2. NodeValidator (Optional)

Step 2 (Optional) - Validates and corrects node names. Uses semantic search and optionally an LLM to fix invalid node references. Recommended to ensure workflow generation succeeds if Step 1 produces errors.

By DanielPFlorian·Created 9 months ago·Updated 9 months ago· 34

workflow_edges (refined)
llm_prompts (debug)
candidate_nodes (debug)

◄instruction►

◄workflow_edges—►

◄use_llm_refinementfalse►

◄refine_model_pathQwen2.5-7B-Instruct-q8_0.gguf►

◄embedding_model_pathparaphrase-multilingual-MiniLM-L12-v2►

◄catalog_directorycatalog►

◄dtypeauto►

◄device_preferenceauto►

◄attn_implementationauto►

◄auto_gpu_layerstrue►

◄n_gpu_layers-1►

◄max_new_tokens4096►

◄context_size4096►

◄use_mmaptrue►

◄use_mlockfalse►

◄n_batch512►

◄n_threadsAuto►

◄temperature0.95►

◄top_k5►

◄top_p0.70►

◄seed0►

CategoryWorkflowGenerator

Inputs (21)

Name	Type	Default	Description
instruction	STRING		Optional context for LLM refinement. Connect 'instruction' output from Step 1 or add your own instructions to guide refinement.
workflow_edges	STRING		Input diagram from Step 1 (WorkflowGenerator).
use_llm_refinement	BOOLEAN	false	Enable LLM-based correction for better accuracy (slower). False uses only semantic search.
refine_model_path	COMBO	Qwen2.5-7B-Instruct-q8_0.gguf	LLM model for refinement (if enabled).
embedding_model_path	COMBO	paraphrase-multilingual-MiniLM-L12-v2	Embedding model for semantic search.
catalog_directory	STRING	catalog	Directory containing node catalog files.
dtype	COMBO	auto	Data type for refinement model (HuggingFace).
device_preference	COMBO	auto	Device preference for refinement model.
attn_implementation	COMBO	auto	Attention implementation (HuggingFace).
auto_gpu_layers	BOOLEAN	true	Auto-calculate GPU layers based on available VRAM (prevents OOM errors) (GGUF).
n_gpu_layers	INT	-1-1–1000	Number of GPU layers (GGUF). -1 for all.
max_new_tokens	INT	40961–16384	Max tokens for refinement.
context_size	INT	4096512–32768	Context window size (GGUF).
use_mmap	BOOLEAN	true	Use memory-mapped loading (GGUF).
use_mlock	BOOLEAN	false	Lock memory to prevent swapping (GGUF).
n_batch	INT	51232–2048	Batch size for prompt processing (GGUF).
n_threads	COMBO	Auto	CPU threads for inference (GGUF).
temperature	FLOAT	0.950–2	Sampling temperature.
top_k	INT	51–20	Number of similar nodes to consider.
top_p	FLOAT	0.700–1	Top-p sampling.
seed	INT	00–4294967295	Random seed.

Outputs (3)

Name	Type	Description
workflow_edges (refined)	STRING	Refined workflow diagram with corrected names.
llm_prompts (debug)	STRING	Prompts sent to LLM (debug).
candidate_nodes (debug)	STRING	Candidate nodes found (debug).