What's New in MDDB 2.19

MDDB 2.19 is about the assistant that answers questions over your collections. Until now its behaviour lived in a TOML file next to the chat service: changing one word of the prompt meant editing a file and restarting. Now the assistant is data in MDDB, edited in the panel, and the chat service picks up a change at the visitor's next message.

The release also changes the licence, lets you switch off the modules you do not use, and closes the gap between what the server can do and what the panel shows. The full list is in the changelog.

A new licence #

From 2.19.0 the server, mddb-cli, the panel and the chat service are under the Business Source License 1.1. Using MDDB in your company, on your servers, for your own products, stays free. Offering MDDB itself to other people as a hosted or embedded service that competes with Tradik's needs a commercial licence. Each version becomes Apache 2.0 four years after its release.

The client libraries, the integrations, the proto files and the chat widget stay BSD-3-Clause, so code you write against MDDB is not affected. Releases up to 2.18.10 stay BSD-3-Clause. LICENSING maps every directory to its licence.

The assistant is configured in the panel #

The panel has a Chat page. A scenario is one assistant: what it is for, what it must never do, how it answers questions outside its scope, which collections it may search, its tone, model and limits. The page shows the system prompt exactly as the server composes it, with its size in tokens.

The chat service loads the scenarios from MDDB and checks for changes every ten seconds. The scenarios in its TOML file still work as a fallback, so an existing deployment keeps running unchanged.

A scenario's collection list is now enforced where the tools run. Before 2.19, a model that was confused or prompt-injected could ask the search tool for any collection the chat service could read.

Skills #

A skill is a reusable block of instructions in the SKILL.md format: a name, a description and the instructions. Attach skills to scenarios in the order you want. They go into the prompt after the guardrails, under a line saying the guardrails win, so a skill cannot override them.

You can paste a skill or import it from a URL. An import shows the full text before anything is saved, and a GitHub link is pinned to a commit. It is fetched again only when you press Update from source, which shows a diff first. Every change keeps the previous version.

Signed-in users keep their conversation #

A host application that already knows who its user is can sign a short-lived token for the chat. With that token, the conversation is stored in MDDB. The user can close the tab, switch devices, or meet a restarted chat pod, and the assistant still remembers what was said. Users can list their conversations, start a new one, or delete all of them. A host can export or delete everything of one user in one call.

Visitors without a token are anonymous, as before. A dropped connection no longer loses their conversation either: it waits for them until the session expires.

What people ask, and which answers are bad #

A scenario can log its conversations. Logging is off by default, and e-mail addresses, tokens and card numbers are masked before anything is stored. The Log tab on the Chat page lists turns with filters for thumbs-down, searches that found nothing, refusals, scenario and date. Each turn shows the question, the answer, the searches it ran, what they returned and how it was rated. Visitors can add a comment to a thumbs-down.

The Docs gaps view groups questions whose searches found nothing or found only weak matches, most frequent first. It is a list of what to write next.

Knowledge from other MCP servers #

A scenario can also use read-only tools of other MCP servers, for things a collection cannot hold well: order status, open tickets, the state of a service. MDDB checks every tool when you save the scenario and refuses any tool that is not declared read-only. The calls go through MDDB, so the chat service never sees the credentials, and the results reach the model marked as outside data. It is off until you turn it on.

Switch off what you do not use #

The new Features page lists the optional modules: chat, geo, temporal, the spell checker, automation, webhooks, schemas, aliases, code intelligence and more. A module that is off disappears from the menu, its API answers "feature disabled", and its MCP tools leave the tool list. Data stays where it is. The page groups modules like the menu and can turn a whole group on or off at once. An operator can force modules off with MDDB_FEATURES_FORCE_OFF.

The menu itself is now grouped and has a search box, because it had grown to 34 entries.

The panel reaches everything #

An audit of the API against the panel found 24 endpoints the panel never called. They all have a screen now: agent memory, collection schemas, bulk import jobs, personal API keys, facets, duplicate detection, Wikipedia import, batch delete, checksums, backup now and restore, TTL and classification in the document viewer, and more tools on the geo, temporal, spell and code pages. Wiring them found nine server bugs, all fixed in this release.

Upgrading #

Nothing to migrate. Every new field is optional, and every new module is on by default except external MCP servers.

A few things behave differently:

  • The licence changes as described above.
  • mddb-chat's session.end webhook fires when a session ends or expires, no longer when the visitor's connection drops.
  • mddb-chat's /health answers 503 until it has reached mddbd, so a container is unhealthy until then. It no longer exits when mddbd is not up yet.
  • mddb-chat needs [mddb] http_url when its gRPC address does not use port 11024. Before, it guessed the login address and got it wrong.
  • /v1/delete-batch answers notFound. The old not_found is still sent and will be removed in a later release.

Download 2.19.1 (the first published build of 2.19) from GitHub releases, or run mddb-cli self-update.