release-showcase-2.4.0

We have enhanced the semantic search example by suggesting search terms to the user and adding tickets in different languages to the dataset. This shows that OpenAI's LLMs are capable of handling non-English language tasks.

As of this release, we support the text-embedding-3-large model in all embeddings examples. This model is particularly suitable for tasks involving multiple languages.

In addition, selected configurations are now saved in the database for each example and operation, in order to ensure consistency and reduce repeated setup effort.

release-showcase-2.3.0

We created an additional example where we show semantic search through a dataset of IT helpdesk tickets. This example demonstrates how the OpenAI connector combined with vector database logic can be leveraged to perform semantic search on data records in an application. Two variations are demonstrated, one where a single column is embedded for the search, the other one uses a multi-column approach.

Furthermore, the values of the OpenAI model enumerations have been updated. We added gpt-4-turbo-preview, gpt-4-1106-preview and text-embeddings-3-small and we removed legacy models gpt-3.5-turbo-16k and gpt-4-32k.

Lastly, the database connector module has been updated to the newest released version.

release-showcase-2.2.0

We created two additional examples demonstrating new use cases of vector embeddings. The first example shows how you can identify and visualize clusters within a textual dataset. The second use case implements a semantic search based on similarity of vector embeddings.

Lastly, we made small improvements to the existing examples.

release-connector-2.2.0

We made several small improvements based on community feedback. We removed all ‘Show message’ activities from our exposed operations and replaced them with log messages. Furthermore, we added a Boolean return value called 'Success' to the Embeddings (list input) operation. Lastly, we shortened the names of the exposed operations so they are better displayed in the Toolbox in Studio Pro.

release-showcase-2.1.0

We made a number of small improvements to clarify the working of the underlying interactions with the large language models. In the example implementations “Product description” and “Complexity reduction”, the user prompt and system prompt are now shown in the UI before the model is invoked. These prompts were also improved to protect against prompt injections.
Furthermore, we have extended the chat implementation with optional JSON mode. It is now possible to instruct the model to respond with JSON in the chatbot-like setting, also taking into account the chat history.

release-showcase-2.0.0

We added two example implementations based on the Embeddings API. In a smaller technical example, the Embeddings API is invoked for a list of String inputs and their embedding vectors are retrieved and shown in the UI.
In a second, more complex example, an implementation of retrieval augmented generation (RAG) is demonstrated. For this implementation use case to work, you will need to provide an external PostgreSQL database that has the ‘pgvector’ extension installed. You’ll find detailed instructions on this in the UI of the running app.

release-connector-2.0.0

We have included three new operations that can be used to invoke the Embeddings API and create vector embeddings for a single String or a list of Strings. Furthermore, we included additional microflows in the Advanced folder of all operations that you can use to create request objects.

For Azure OpenAI, we now support authorization with an api-key, besides the existing Microsoft Entra token. Please note that it depends on your organization's security settings, which authorization methods are allowed for your account.

Breaking changes:

We removed the default models per operation from the Configuration entity and changed the Model input parameters in all exposed microflows from Enumeration to String to give developers more flexibility in choosing a model without depending on a fixed list of values. Furthermore, we introduced two new entities: AbstractUsage and ConfigurationTest. ChatCompletionsUsage is now a specialization of AbstractUsage. ConfigurationTest was introduced so we could remove entity access from ChatCompletionsSession and ChatCompletionsSessionMessage; it is used in the user flow of testing a newly set up configuration with a simple chat completions call.

release-showcase-1.3.0

We have added two showcase implementations for JSON mode of chat completions: one to generate JSON strings based on user input prompts, plus an example implementation that generates small batches of demo data (Car records) to be stored in the app database.

Furthermore, the side menu is now compatible with Mx10 and we did some general styling improvements.

release-connector-1.3.0

We have included JSON mode in the chat completions operations. This mode forces the compatible language models to always return valid JSON as the response.
For this, we have extended the non-advanced operations for chat completions with the responseFormat input parameter. For existing implementations that rely on a text response, this can be set to Enumeration value "text" or left empty (to let the system itself assume the default). For use cases where always valid JSON is required as a response, the Enumeration value must be set to "json_object".

For image generations operations, developers can now set the ResponseFormat (url vs. base64) field. This value determines how the retrieval of the image will happen in the implementation. The new field is optional, for existing usages the value can be set to URL or left empty and let the API assume the default value.

Lastly, we have improved the user experience when creating configurations.

release-showcase-1.2.1

New Years Cleaning: We removed unused Marketplace modules and cleaned up the jar files in the userlib folder so that the Showcase app is now out of the box compatible with Mendix Studio Pro 10.6.0.