Tool Descriptions v2 + Schema Parsing + New Model

[cmac86]

Three updates that work together.

Structured Tool Schemas

CAAL now parses structured schemas from tool descriptions and passes them to the model as proper function definitions — parameter names, types, required fields. Previously, tools were described as prose blocks. This cuts prompt tokens almost in half (8,700 → 4,500) and gives models the structure they need to route correctly.

This matters most on constrained hardware where every token costs time.

Updated Tool Descriptions

All tools in the registry have been updated with v2 descriptions — clearer routing rules, explicit "use when" triggers, and structured schemas. If you've installed tools from the registry, pull the latest versions to get the updated descriptions.

ESPN and Weather tools have now been consolidated into single tool-suites each.

caal-ministral Round 7

ollama pull coreworxlab/caal-ministral:latest

Trained on the v2 descriptions with structured schemas. 92% on our 85-test suite (args, single-turn, multi-turn, conversational). The previous model was trained on prose descriptions — this one matches what CAAL now sends to the model at inference.

Why This Matters

Base Ministral 3B can now call tools correctly with zero fine-tuning just from the structured schemas. That was not possible with the old prose descriptions. Smaller models benefit the most from this — less ambiguity, less prompt to process. They still need fine-tuning - I'll be doing more testing on this soon.

Video

New video dropping this week covering CAAL on low-power hardware. More on that soon.