release-connector-3.0.0

The OpenAI connector now reuses many generic entities and operations from the GenAI Commons module. This makes it easier to build vendor-agnostic applications and swapping models and integrates well with the newly released Conversational UI module. Updating will cause errors in existing chat completions implementations. To mitigate those, you’d need to change the input parameters of the operations and the post-processing of the response. Examples of how this can be done can be found in the OpenAI Showcase App. Lastly, the complex “Chat Completions (advanced)” operation was removed and can be replaced with the regular “Chat Completions (with history)” operation.

release-connector-2.7.1

Added userlibs that were missing in release 2.7.0.

All chat completions operations now support vision which enables models like GPT-4 Turbo to interpret and analyze images, allowing them to answer questions and perform tasks related to visual content. For Chat Completions without History the ImageCollection is an optional input parameter, while for Chat Completions with History the ImageCollection can optionally be added to individual user messages in ChatCompletionsSession_AddMessage.

Additionally, the chat completions operations have a new optional input parameter called “MaxTokens” to control the maximum number of tokens to generate in the chat completion.

release-showcase-2.7.0

The Showcase app has been updated to include the OpenAI Connector v2.7.0 and the PgVector KnowledgeBase v1.2.0.

We added a data manipulation page to the semantic search ticket example. This illustrates how changes to Mendix objects influence their corresponding records in a knowledge base. The underlying microflows makes use of operations from the PgVector Knowledge Base.

Furthermore, we created a new example demonstration chat completions with vision where users can either take a picture with their webcam or upload an image from their drive and ask questions about it.

Lastly, we removed outdated OpenAI models from the models enumeration.

release-pgvectorknowledgebase-1.2.0

We added a new operation that replaces existing chunks in the knowledge base for chunks based on a Mendix Object. This allows to keep the keep the knowledge base in sync with the application database. Similar to the Repopulate Knowledge Base operation, you can use Chunk_Create to create the input for the Replace operation.
Additionally, we added an operation that inserts a list of Chunks into an existing knowledge base.
Lastly, two new operations have been added that can be used to delete chunks and their corresponding labels from the knowledge base by either providing a Mendix object or a list of Mendix objects.

release-connector-2.7.0

All chat completions operations now support vision which enables models like GPT-4 Turbo to interpret and analyze images, allowing them to answer questions and perform tasks related to visual content. For Chat Completions without History the ImageCollection is an optional input parameter, while for Chat Completions with History the ImageCollection can optionally be added to individual user messages in ChatCompletionsSession_AddMessage.

Additionally, the chat completions operations have a new optional input parameter called “MaxTokens” to control the maximum number of tokens to generate in the chat completion.

release-showcase-2.6.0

The Showcase app has been updated to include the OpenAI Connector v2.6.0 and the PgVector KnowledgeBase v1.1.0.

We've added an example to illustrate the new OpenAI Connector function calling capabilities, which enable the LLM to intelligently decide when to call a predefined function microflow to gather additional information to include in the assistant response. In this example, users can ask questions about tickets in the application database.

Additionally, we have replaced parts of the implementations of the two similarity search examples by the new operations from the PgVector Knowledge Base. The association between the chunk objects and the Mendix objects for which the chunks were created is now set automatically when the chunks are retrieved.

release-pgvectorknowledgebase-1.1.0

We added two new operation to retrieve chunks from the knowledge base and automatically associate them to the Mendix object for which the chunks were created (if applicable). This can be done as a nearest neighbor search or as a regular retrieve with offset and limit. Additionally, we added an operation to delete a knowledge base.

release-connector-2.6.0

All chat completions operations now support function calling, which enables the LLM to intelligently decide when to call a predefined function microflow to gather additional information to include in the assistant response. Currently, function microflows are limited to one input parameter of type string and must return a string.

Additionally, chat completions operations have a new optional input parameter called “Temperature” to control the randomness of the response.

After updating to this version, existing implementations with the chat completion operations will have two new optional input parameters “FunctionCollection” and “Temperature” that can be set to empty if not needed.

Lastly, we expose helper micoflows to create inputs for the connector operations in the toolbox in Studio Pro so that they can be dragged and dropped directly from there.

release-showcase-2.5.0

We have replaced all vector database interaction logic regarding storing and retrieval of knowledge base chunks by operations of the new PgVectorKnowledgeBase module. This means that no custom query coding is required anymore by developers. As a result the example queries have been removed from the pages. Also it is possible now to store connection details for multiple vector databases in the same Mendix app.

release-pgvectorknowledgebase-1.0.0

Initial release. We included the necessary operations to populate knowledge bases and to retrieve data (with and without vector similarity calculation).