Sophia is an AI-powered knowledge base assistant built on top of Ulabase. It turns static documentation (PDFs, Markdown files, HTML pages, Word documents, spreadsheets, and more) into a conversational agent that answers questions and cites the source documents it draws from. It uses RAG (Retrieval-Augmented Generation) combined with large language models (Claude by Anthropic, Amazon Bedrock) to find and synthesize relevant content from your knowledge base.

Past releases were about engineering work inside Sophia. This one is about the product: its own domain, the launch of the commercial offering, a clearer name for what users configure, and the operational pieces a customer needs, which are multi-tenant administration, invite-based onboarding and legal compliance.


A New Home at bysophia.ai

Sophia now lives at bysophia.ai. The previous addresses on sophia.restheart.com continue to work, and new development happens on the new domain. The two public instances we run are reachable at:

The API endpoint moved to api.bysophia.ai, together with the MCP server endpoint and the JWT issuer. The previous sophia.restheart.com addresses remain active for unauthenticated access: agents, chat, and MCP all work there. Authentication now requires bysophia.ai. To migrate, update the base URL.

A New Home at bysophia.ai


Sophia AI Is Now Commercially Available

This month we also launched the commercial offering. Companies can now buy Sophia, and we have published a pricing model for both the SME and Enterprise segments.

For small and medium businesses the monthly fee is €150 for the first agent and €100 for each additional one. It includes infrastructure, software updates, and weekday business-hours support. We can also assist with initial setup and document ingestion if needed.

For larger organizations the RESTHeart + Sophia AI Enterprise subscription bundles one or more RESTHeart production instances with up to 5, 20, or unlimited Sophia agents, with progressively tighter SLAs on response time and a larger number of authorized contacts.

We price per agent on purpose. A single agent serves an entire team or audience, so the number of people who open it says little about its value. Customers also plan a deployment agent by agent.


"Contexts" Became "Agents"

The most visible product change in May is also the smallest in code: we renamed the user-facing entity from Context to Agent.

Users build assistants: an agent for tax law, an agent for the RESTHeart docs, an agent for an internal product team. The old name described the implementation, a grouping of prompt, scope, and options. The new one describes what the user builds. Every URL, every label in the admin panel and every reference in the docs now says agents.

The MCP tools got the same treatment. The redundant sophia_ prefix is gone (sophia_search is now search). The two interactive-artifact tools have been renamed from render_input / render_output to ask and show, names that describe what the agent is doing. Models pick the right tool more reliably when the name carries the intent.


Multi-Tenant Administration

Sophia now supports multiple tenants in a single deployment. Alongside the existing global admin, there is a new tenant-admin role. It has the same administrative capabilities (agents, knowledge, users, tokens, chat history), scoped to a single tenant's data. Tenants don't see each other.

We needed this to offer Sophia as a managed service. Each customer can now administer their own agents, invite their own users, and inspect their own costs, while we, as platform operators, retain a global view. The isolation is implemented entirely with RESTHeart's native readFilter / writeFilter ACL rules, with the tenant injected from the user's JWT.


User Invitation, Password Reset, and Profile Page

User onboarding has been rebuilt on the pattern Ulabase already uses. Admins no longer set a user's password out of band. They send an invite: the user receives an email with a single-use token, picks their own password, and accepts the terms of service in a single step. The same mechanism powers the self-service password reset flow. Auto-registration remains disabled by design: Sophia is invite-only.

Around this flow we added server-side password strength enforcement, brute-force protection on login and rate-limited password reset requests. An admin can resend a pending invite without recreating the user. Emails go through Ermes, the same SMTP wrapper Ulabase uses in production.

Every authenticated user, regardless of role, now has a profile page at /profile. There they see who they are logged in as, change their password, log out, and check which agents they have access to. Until last month this page did not exist.


Legal Compliance and Consent Tracking

With the commercial launch we needed Terms of Service and a Privacy Policy. The policies are now hosted on restheart.com (shared infrastructure with Ulabase) at versioned, immutable URLs like /legal/terms/sophia/1.0/. Sophia stores only the consent record: which versioned URL the user accepted, when, and from which IP.

A new server-side interceptor blocks any request from a user whose accepted policy version is outdated, after a major policy change. Users see a re-acceptance screen, agree, and resume, with no support intervention.


Cheaper Turns with Model Profiles

Per-phase model usage breakdown for a single turn

April's release cut cost through prompt caching and the search preamble. May adds a new mechanism: model profiles, which assign different models to different phases of the agent loop.

A single turn calls the LLM up to six times: planning, the main reasoning loop, synthesis, and the terminal tool calls that save context or render an artifact. The main loop benefits from prompt caching and needs the primary model. The terminal calls, context_save and ask, together account for roughly 73% of the input cost of a typical turn, and nothing downstream reuses their prefix. They do not need a premium model.

The new economy profile keeps the main loop and planning on the agent's chosen primary model (e.g. Claude Sonnet 4.6) and routes the terminal phases to a cheaper fallback like Nova Lite. In our telemetry this cuts the input cost of a typical turn by 60–70%, and we measured no regression in answer quality.

An ultra-economy profile is also available for high-traffic agents that can tolerate a small accuracy loss on planning. Per-phase token usage is now tracked separately, so administrators can see where the savings come from.


Self-Introspection for the Agent

Two new tools, interactions and interaction, let the agent look back at its own chat. The first returns a lightweight manifest of all previous interactions in the current conversation. The second returns the full step trace of a specific one, including the HTML artifacts the agent produced earlier.

For the user, the agent can now reference an artifact it generated three turns ago without generating it again. It can also answer "what did you tell me about X earlier?" without loading the entire transcript back into context. The tools are exposed only inside Sophia's own loop. MCP clients such as Claude Desktop manage their own history, so the tools are absent from MCP.


Looking Ahead

Next month we work on the commercial side: onboarding the first paying tenants, improving the public home page, and finishing the re-engineering of document injection, which should further cut cost on agents with very large knowledge bases.

Build your app on Ulabase

Your backend is ready in minutes, and you spend your time on your app. The Free plan needs no credit card.