Ollama

SchemaRouter can use a locally running Ollama model as a bounded decision backend. This does not turn SchemaRouter into an Ollama agent system: Ollama does not own the tool loop, does not construct arbitrary execution plans, and does not execute registered tools. It only selects from finite option IDs already authorized by SchemaRouter.

The integration uses Ollama structured-output format and the existing httpx dependency, so no Ollama Python SDK is required.

Prerequisites

Run Ollama separately and make sure the model is already available to that server. SchemaRouter does not pull models automatically.

pip install schemarouter

Configure a local backend

from schemarouter import DecisionPolicy, SchemaPlanner
from schemarouter.integrations import OllamaDecisionBackend

backend = OllamaDecisionBackend("your-installed-model")

planner = SchemaPlanner(
    registry,
    decision_backend=backend,
    decision_policy=DecisionPolicy(
        enabled=True,
        endpoint_selection=True,
        fallback="deterministic",
    ),
)

The default API base URL is http://127.0.0.1:11434. Trusted application code may explicitly supply another absolute HTTP(S) base URL.

Bounded structured output

For each request, SchemaRouter sends Ollama a JSON Schema whose option_id is constrained to an enum of the finite locally authorized option IDs. The model receives the query, max_selections, option IDs/labels/descriptions, and optional bounded context. DecisionOption.metadata is never forwarded.

After the server responds, SchemaRouter validates the result again. Unknown or duplicate IDs, selection-count overflow, invalid scores, malformed JSON, and inconsistent abstention fail closed before the result can affect planning.

Structured generation is a reliability aid, not the authority boundary. Local SchemaRouter validation remains authoritative even if a server or model ignores the supplied schema.

Sync and async

backend = OllamaDecisionBackend("your-installed-model")

async_backend = OllamaDecisionBackend(
    "your-installed-model",
    async_mode=True,
)

Both paths use non-streaming /api/chat requests. Temperature defaults to zero. Trusted local code can pass additional generation settings through options={...}.

Context privacy

Request context is included by default. Disable it when the context is unnecessary or sensitive:

backend = OllamaDecisionBackend(
    "your-installed-model",
    include_context=False,
)

Benchmark a local model

python scripts/benchmark_decision_routing.py --corpus benchmarks/decision-routing-v1.json --ollama-model your-installed-model --json-out artifacts/ollama-decision-benchmark.json

Use --ollama-base-url when a trusted Ollama endpoint is not on the default loopback address. Ollama token counters are retained in benchmark metadata when present.

Reported selection scores are model self-assessments. SchemaRouter does not treat them as calibrated probabilities.

Scope

This backend provides a practical local/open-model path behind the generic DecisionBackend contract. It does not claim equivalence to Jev/System One or to research routers such as TRINITY/TinyRouter. Those systems have different training objectives and need separate measured evidence before quality comparisons are made.