Changelog
All notable updates and changes to the LoyJoy Conversational AI and Marketing Platform are documented here.
unreleased
release-2025-08-24
- Legacy NLU ist now configured as a billing product, not via a feature flag in the tenant.
- Added support for OpenAI-compatible
/v1/chat/completions
endpoints. Removed vLLM, Fireworks.ai and Together.ai, as those now can be configured as such OpenAI-compatible endpoints. - All AI models provides with exception of Cohere and Ollama now are supported by the
Agentic AI
process module. This now enables complex agentic AI tool-calling and reasoning implementations with all kinds of LLM providers.
release-2025-08-23
- Added instructions to
Agentic AI
process module as a replacement for system message. - Process module
ReachFive
can now read, assure and create up to 3 consents in one request. - Performance optimizations in the widget for very long conversations.
- Removed the setting that certain tenants might have AI capabilities disabled.
release-2025-08-22
- Refactored AI model select, which now shows pricing information and deployment region, such as European Union.
- Added Scaleway gpt-oss-120b
- Updated Anthropic models to newest versions.
release-2025-08-19
- The MCP server now requires the configuration of an
Agentic AI
process module in the agent. This module allows to configure the tools that should be made available via the MCP server. If no suchAgentic AI
module exists, the MCP endpoint will throw an HTTP status 400 error. - Added several model checking rules that ensure the correct configuration of the
Agentic AI
process module.
release-2025-08-18
- Process module
AI Agent
now represents tool configurations in an accordion with detailed configuration options for each tool. - The variables REST API is now disabled by default. It will throw an HTTP status 403 when called, and can be enabled in the tenant by LoyJoy support on request. This is due to security by default, as most tenants do not require this API.
- Added LLM provider
Mistral
as another European model provider.
release-2025-08-12
- Enhanced messages screen for AI messages in the knowledge tab, including expandable cards for system messages for a more compact view, chat bubbles for user and assistant messages etc.
- Web pages in knowledge sources now also like web sites can have settings for CSS selectors etc.
AI Agent
process module- Now can configure tools.
- In case OpenAI or Anthropic ist selected as the LLM model the tool
web_search
is available. - Added tool calling for Mistral and Anthropic additionally to OpenAI and Azure.
- Process modules
GPT gateway
does not offer configuration of system and user message any more for stability reasons. - Added
GPT-5
model family (GPT-5, GPT-5 Mini, GPT-5 Nano) for OpenAI - An empty list of sources screen now shows an empty message.
- Fix for thumbs up/down at AI messages
release-2025-08-05
- In
GPT Knowledge
process module:- Removed text splitting for multiple topics in one user request, as this now can be addressed by agentic AI.
- Removed reranking for performance reasons, as this optimization is not necessary anymore due to stronger AI models.
- Raised the default number of chunks from 16 to 20.
- AI knowledge messages screen
- Tools list and tool call parameters are now visible in the AI messages screen.
- In the RAG query now the query translations are visible, when multiple query languages have been selected in the
GPT Knowledge
process module.
- Prepared the AI completions stack for web search capabilities like ChatGPT. This currently is only available with OpenAI, not with Azure.
- Function
knowledge
: Allow specifying one folder to include in the chunk search. E.g.${knowledge(user_answer, "FOLDER_NAME")}}
.
release-2025-08-04
- Process module
GPT Prompt
now allows to include the chat history, so that the prompt can include the chat history into the analysis to produce an analysis result. This is preferable over printing${chatHistory()}
into the system message or user message, as${chatHistory()}
does not include the latest chat message of the customer, which triggered the processing of theGPT Prompt
process module. Printing${chatHistory()}
into the system message or user message in theGPT Prompt
process module should be avoided. - In process modules
GPT Knowledge
andGPT Smalltalk
the prompt and system message have been unified into only the system message. The user message is now solely for the actual request of the customer.
release-2025-08-03
- Removed process module
GPT Knowledge Follow-up
, as theGPT knowledge
module already covers follow-up questions, automatically. - Question rephrasing for the RAG pipeline in
GPT knowledge
is now executed withGPT-4o-Mini
, if possible.
release-2025-08-02
- Optimized performance of knowledge messages list.
release-2025-08-01
- In the list of knowledge messages, added an error text in case an error has occurred in a GPT chat.
- In process module
GPT Knowledge
:- RAG queries now are always rephrased to make sure that relative follow-up sentences always work with the RAG pipeline.
- When checking the current Web page as part of the RAG pipeline, now a max size of tokens is ensured to prevent context window overflow.
release-2025-07-28
- In knowledge messages now all questions and answers of a conversation are shown in a unified view when clicking on AI communication in the top right corner. This allows to see all questions and answers in one place, which is useful for debugging and understanding the AI conversation flow.
- The new
Agentic AI
process module allows to create agentic AI processes, which can be used to build complex AI workflows. Agentic AI processes are based on the Model Context Protocol (MCP) and allow to integrate external data sources, such as databases or APIs, into the AI process. The process module currently is in beta and will grow in the coming days and weeks with a growing set of MCP tools.
release-2025-07-16
- Added edit icon to GPT prompt fields, so that prompts can be modified more easily.
- Added option to set chat height to 100%. This is especially useful when rendering the chat with an avatar (HeyGen, D-ID).
- In
GPT Knowledge
module, added URL filters for Web search. This limits the Web search on certain URLs. - In
GPT Knowledge
module in Sharepoint added additional meta data fields. This enables to include additional information such as the author in the search.
release-2025-07-14
- The variables API now supports filtering by
created_at
field. This allows to retrieve variables created after and before a specific date.
release-2025-07-07
- Added new Azure OpenAI models
azure-gpt-4.1
,azure-gpt-4.1-mini
,azure-gpt-4.1-nano
,azure-o4-mini
release-2025-06-26
- Added new OpenAI models
openai-gpt-4.1
,openai-gpt-o3
andopenai-gpt-o4-mini
. - Model Context Protocol (MCP) can now be used also with O-Series models, such as
gpt-o3
andgpt-o4-mini
. - Translations in LoyJoy Manager from one language to another language are now based on
gpt-4o-mini
instead of Google Translate. This allows to tackle more complex translations like translating from English to German (Siezen).
release-2025-06-24
- Products from the
Product database
can be used as answer options in theQuestionnaire
module. This allows to create questionnaires with product options, e.g. for product selection or product feedback. - Added native support for D-ID and HeyGen avatars in the LoyJoy chat widget. This allows to use animated avatars in the chat UI, which can be customized with different styles and animations.
- The
GPT Gateway
module now reports Azure content filtering information instead of returning a timeout error. - The conversations picker is now hidden by default in the chat header and Home View. It can be enabled the the branding header settings and in the home view as a widget.
release-2025-06-21
- Home View items (Roster and Groups) can open articles now. This is useful if you want quick accesss to imprint or data privacy policies
- The role system has been extended to allow for more than 10 roles. This allows for more fine-grained access control. Also the runtime API now supports these roles.
- Roles can be added and removed in the manager. This allows for more flexible role management on existing users.
- The
GPT Knowledge
module now supports the Model Context Protocol (MCP) to provide context to the LLM. This allows for integration of external data sources, such as databases or APIs, into the LLM's context. MCP is a standardized way to provide context to LLMs and is supported by many LLM providers. - Update CleverReach integration now allows sending forms and adding inactive subscribers (for double opt-in).
release-2025-06-13
- Persistent quick replies can open articles. This is useful if you want quick accesss to imprint or data privacy policies
- Missing ALT texts for images are now reported by a model checking rule. This helps to ensure that all images have alt texts for accessibility purposes.
release-2025-05-14
- Live chat: The agent is now able to upload and send files
- Users can now have multiple roles
release-2025-05-08
- Accessibility: The LoyJoy chat is now barrier-free and screen reader compatible. This fulfills the requirements of the Barrier-Free Reinforcement Act and sends an important signal for inclusion.
- Automated AI testing: Upload questions as .xslx or .csv and the answers will be generated and made avaibale for download
- It is now possible to change the font color of the header, primary buttons and secondary buttons
- Form module: The height of the input field are now determined automatically
API-Client
module: Ability to change header name and prefix in OAuth2 mode- Fixed language labels in language selection module
- Avatars based on D-ID and HeyGen can now be rendered in the LoyJoy chat background.
- Process module
Message
now support to configure an AI label in case the AI is used to generate the message. - The chat UI now checks for URL query parameters
loyjoy-process
andloyjoy-view
to start a specific process or view. This is helpful for websites that do not allow to change the process ID in LoyJoy widgets on their web pages (e.g. Volksbanken) and allows to integrate different agents on such websites. - The z-index setting of the Chat UI was evaluated also in LoyJoy manager and on LoyJoy sites. This has been removed, as it is not needed there. The z-index setting is only applied to the LoyJoy chat widget on customer websites.
release-2025-04-20
- You can now generate Alt texts for images using LLMs
release-2025-04-19
- Knowledge Home UI in manager was adjusted
- Show start and enddate of giveaway in Data tab
- Intelligent insights uses percentage
release-2025-04-16
- The sorting mode of agents in the manager is now stored across sessions
- New Events in
Email
module:EMAIL_EXISTS
,EMAIL_ASKED
,EMAIL_NOT_ASKED
- Products: Added delete all button to quickly wipe the product database
- GPT recommender module is now automatically connected with the LoyJoy product database
release-2025-03-10
- Show product galleries in live chats
- Add option to upload products as CSV or XLSX to the LoyJoy product database
- Add module "GPT Recommender": Enables to rapidly build smart product recommenders based on your product data
- Clicking on a quick reply now consistently leads to a postback.
- It is now possible to use emojis e.g. in a Decision Gateway.
release-2025-02-20
numParticipations()
andnumParticipationsBpmnProcess()
now return the number of participations of the logged in user.- GPT answers on documents attached to the chat UI now use more chunks. This enables questions like
summarize this document
. - Preparations for hosting in other data centers than Google Cloud.
release-2025-02-11
- GPT Knowledge now supports the inclusion of the current web page on which the user is currently located as context.
- GPT Knowledge is now able to rephrase followup questions of the user.
- GPT Knowledge is now able to decompose the user question into multiple subquestions to enhance RAG context search.
release-2025-01-31
- Added new functions
- stringEndsWith
- stringStartsWith
- stringLength
- stringIndexOf
- stringLastIndexOf
- stringSubstring
- stringStrip
release-2025-01-23
- TimerEvents: Allows interpolation of ISO8601 with functions/variables.
- Templates that contain timers do not need to be updated every year.
- GPT citations with Cohere
- Adds inline references with square brackets e.g. [1] to AI generated answers when
Cohere Command-R
is selected as GPT provider
- Adds inline references with square brackets e.g. [1] to AI generated answers when
release-2025-01-22
- Modul paste button added to each gateways, loops and more
- Added "Add to cart" button in Product feed module
- The Markdown representation is now removed before embedding a chunk
release-2025-01-21
- Copy tokens can now contain products and sources. This allows templates to be copied with products and sources, e.g. for a product feed template or a GPT Knowledge template.
release-2025-01-20
- Skip values in the
Questionnaire
module are now empty by default - Added a model checking rule for gateways without branches
release-2025-01-19
- StartEvents can now be interrupting: After entering a question into the footer, and receiving an answer, the chat might not jump back to the last process step, but instead wait for the next user question.
- You can now select the GPT model on each module. This allows to select e.g.
GPT-4o
for a specific module due to AI strength requirements, andGPT-4o-Mini
for all other modules in tenant settings. - Added new function
displayLanguage()
, which returns the language of the chat, such asEnglish
oderGerman (Siezen)
. - All GPT modules no longer implicitly inject the language of the user into the prompt. Instead the language has to be set explicitly in the prompt with help of the new function
displayLanguage()
- Added filter by agent for consents in data view.
- Fixed tooltips in process editor
release-2025-01-13
- Added web search to process module
GPT Knowledge
. This allows to search the web for answers to a question, additionally to the internal LoyJoy knowledge base. The web search is based on multiple search engines and can be configured in the module settings. - When reranking is enabled in process module
GPT Knowledge
, now the 1/3 of the sources are chosen always, and 2/3 are chosen based on reranking. This makes the source selection more robust. - Deprecated GPT-3.5, as GPT-4o and GPT-4o-Mini are better alternatives.
release-2025-01-03
- Website crawling now converts HTML to Markdown instead of plain text. This allows to preserve the formatting of the website and links.
- In process editor it is now possible to dragscroll with the middle mouse button
- Mappings now are based on JSON pointers, XPath and the LoyJoy expressions engine. This allows arbitrarily complex mappings to read and write API data. An example for a complex mapping would be to collect all object keys or values from JSON.
- Process module
GPT Knowledge
now allows to configure the number of chunks to be used for the AI answer (max. 32 chunks). Also chunks ordering is optimized before sending it to GPT to ensure that related chunks are joined together. This allows to get more relevant answers from the AI. - Chat widget:
- The chat footer now adapts to the text length by adding rows and a scrollbar if the text is too long.
- The chat shows only the 4 best sources for a GPT answer, however now allows to click on "Show all sources" to see all sources.
- Added params
message
andsuggestions
to the widget API, which allows to set the message and suggestions of a chat widget. An example for themessage
param is the search on loyjoy.com, which upon a search query triggers the chat with the search query as the first chat message. - The max length of the chat footer text can now be configured in branding. This allows to enter longer texts in the footer for internal use cases.
- Comboboxes now prevent entering incorrect values
- Border colors of inputs in the chatUI are now in brand color
- Added products screen to LoyJoy manager, which allows to add and edit products.
- Added functions
chatHistoryLive
to get the live chat history, e.g. to send it via email to a CRM or ticket system after live chat has ended.locationHref
to retrieve the current complete URL of the widget.nl2br
to convert newlines to<br>
tags, e.g. to send text with newlines via an HTML email.