New Nodes

LLM Control Nodes

LLM nodes controlling the flow execution.

Tested local models:

• Gemma 3 - 4b
• Gemma 3 - 12b
• Qwen2.5 - Coder 7B
• Llama 3.2: Dark-Champion
• Llama-3.1-8B
• DeepSeek R1: Llama 8B
• DeepSeek-R1-Qwen3-8B

LLM Branch Node

• Like the Control > Branch Node but with integrated decision making about the input being either True or False
• Useful to route a execution flow between two possible downstream branches.
• Node execution might fail if the LLM-output cannot be parsed according to the decision data schema.

With Structured Output Node

• Let LLMs generate structs as outputs. Useful for API-calls or anything else that requires deterministic information formatting (e.g. dispatch events)
• Effectively, this is a forced single-tool call configuration.
• Node execution can fail if the LLM produces an output that cannot be parsed as JSON or if the JSON produced violates the specified schema.
• If node execution succeeds, however, the output is guaranteed to be valid JSON data that aligns with the specified schema.

Make Function Schema Node

• Function call definitions or JSON Schemas are tedious to write by hand so this is an utility node to help you out.
• Node execution can fail if the LLM produces an output that cannot be parsed as JSON schema.
• If node execution succeeds, however, the output is guaranteed to be a valid OpenAI-like Function Call Definition with valid JSON schema in the "parameters" section.

Additional Utility Nodes

Parse With Schema Node

• Like utils > types > From String Node but with additional schema validation
• Input strings must not only be JSON-serializable but also follow the provided schema
• Schema definitions can either be JSON schemas or OpenAI function definitions.
• Produces detailed error messages in case of violation.
• Additionally, this module bundles JSON- and schema-related utility functions.

Shuffle Node

• Randomly shuffle array elements

Documentation

• Restructure readme + add new images

Fixes

• Fixes an issue that caused model name hints having no effect for model loading due to an early return
• Minor refactoring of bit.rs in core library
• Log name of loaded model in Find Model node for better debugging