Skip to content

AI Support 2.2.0

Choose a tag to compare

@mlopezFC mlopezFC released this 20 Oct 22:04
· 5 commits to main since this release

A large release. The assistant now reaches beyond the application: the chat bubble can be embedded in
any website, it talks to Anthropic and Gemini models as well as OpenAI and Ollama, and it records,
hears and transcribes voice notes.

Embedded chat

  • The chat bubble is now published as the <ai-support-chat> web component, so it can be placed
    on any site — a static marketing page included — and answered from the application like any other
    conversation.
  • Web channels tie each embedding site to an assistant. Only a site with an enabled web channel
    can load the chat, and turning the channel off takes the chat off the site.
  • Visitors are anonymous: each browser gets its own cookie and is named visitor- plus a short hash.
  • The embedded chat is built in the embedding page's language, sends the page's context with each
    message, and can be dragged to a corner, which it remembers.
  • Hosts can switch off optional features and suppress notifications, and the chat works inside a
    shadow root.
  • Visitor sessions are purged after a while, with bounds on how many messages a visitor sends and on
    the length of messages and context.
  • Setup: serve it from a host on the site's own domain, let /web-component/** through the security
    configuration, and have Vaadin scan com.appjars.aisupport.flow.embed. The README has the full
    procedure.

Models and providers

  • Anthropic and Gemini models are now supported, alongside OpenAI, OpenAI-compatible servers and
    Ollama.
  • Each LLM now stores the content types its model accepts, prefilled from a built-in model
    catalog and editable in the LLM form.
  • An LLM's configuration can be tested from its form.
  • An attachment a model does not accept is now sent as text, or as a note when it cannot be.
  • Token estimates now cover OpenAI models the tokenizer does not know.

Voice notes and transcription

  • Users can record and play voice notes where the chat offers attachments.
  • Audio is transcribed by the system LLM for models that do not hear it, and the transcript is
    stored with the attachment, shown under the audio, and used wherever the text of a message is read
    — retrieval, moderation and the session summary.
  • Audio is sent in the formats each provider reads, and its tokens are estimated by how long it
    plays.

Attachments and chat memory

  • The chat now accepts Markdown, CSV and Excel attachments.
  • Documents a model does not read are sent as their text, long PDFs included, with PDF pages capped
    per turn.
  • A turn's attachments are fitted to a token budget, and the uploader is told when text was cut.
    Administrators are told when a channel attachment could not be fully read.
  • The memory budget now defaults to 25,000 tokens, and a turn the model rejected is kept out of
    the memory.
  • Model errors reach the inspector with their reason and without credentials or quoted secrets.

Private messages and session summaries

  • Every assistant prompt is now wrapped in fixed rules, and private messages are kept out of
    answers everybody sees and out of users' message search.
  • Administrators can send private messages from the chat bubble, and switch a session between
    automatic and manual mode.
  • Session summaries are now generated in batches on their own executor, instead of after every
    answer.

Channels

  • Host applications can now add channel types of their own, with no persistence code.
  • Repeated or concurrent channel messages are kept in one session, a channel is deleted and its
    sessions unlinked in one transaction, and WhatsApp Coexistence echoes are read as the business's
    own messages.
  • Replies longer than a channel accepts are sent in parts, a channel is marked failed when its
    provider answers 401 or rejects an upload, and WhatsApp media is streamed so the size limit stops
    large files.

Administration views on phones

  • The administration views, filters, dialogs, content viewer, vector search, inspector and tools
    picker now fit on phones, with dialogs shown full screen and a row's hidden columns shown on a
    tap.
  • Session and message search results are shown in tabs with their counts, the extra document actions
    are gathered in a menu, and the message input's actions collapse into an overflow menu.

Fixes

  • The embedding store database now defaults to spring.datasource.url.
  • Enum columns no longer generate CHECK constraints, and the logging backend is left to the host
    application.
  • AI Support starts without an application UserProfilePictureProvider, and the queries of the
    largest tables are indexed.

Upgrading

Databases created by earlier versions keep check constraints and a narrow configuration column that
ddl-auto=update never changes. Run once, before creating an Anthropic or Gemini LLM, a web channel
or a channel of your own type:

ALTER TABLE as_llm DROP CONSTRAINT IF EXISTS as_llm_type_check;
ALTER TABLE as_llm_exchange DROP CONSTRAINT IF EXISTS as_llm_exchange_llm_type_check;
ALTER TABLE as_llm ALTER COLUMN configuration TYPE text;
ALTER TABLE as_channel DROP CONSTRAINT IF EXISTS as_channel_channel_type_check;

Transcripts are stored in the new transcript column of as_message_attachment, which
ddl-auto=update adds; a schema managed otherwise needs it added once.

  • memory.max-tokens now defaults to 25000 instead of 10000. Set it to 10000 to keep the
    previous size, and lower it for a small-context model.
  • Answering no longer regenerates the session summary. The chat views request summaries themselves;
    code that adds messages outside them calls AssistantService.summarizeSession.
  • ChannelService, ChannelDao, MessageAttachmentService and MessageAttachmentDao gain methods.
    Applications that implement any of these interfaces themselves must implement them.
  • A session open before the upgrade keeps the private messages its memory already holds until it is
    closed. DELETE FROM as_chat_memory clears them right away, at the cost of the earlier context of
    those conversations.
  • The enum-typed channel accessors and PromptDto.system are deprecated; ChannelDto.getType()
    returns the type key of every channel.

Requires Java 21, Spring Boot 4.x, and Vaadin 25.2 for the UI layer, and appjars-utils 2.1.0.
The backend and service layers of an AppJar do not require Vaadin.

Free mode: every AppJar can be used free of charge. Free mode restricts how much you can run
through it, and holds some capabilities back for a licensed installation. Installing a license lifts
the restrictions with no code changes. What free mode allows for each AppJar is listed here:
https://docs.appjars.com/licensing/#free-mode-limits

Getting started: https://docs.appjars.com/ai-support/getting-started/
Documentation: https://docs.appjars.com/ai-support/overview/
Pricing: https://www.appjars.com/catalog/ai-support/