Edge-model research
Reason broadly, extract narrowly
A reasoner chooses a tool; a compact schema-constrained model extracts only that tool’s arguments.The split pipeline
1. Trigger
The reasoner emits `<TOOL:name>` without arguments.
2. Encode
Frozen GIST-small embeds the tool name and conversation.
3. Project
A learned projector creates soft-prefix embeddings.
4. Decode
SmolLM2-360M emits JSON under schema masking.
5. Execute
`@neosyntropy` validates, invokes, and logs the tool.
Why separate extraction?
The reasoner remains responsible for conversational judgment. The smaller decoder receives a narrow task with a machine-checkable output space.
Guidance or llguidance masks invalid tokens during generation, while Pydantic validates the completed arguments before invocation.
The local tool decorator
Pipeline details come from slm-ntk/README.md and its runtime, decoder, and guided-json modules.