Multi-Agent Conversational Assistant
Client
Digital Transformation consulting company
Sector
IT Services
Revenue
€ 2M
Client
Digital Transformation consulting company
Sector
IT Services
Revenue
€ 2M
From challenge...
In asynchronous learning paths, students and workers often lack immediate support when they need clarification on course content. This can interrupt the learning flow, reduce engagement and increase repetitive requests to teachers and tutors.
The challenge was to introduce an assistant capable of providing reliable answers based on educational materials, while reducing the hallucination risk typically associated with LLM-based chatbots. From a user experience perspective, the system also needed to handle free-form inputs and advanced conversational flows, including editing previous messages and continuing the conversation coherently from that point.
...to result
The conversational assistant integrated into WhoTeach enables immediate, contextualized and personalized learning support. Through a RAG pipeline, answers are generated from the official materials uploaded by teachers, improving quality, relevance and control over course-specific content.
Personalization based on the VARK model allows the assistant to adapt the response format to the user’s learning profile, offering structured explanations, summaries, examples, diagrams, images or audio. The result is a smoother learning experience, with just-in-time support throughout asynchronous courses.
Results achieved: 30% reduction in the average time needed to obtain a useful answer on course content; 20% reduction in repetitive requests to teachers and tutors on topics already covered in the materials; 15% increase in module completion thanks to immediate and personalized support.
From challenge...
In asynchronous learning paths, students and workers often lack immediate support when they need clarification on course content. This can interrupt the learning flow, reduce engagement and increase repetitive requests to teachers and tutors.
The challenge was to introduce an assistant capable of providing reliable answers based on educational materials, while reducing the hallucination risk typically associated with LLM-based chatbots. From a user experience perspective, the system also needed to handle free-form inputs and advanced conversational flows, including editing previous messages and continuing the conversation coherently from that point.
....to result
The conversational assistant integrated into WhoTeach enables immediate, contextualized and personalized learning support. Through a RAG pipeline, answers are generated from the official materials uploaded by teachers, improving quality, relevance and control over course-specific content.
Personalization based on the VARK model allows the assistant to adapt the response format to the user’s learning profile, offering structured explanations, summaries, examples, diagrams, images or audio. The result is a smoother learning experience, with just-in-time support throughout asynchronous courses.
Results achieved: 30% reduction in the average time needed to obtain a useful answer on course content; 20% reduction in repetitive requests to teachers and tutors on topics already covered in the materials; 15% increase in module completion thanks to immediate and personalized support.
The solution
We introduced a generative AI conversational assistant integrated into the WhoTeach platform and designed to support students and workers during asynchronous learning paths. The assistant retrieves the most relevant content from teacher-uploaded materials through a RAG pipeline, generating answers grounded in the course knowledge base.
The solution preserves conversational context, guides users in understanding course topics and personalizes the response format according to the VARK model. The multi-agent architecture orchestrates specialized components for input understanding, retrieval, response generation, format recommendation and multimedia output production.
What we implemented:
- RAG pipeline to retrieve relevant content from official course materials
- Grounded answer generation to reduce hallucination risk
- Multi-agent orchestration for input understanding, routing, retrieval and output generation
- VARK-based personalization with structured explanations, lists, examples, diagrams, audio and images
- Conversational state and chat context management
- Ability to edit previous messages and continue the conversation coherently
- Cloud-native service architecture, scalable and integrated with WhoTeach
Technology stack
The solution is built on a cloud-native service architecture, with containerized deployment on Google Cloud Run and REST APIs exposed through FastAPI. PostgreSQL on Cloud SQL manages the student profile and conversational state, Firestore supports message management, and Google Cloud Storage stores generated assets such as audio, images and diagrams.
The AI layer combines LangGraph and LangChain for multi-agent orchestration, OpenAI GPT-4o for response generation and ChromaDB as the vector database supporting the RAG pipeline. Retrieval uses text-embedding-3-small embeddings and semantic chunking, while Google Cloud Text-to-Speech and Vertex AI Imagen 4 Ultra generate audio and image outputs to support the explanation.