Sophia uses Claude (default) from AWS Bedrock to deliver intelligent, context-aware conversations. RAG combines real-time document retrieval with advanced language processing. Administrators customize prompt templates per agent, and the system maintains conversation context either via classic history injection or via a persistent agent-context document (see Agentic Context Management below).

When enabled, Sophia iteratively uses tools — searching the knowledge base, reading files, listing tags or paths — before composing the final response. The loop runs up to maxAgentIterations rounds and tool events stream to the client when streamThinkingEvents is enabled. Optional features include:

A persistent markdown document keyed by chatId, replacing classic history injection. At turn start the saved context is auto-loaded as a synthetic context_load tool result; at turn end a forced synthetic call summarises the turn via context_save, context_append, or context_skip. This frees the main loop from history-token overhead and keeps context coherent across long sessions.

Adds show (in-iframe HTML artifacts) to the model’s tool set during the main loop, and a forced post-answer ask call that constrains the model to attach 2–4 follow-up buttons. Per-agent, the description of both tools can be overridden to teach the model domain-specific patterns. A per-message toggle lets the user activate it for individual queries even when the agent default is off.

Documents are uploaded into MongoDB GridFS (docs.files). Apache Tika parses prose formats — .txt, .md, .html, .xml, .pdf, .docx, .xlsx, .pptx, .rtf, .epub, .odt, .ods, .odp, .pages, .numbers, .key — and code-aware splitters handle source files (.java, .kt, .py, .js, .ts, .go, .rs, .swift, .cs, .cpp, .c, .h). Each file is split into segments, embedded with Titan Embed Text v2, and persisted in textSegments. The admin panel lists documents in flat and tree views with filename / tag / path filters; failed indexing surfaces an inline retry button.

Agents are the core abstraction for knowledge partitioning. Each agent specifies a set of tags that act as mandatory filters on every vector search — users in one agent cannot access documents belonging to another. Combined with private-agent access control (see below), this provides enterprise-grade data isolation.

Authentication uses cookie-based sessions (POST /token/cookie) for the web interface and JWT bearer tokens for API and MCP clients. JWTs carry custom claims (contexts, tags) used for fine-grained authorization. Private agents require a JWT containing the agent id in the agents claim — public agents are reachable without authentication. API tokens are revocable and verified on every request by a low-priority vetoer.

A web-based administration interface provides full management capabilities:

WebSocket messaging via MongoDB Change Streams pushes updates to the chat client. Streaming response delivery shows tokens as they are produced. In agentic mode, tool execution events are streamed (tool_start, tool_result, text_start, model_fallback, render_step) with arguments, summaries, and durations. A streaming watchdog and document-recovery routine keep the UI consistent if the WebSocket drops.

RESTHeart on Java 25, MongoDB Atlas with Atlas Vector Search, LangChain4j integration to AWS Bedrock (Claude / Nova / Nemotron) and Titan Embed v2. Apache Tika for document parsing. Authentication via mongoRealmAuthenticator + jwtTokenManager with cookie support and OAuth client_credentials (for MCP).

Angular 21 standalone components, signals, Tailwind CSS v4, spartan-ng. Three.js powers the optional avatar orb and animated face. Reactive Forms drive the admin editors. Charting via lightweight inline SVG.

Sophia is delivered as a managed RESTHeart Cloud — dedicated service. SoftInstigate handles infrastructure, scaling, monitoring, and updates. As a cloud customer you administer your own instance through the admin panel — see the Administrator Guide.