Key facts
| Router model | model="plugsky-fusion" selects per request (live) |
| Strategies | cost_saver for admin volume, balanced for interactive tools, max_quality for analysis |
| Routing inputs | Token count, tool use and prompt heuristics plus workspace strategy |
| Overrides | Fixed model names bypass routing for reproducible clinical-adjacent flows |
| PII handling | No-PII mode auto-redacts identifiers before inference |
| Residency | Region pinning; VPC, on-prem and air-gapped deployment available |
| Audit | Model, strategy, rule and request metadata logged and exportable |
| Roadmap | Classifier routing /v1/plugsky/route (model=auto) is coming soon |
TL;DR
- Treat admin triage and clinical reasoning as different routing classes.
- Give interactive clinician tools a latency budget and a balanced strategy.
- Redact PHI before inference and keep the workspace region-pinned.
- Pin a model where a workflow must reproduce exactly.
- Review the decision log with clinical informatics before scaling.
How it works, step by step
- Classify workflows: administrative extraction, patient communication drafts, clinician-facing search, analysis support.
- Assign strategies per class and set a latency budget for anything a clinician waits on.
- Enable redaction and confirm the PII mode your policy requires.
- Pin fixed models for workflows that must be reproducible for review.
- Pin the workspace to one region and configure retention.
- Audit the routing log monthly and widen scope only when escalation patterns look clinically sensible.
Try it yourself
Latency budgets for clinical tools
Clinicians abandon tools that make them wait. Routing lets you treat latency as a per-workflow constraint: clinician-facing search and lookup run balanced or a pinned fast model, while asynchronous summarisation and coding support can use max_quality because nobody is staring at a spinner.
The router inspects token count, tool use and prompt heuristics, so trivial lookups stay on cheap fast paths and genuinely hard cases escalate. Measure with the latency tester on your own prompts before committing to a strategy; averages hide the tail that clinicians actually feel.
PHI-aware routing
Every routed request is still a prompt containing clinical context, so redaction comes before routing decisions matter. No-PII mode filters and auto-redacts identifiers, detect-only tags them, and passthrough leaves controls to your application; most clinical deployments choose no-PII plus their own gateway pass.
- Residency: pin inference, embeddings and logs to one region so patient data stays in-jurisdiction.
- Retention: request and response bodies are not stored, and prompts are not used for training.
- Keys: project-scoped keys keep environments and departments separate.
Evaluation and audit before scale
Routing changes what a workflow sees, so it must be evaluated like a clinical process change. Build a case set with known outcomes, run it under the proposed strategies, and compare escalation rates, tool selection and output quality. The decision log attributes each result to a specific model, which turns quality review into something clinical informatics can actually do.
Pilot one department with humans in the loop, document the evaluation, and scale only when the evidence supports it. The free plan with plugsky-micro and plugsky-lite is enough to build the pipeline; paid tiers or the 14-day full-access trial cover the scale-up.
Honest comparison
| Concern | Plugsky routing | Single strong model | Manual tiering |
|---|---|---|---|
| Admin volume | cost_saver strategies | Premium on every call | Engineer per call |
| Interactive tools | Balanced with latency budget | Fixed latency | Ad hoc |
| Analysis support | max_quality where quality leads | Same model everywhere | Manual escalation |
| PHI controls | Redaction modes plus your gateway | Provider-dependent | You build it |
| Evidence | Model, strategy and rule logged | Single-model logs | Custom logging |
Frequently asked questions
Does routing risk patient safety?
Routing selects models; clinical accountability stays with people. Use routing for extraction, search and summarisation support, keep humans in the loop, and pin models where reproduction matters.
How do we keep clinician tools fast?
Set a latency budget, use balanced or a pinned fast model for interactive tools, reserve max_quality for asynchronous work, and verify with the latency tester.
Is PHI handled safely?
Redact before inference with no-PII mode or your own gateway, pin the workspace to one region, and rely on the platform's no-retention behaviour for request bodies.
Can we force one model for a workflow?
Yes. An explicit model name bypasses routing entirely, giving reproducible behaviour for workflows that need it.
Can it run on-prem?
Yes. The same routing layer runs in VPC, on-prem and air-gapped deployments.
How do we evaluate a routing change?
Replay a historical case set under the new strategy, compare escalation and quality per model using the decision log, then have clinical informatics sign off.
How do we start?
One workflow on the free plan, measured against historical cases, then paid tiers or the 14-day full-access trial when you scale. See the live pricing page.