Routing¶

ASHA v0.4.2 - task-based model selection via Agent.

The standalone ModelRouter class was removed in v0.4.1. Routing is handled by SmartRoutingAdapter when you pass routing_config to Agent.

Smart routing¶

from asha import Agent

agent = Agent(
    routing_config={
        "chat": "gpt-4o-mini",
        "analysis": "gpt-4o",
        "code": "gpt-4o",
    },
)

response = agent.run("Analyze Q1 revenue", task_type="analysis")

task_type selects which entry in routing_config to use.

Fallback providers¶

agent = Agent(
    model="gpt-4o-mini",
    fallback_providers=[
        {"provider": "anthropic", "model": "claude-3-haiku-20240307"},
        {"provider": "ollama", "model": "llama3"},
    ],
)

Use agent.run_with_fallback() for explicit fallback behavior.

Local model (AshaFit)¶

agent = Agent(
    local_model="auto",
    sample_prompts=["Analyze customer feedback with PII."],
    privacy=True,
)

AshaFit picks a local model from your prompt corpus and hardware. Preview API - see local-advisor.md.

wrap_llm auto-select¶

from asha.integrations import wrap_llm

client = wrap_llm(
    ollama_client,
    auto_select_local_model=True,
    sample_prompts=["Typical prompts from your app..."],
)

Provider auto-detection¶

Agent infers provider from model name when provider is omitted:

Pattern	Provider
`gpt-*`	openai
`claude-*`	anthropic
`gemini-*`	gemini
`grok-*`	grok
`org/model`	huggingface
`mock`	mock
Other	ollama

Override with provider="openai" etc.

What was removed¶

Removed (v0.4.2)	Replacement
`ModelRouter`	`Agent(routing_config=...)`
`RoutingStrategy` enum	Task-type dict keys
IR-based `router.route(ir)`	Not exposed in public API

Status¶

Smart routing is preview - routing_config shape may evolve before 1.0.0.