Typed Decision Node

Use Typed Decision to choose a label, score text against a rubric, or estimate whether a statement holds. Model selects the weights while the question inputs and result shape stay the same. Existing Laya nodes keep Laya as their default model.

Choose model weights

This node runs ONNX models. The dropdown includes these models:

Model	First use
mizchi/laya-multilingual-onnx	Downloads missing files, about 681 MB for the complete bundle. Default for existing flows.
fastino/GLiNER2.5-Decide	Downloads a pinned community ONNX conversion, about 1.76 GB.
fastino/GLiNER2.5-multi-Decide	Downloads a verified ONNX export from the Flow-Like CDN, about 1.13 GB.
fastino/GLiNER2.5-Decide-1B	Downloads a verified ONNX export from the Flow-Like CDN, about 4.14 GB.
fastino/gliner2.5-multi-v1	Downloads a pinned community ONNX conversion, about 1.13 GB.
fastino/gliner2.5-base-v1	Downloads a pinned community ONNX conversion, about 749 MB.
fastino/gliner2.5-small-v1	Downloads a pinned community ONNX conversion, about 293 MB.
custom	Loads a complete bundle that you provide. Missing files produce an error.

The Decide preset uses the nishparadox/gliner2.5-decide-onnx float32 conversion. The small, base, and multilingual v1 presets use the encoder and classifier from codesoda/gliner2-onnx. Multi-Decide and Decide-1B use Flow-Like ONNX exports from the CDN. Download URLs use fixed revisions, and downloaded files are checked against SHA-256 hashes.

Connect Model Directory to a FlowPath, a reference to a directory in your flow’s storage. Downloadable GLiNER models use a separate .decision-cache/<model-name> directory for each preset. Existing cached files are reused. Laya keeps its .laya-cache layout and also recognizes its older cache layout. Loaded model sessions are reused within an execution cache.

Ask a question

1. Set Text to the content or serialized JSON state to evaluate.
2. Write the question in Instructions.
3. Select Question Type and fill in its inputs:

Question type	Inputs	Answer
choice	Choices: unique labels, such as billing and technical support.	Choice: the selected label.
score	Levels: rubric descriptions in order, starting at level zero.	Score: the expected rubric level, which can be fractional.
noul	Optional False Description and True Description.	P(True): the probability that the statement holds.

The result also contains option probabilities, confidence, and the number of input tokens. act_probability contains Laya’s separate action-head probability and is null for GLiNER. Confidence summarizes the option distribution; it does not measure real-world accuracy.

Export weights

All presets download automatically. To export a different checkpoint or your own fine-tuned weights, use the export tool from the Flow-Like repository. With uv installed, run this from the repository root:

uv run --python 3.12 tools/export-decision-model.py \
  --model fastino/GLiNER2.5-multi-Decide \
  --output ./decision-model

Use any supported GLiNER2 checkpoint or local checkpoint directory as --model. The output directory must not already exist. The tool installs its Python dependencies, exports on CPU in float32, and checks the exported result before publishing the bundle. Exporting needs several times the checkpoint size in memory and disk space; the 1B model needs multiple gigabytes.

Place the resulting directory in storage accessible to the flow and connect its FlowPath to Model Directory. Choose custom in Model. A local bundle can also override a preset when its source_model matches that preset.

Use custom weights

Select custom and point Model Directory at your complete bundle. The node detects the format from its configuration and performs no model downloads.

• A GLiNER classification export contains decision_config.json, model.onnx, tokenizer.json, and model.onnx_data when the export uses external tensor data. The export tool also accepts a local GLiNER2 checkpoint directory as --model.
• A Laya bundle contains model.onnx, rl_agent_config.json, and either tokenizer/tokenizer.json or tokenizer.json.

For compatibility, saved flows retain the internal node name onnx_laya and the FlowScript name onnx::laya.

T

AI/ML/ONNX/NLP

Answer a choice, rubric score, or true-probability question about text with Laya or GLiNER2.5. Select ONNX model weights and connect a Model Directory to cache downloads and reuse them across runs. Select custom to load your own complete bundle without downloading missing files.

onnx_layaonnx

Inputs7

Outputs4

Security exposureUnrated

Packageonnx

Ratings

Scores range from 0 to 10. Higher values mean more impact, exposure, or operational weight.

No score metadata has been set for this node yet.

Input Pins

7

Input

Execution

exec_in

Initiate execution

Model Directory

Struct

model_dir

FlowPath directory for downloaded ONNX models and reusable cache. Place your exported weights here for custom. Custom requires a complete Laya or GLiNER2 classification bundle and never downloads missing files

SchemaFlowPathFlowPathFlowPath3 fields

pathstringrequired

store_refstringrequired

cache_store_refstring | null

Schema enforced

Model

String

model

Model weights to use. ONNX presets download once and reuse the Model Directory cache. Custom loads your own complete bundle

Default mizchi/laya-multilingual-onnx

mizchi/laya-multilingual-onnxfastino/GLiNER2.5-Decidefastino/GLiNER2.5-multi-Decidefastino/GLiNER2.5-Decide-1Bfastino/gliner2.5-multi-v1fastino/gliner2.5-base-v1+2 more

Text

String

text

Text or serialized JSON state to evaluate

Instructions

String

instructions

Question to answer about Text

Question Type

String

question_type

choice selects a label; score returns an expected rubric level; noul returns P(true)

Default choice

choicescorenoul

Choices

String Array

criteria

Unique labels to choose from

Default []

Output Pins

4

Output

Execution

exec_out

Decision complete

Result

Struct

result

Typed answer, option probabilities, confidence and token count. act_probability is Laya's action probability and null for models without an action head

SchemaLayaResultLayaResultLayaResult8 fields

question_typeLayaQuestionTyperequired

enum "choice", "score", "noul"

choicestring | null

Selected label for a choice question.

scorenumber | null

Expected zero-based rubric level for a score question.

format double

noulnumber | null

Probability that the statement holds for a noul question.

format double

probabilitiesArray<LayaProbability>required

itemsLayaProbabilityarray item

labelstringrequired

probabilitynumber:doublerequired

format double

confidencenumber:doublerequired

One minus normalized entropy; for noul, the larger of P(false) and P(true).

format double

act_probabilitynumber | null

Laya's probability of action index zero; absent for models without an action head.

format double

input_tokensinteger:uintrequired

format uintmin 0

Choice

String

choice

Selected choice label

Confidence

Float

confidence

Normalized entropy confidence, or max(P(false), P(true)) for noul

Node Info

Internal name

onnx_laya

Category

AI/ML/ONNX/NLP

Version

3