Upload documents in many formats — prose (.txt, .md, .html, .xml, .pdf, .docx, .xlsx, .pptx, .rtf, .epub, .odt, .ods, .odp, .pages, .numbers, .key) and source code (.java, .kt, .py, .js, .ts, .go, .rs, .swift, .cs, .cpp, .c, .h). Sophia processes them into searchable segments with vector embeddings. The admin panel offers inline and batch tag editing, directory autocomplete on upload, and one-click retry for documents that failed indexing.

Organize your knowledge into isolated agents, each with its own prompt template, document tags, RAG configuration, and agentic behaviour. Agents ensure strict data separation — users in one agent cannot access documents from another. Serve different audiences (customers, internal staff, partners) from a single Sophia instance.

For complex questions, Sophia operates as an autonomous agent: it iteratively calls tools (knowledge search, file retrieval, persistent context save/load), reasons over the results, and composes a comprehensive answer. Users see the reasoning process in real time. Per-agent options let you tune the loop: max iterations, search-and-fetch preamble, compact search payloads, agentic context management, prompt caching, and extended thinking.

Enable Collaborative Mode on an agent to let Sophia render rich interactive HTML artifacts directly in the chat — diagrams, calculators, mini-dashboards, REST micro-tools — and to attach 2–4 follow-up buttons after every answer for guided exploration. Both the artifact tool (show) and the follow-up tool (ask) accept domain-specific descriptions you can override per agent.

Switch any agent to Deck View to replace the linear chat with a horizontal card deck — each interaction becomes a swipeable card with prompt above and answer below. Useful for guided tours, lesson decks, and demo flows.

Cookie-based authentication secures the web interface, while JWT tokens provide programmatic access for APIs and MCP clients. API tokens can be scoped to specific agents and tags, ensuring knowledge segregation extends to every access path. Private agents require either an OAuth flow (Claude Desktop) or an API token; public agents work without authentication.

Sophia is invitation-only: there is no public signup. Administrators add users by email, and Sophia sends a transactional activation message through AWS SES. The recipient picks their own password, accepts the legal policies and is automatically logged in — administrators never see nor handle passwords. A status lifecycle (invited → active, plus disabled for revocation) is enforced at login time by a server-side vetoer.

A single Sophia instance can host multiple tenants (organisations) without any data leak between them. Each agent, knowledge document, user, chat session and API token is scoped to a tenant via server-side ACL filters. A tenant-admin role manages every resource of their own tenant; a global admin spans all tenants. Public agents and shared knowledge can be opted into explicitly.

Terms of Service, Privacy Policy and Cookie Policy are versioned and hosted externally on restheart.com. At first login Sophia presents a blocking modal capturing explicit consent — including the separate approval of art. 1341 c.c. clauses required under Italian law. Consents are recorded on the user document with version, URL, timestamp and IP address as immutable audit evidence. A major-version bump of any policy automatically forces all users to re-accept before continuing.

Sophia exposes a Model Context Protocol (MCP) server, allowing AI clients such as Claude Desktop, Cursor, Claude Code, and VS Code to connect directly to your knowledge base. The admin panel generates ready-to-paste configuration snippets and the server supports OAuth client_credentials for transparent re-authentication.

See Sophia MCP Server for setup instructions.

Track token usage and dollar cost per chat, per agent, per model from the Costs dashboard. The agent editor shows a live estimate of cached prefix size and per-interaction cost based on the chosen model and prompt-caching support. When a Bedrock model is throttled or fails, Sophia automatically falls back to a configurable list of secondary models — the user sees a model_fallback event and the answer continues seamlessly.

The responsive chat interface works across devices and can be embedded into existing websites via iframe. Hint questions can be shown as clickable chips at session start. Multi-language support (English/Italian) and customizable themes let Sophia match your brand.