Manual editing
Add blocks, drag models, connect port dots, pan blank canvas, preview flow, and save/load JSON.
Editor
The canvas is the workflow.
The builder combines a block sidebar, auto-growing canvas, and a right sidebar with Execution and AI Builder tabs. Sidebars resize, scale text, and retract into reopen rails instead of disappearing.
Example presets
Packaged examples like quick answer, triage router, quality gate, draft review, and research synthesis.
Safe graph state
IDs normalize on load/generation, far blocks expand the canvas, duplicate edges are ignored, and direct model-to-model links are blocked.
AI Builder
Describe the pipeline. NativeLab validates the JSON.
The AI Builder tab sends a compact schema guide to the active model, checks context before the call, extracts JSON, retries once if needed, fills active-model placeholders, validates, saves, and lets you test.
Prompt workflow
Load a model
Any active local, Ollama, HF, or API model can be used.
Describe the graph
Name the output JSON and ask for the workflow you need.
Load and test
Saved pipelines load back onto the canvas through normal persistence.
Smart context
- Empty canvas prompts go as-is
- Existing canvas is attached as JSON
- History is stored under localllm
- /get_data prints current canvas state
- /context compacts builder history
- Context overflow is blocked before sending
Blocks
20+ blocks across generation, context, and logic.
Use deterministic blocks where possible and LLM logic where language judgment is useful.
I/O
Input, Output, Model, Intermediate with live streaming output tabs.
required path
Context
Reference, Knowledge, and PDF Summary blocks inject reusable material.
documents
Deterministic logic
IF/ELSE, SWITCH, FILTER, TRANSFORM, MERGE, SPLIT, and safe Custom Code.
fast
LLM logic
LLM-IF, LLM-SWITCH, LLM-FILTER, LLM-TRANSFORM, and LLM-SCORE.
agentic
Execution
Same guardrails everywhere.
Manual, generated, loaded, and CLI-run pipelines share validation, graph normalization, deterministic helpers, and centralized LLM error handling.
Validation layer
Required blocks, valid endpoints, model refs, context/PDF metadata, LLM instructions, and custom code checks.
Native graph core
Optional C helpers accelerate ID normalization, loop checks, route selection, transforms, merges, and validation records.
Error dialogs
Context-window and engine errors surface as user-facing dialog text, not just logs.