Sophia utilizes the Claude model from AWS Bedrock to deliver intelligent, context-aware conversations. The system implements a Retrieval-Augmented Generation (RAG) approach that combines real-time document retrieval with advanced language processing. Administrators can customize prompt templates to tailor the AI’s behavior and tone, while the system maintains conversation context through intelligent chat history management.

When enabled on a context, Sophia operates as an autonomous agent that iteratively uses tools — searching the knowledge base, reading files, and gathering information — before composing a final response. The agent loop runs up to a configurable number of iterations, and tool execution events can be streamed to the client in real time so users see the reasoning process as it happens.

The platform supports comprehensive document management with compatibility for TXT, MD, PDF, and HTML files. Documents are automatically processed through vector embeddings using Amazon Titan Embed Text v2 for semantic search. The admin panel provides a web interface for uploading, browsing (flat and tree views), filtering, and deleting documents.

Contexts are the core abstraction for knowledge partitioning. Each context defines a set of tags that act as mandatory filters on every vector search — ensuring that users in one context cannot access documents belonging to another. This provides enterprise-grade data isolation without requiring separate deployments.

Authentication uses cookie-based sessions for the web interface and JWT bearer tokens for API and MCP access. Role-based access control distinguishes between admin users (who manage contexts, knowledge, and tokens) and regular users (who interact with the chat). API tokens can be bound to specific contexts and tags, ensuring that programmatic access respects the same knowledge segregation boundaries.

A web-based administration interface provides full management capabilities:

WebSocket-based messaging with MongoDB Change Streams provides live updates. Streaming response delivery allows users to see AI responses as they’re composed. In agentic mode, tool execution events are streamed in real time, showing tool names, arguments, results, and durations.

The backend is built on RESTHeart framework running on Java 25. MongoDB serves as the primary database with Atlas compatibility for cloud deployments. LangChain4j integrates with AWS Bedrock for Claude LLM access and Amazon Titan for embeddings. Security is handled through JWT authentication with cookie-based session support.

The frontend is built with Angular 21 using standalone components and signals for state management. Tailwind CSS v4 provides styling, with spartan-ng UI components. The admin panel uses responsive layouts with mobile support.

Sophia deploys as a managed RESTHeart Cloud service. The system uses Docker containers for consistent environments. As a cloud service, all infrastructure concerns — scaling, monitoring, updates — are handled automatically.