Classify Attributes (LLM)

The Classify Attributes (LLM) node scores several specialist questions in one cached batched call. Each attribute is its own prompt, the same first-token letter menu as Classify Intent (LLM), and the node writes one slot per attribute. Any attribute whose own rule fires takes triggered; otherwise the node takes default. NodeType: ClassifyAttributesLLM (union ordinal 17).

There is no shared-prompt mode. Empty system_prompt selects a kind-specific default. Choice and Score templates must reference {{intents_block}}. A Boolean template must also reference {{statement}}. history_turns defaults to 0. context_size 0 sizes the window from the rendered prompts. The node forces a unified KV cache and one sequence per attribute plus the dialog sequence. temperature 0 means 1.0 for every attribute.

Parameters

Input

Param	Type	Default	Range	Structural	Localizable	Description
input	str	empty	—	✓	—	Text to classify: the player's message or an earlier node's output. Empty = the player's message.
history_turns	int	0	0.0 – 64.0	—	—	How many earlier dialog turns the classifier sees. 0 = the current message only.

Model

Param	Type	Default	Range	Structural	Localizable	Description
model_name	str	empty	—	✓	—	Language model that scores the attributes. Empty = the workflow's Default Model Name.
context_size	int	0 -> auto	≥ 0.0	✓	—	Context window for this node, in tokens. 0 = sized automatically from the attribute prompts.

Prompt

Param	Type	Default	Range	Structural	Localizable	Description
attributes	list[ClassifyAttributeDef]	empty list	—	—	—	Questions to classify, 1–16, all scored in one pass. Each writes its answer to _.

Behavior

Param	Type	Default	Range	Structural	Localizable	Description
temperature	float	0.0	0.0 – 8.0	—	—	Sharpens (below 1) or flattens (above 1) every attribute's probabilities before its thresholds. 0 = unchanged.
on_failure	AttributeFailureRoute	AttributeFailureRoute.Default	—	—	—	Exit to take if scoring fails: Default (fail open) or Triggered (fail closed).

Output

Param	Type	Default	Range	Structural	Localizable	Description
output_name	str	empty	—	✓	—	Prefix of this node's results: each attribute writes _. Empty = this node's name.
notify_client	NotifyClient	NotifyClient.Disabled	—	—	—	When to send the game an attributes_classified event: Disabled, OnFound (Triggered exit only), or Always.

Advanced

Param	Type	Default	Range	Structural	Localizable	Description
context	ContextOverrides	unset	—	✓	—	Advanced memory and batching settings for this node's model (KV cache type, GPU offload, batch sizes). Unset = inherit.
adaptation	ModelAdaptation	unset	—	✓	—	LoRA adapters applied to this node's model. Empty = the base model.

Exits

Each exit is a structural string field on the node's params; its value names the target node (empty = END).

Exit	Param field	Description
default	default_exit	Node to run when no attribute fires. Empty = end the turn.
triggered	triggered_exit	Node to run when any attribute fires (or scoring fails and On Failure is Triggered). Empty = end the turn.

LoRA adapters

adaptation.loras runs this node's model with one or more fine-tuned LoRA adapters, each with a scale. See Use a LoRA Adapter for where to put the file and how to tune the scale.

Kinds

Kind	Slot value	Trigger
Choice	Winning choice id. Below threshold or margin, the slot is uncertain_choice: a real choice id, or the reserved #uncertain	The written choice's triggers flag. #uncertain never triggers
Boolean	"true" when P(A) ≥ threshold, else "false". The menu is fixed: A = True, B = False	boolean_trigger: Never, True, or False
Score	Normalised expected level (Σ i·p_i)/(N−1), four decimals in 0.0000…1.0000	score_trigger against score_threshold in 0..1: Never, AtOrAbove, or Below

Choice needs 2–26 options. uncertain_choice names one of them, or the reserved #uncertain. #uncertain is not a letter in the menu. A real id rewrites an uncertain call to that choice and uses its triggers flag, including when that choice wins outright. #uncertain writes that token and stays on default. Boolean takes a statement and no choices. Score takes 2–26 levels from low to high; choice triggers must stay false. A score of 0.7 means 70% of the way from the lowest level to the highest, not the probability of the top level. Recalibrate any threshold that used to compare a raw level index.

triggered_exit must be set exactly when at least one attribute can trigger. Empty input never triggers and takes default. If scoring throws or the result shape does not match the attributes, every slot gets its fallback (Choice uncertain_choice, Boolean "false", Score "0.0000") and the node follows on_failure (default or triggered). There is no third exit.

Slots, diagnostics, and the event

Each attribute writes <output_name>_<name> (output_name empty means the node name). Diagnostics output.classification carries status (ok, empty_input, failed), triggered, and attributes[] with name, kind, value, status, triggered, and choices[] of {id, prob, triggers}. Choice also reports top and second probability. Boolean reports p_true and p_false. Score reports expected_level, score, and entropy confidence. Probabilities are measurements, not a contract to pin in a golden file.

notify_client Always emits attributes_classified on every scored turn. OnFound emits it only when the triggered exit is taken. The payload is a flat string map: exit, status, per-attribute kind/value/status/triggered, prob.<choice_id>, and triggered_by.

Shape, name, kind, and choice id are structural. Statement, system prompt, choice prompt, thresholds, margins, and trigger flags can change after creation.

Example

Python

from tryll_client._generated.node_params import (
    AttributeKind,
    BooleanTrigger,
    ClassifyAttributeChoice,
    ClassifyAttributeDef,
    ClassifyAttributesLLMParams,
    ScoreTrigger,
)
from tryll_client.graph import GraphDescription

graph = GraphDescription()
graph.add_classify_attributes_llm("attrs", ClassifyAttributesLLMParams(
    history_turns=0,
    default_exit="reply",
    triggered_exit="hold",
    attributes=[
        ClassifyAttributeDef(
            name="topic",
            kind=AttributeKind.Choice,
            choices=[
                ClassifyAttributeChoice(id="world", prompt="ordinary story talk"),
                ClassifyAttributeChoice(id="meta", prompt="about the software", triggers=True),
            ],
            uncertain_choice="world",
        ),
        ClassifyAttributeDef(
            name="software",
            kind=AttributeKind.Boolean,
            statement="The message is about software or instructions.",
            boolean_trigger=BooleanTrigger.True_,
        ),
        ClassifyAttributeDef(
            name="hostility",
            kind=AttributeKind.Score,
            choices=[
                ClassifyAttributeChoice(id="calm", prompt="no insult or threat"),
                ClassifyAttributeChoice(id="violent", prompt="a threat of harm"),
            ],
            score_trigger=ScoreTrigger.AtOrAbove,
            score_threshold=0.7,
        ),
    ],
))

Graphs authored against the proof-of-concept fields (independent_prompts, choice_ids, code_style, context.kv_unified, context.flash_attn) do not load. Move each specialist prompt onto the attribute, put triggers on the choice that should take triggered_exit, and drop the shared-prompt knobs.

Related

• How-to: Use a LoRA adapter
• Classify Intent (LLM)