Server Enhancements

AI Support

• LLMs now support default system prompts. When provided, these will be placed before any other messages or conversations. See here for more details.
• Semantic Index content ingestion now supports the ability to specify a GenAI Flow procedure to process the content as the default ingest mechanism. See Content Ingestion Flows for more information. Further, note that complex content ingestion is no longer limited by procedure execution time-outs when performed via the Semantic Index Service.
• The Chat Message builders now support image_url prompts containing ResourceReference values. These are used by “multi-modal” LLMs.

Sources

• Added VIDEO source type. This source can fetch images from online cameras, returning them for processing in the Vantiq system. See Video Source Integration for details.

Other

• Nodes now pay attention to the URI’s path when targeting the local server. This may cause issues if you have nodes that have extraneous paths, e.g. https://dev.vantiq.com/ui/ide/index.html instead of https://dev.vantiq.com. Using the URI of the “self” node will fix any issues this causes.
• New Organizations can now be automatically provisioned with resources. This is only relevant on installations where you have system admin permissions. See here for more details.

UI Enhancements

General

• AI Client Assistant is a new feature added to Client Builder for generating client pages using an LLM. See the Client Builder User’s Guide for more information.
• LLM playground now supports exporting conversation to a JSON file and loading conversation from an exported file.

GenAI Builder

• The GenAI Builder now enforces strict typing between connected tasks. For each edge in the GenAI Flow, it must be possible to assign the output type of a parent task to the input type of the child(ren). See GenAI Task Typing for more details.
• The AI Primitive components Loop, Map, Procedure, and Transform have been added.
• The AI Pattern component ReduceDocuments was added.
• GenAI Flows may now access per-flow memory as part of their implementation.
• The LLM component can now specify a schema type when expecting JSON output.
• Support for retrieval with contextual compression (which includes “re-ranking”) has been added via the new resource component SemanticIndexWithCompression and several document compressors.