: elderly individuals and patients with serious illnesses who might wish to leave behind this type of digital footprint. However, upon further analysis, it became clear that the potential audience was much broader, encompassing any end user interested in creating their own personalized digital twin for communication for various reasons.

Solution

To accomplish such a complex task, we proposed a modern and modular architecture.

The key element of the solution was a sophisticated RAG (Retrieval-Augmented Generation) flow. Instead of simply sending the user’s query to an LLM (as simple chatbots do), our system performs a multi-step process:

1. It analyzes the user's query.
2. Instantly performs a semantic search of the vector database (Qdrant) where all the facts about the user and their personality traits are stored.
3. Finds the most relevant pieces of information (e.g., “fact about childhood,” “opinion about music”).
4. “Augments” the user's initial query with this information.
5. Sends this enriched, highly contextual prompt to the LLM (OpenAI) for response generation.

This approach ensures that the AI “remembers” who it is and responds according to its “personality.”