Enable tool calling support

[AnanyaRaval]

Is your feature request related to a problem? Please describe.

I want to use the models available via vector-inference to build an agent for a research project.

Describe the solution you'd like

Launch vllm with tool calling enabled as described here: https://docs.vllm.ai/en/latest/features/tool_calling/

Describe alternatives you've considered

A clear and concise description of any alternative solutions or features you've considered.

Additional context

Add any other context or screenshots about the feature request here.

[Keesan12]

If you wire this through vLLM, one thing worth deciding up front is whether you want to support only the happy-path parser/template combinations first or expose a broader "best effort" tool-calling surface.

The failure modes that usually matter most are:

• chat template says tool calling is enabled, but the selected model template does not actually emit callable JSON reliably
• the model emits structurally valid JSON that does not match the declared tool schema
• the runtime cannot tell whether a generation is a final answer or an intermediate tool call without brittle heuristics

A practical rollout pattern is:

1. gate support behind an explicit capability check per model/template
2. persist the raw tool-call payload for debugging
3. fail closed when the emitted call cannot be validated against the schema
4. make the second-turn tool-result format part of the contract, not an implementation detail

We ended up caring a lot about that in MartinLoop-style agent runs because most of the pain was not "can the model call a tool once" but "can the loop stop cleanly and leave debuggable evidence when the call shape drifts."