Key facts
| Router model | model="plugsky-fusion" fans out and escalates per request (live) |
| Escalation | Default chain runs from plugsky-micro up to plugsky-pro and frontier tiers |
| Strategies | cost_saver, balanced, max_quality and custom rules |
| PII handling | No-PII mode auto-redacts, detect-only tags, passthrough for your own controls |
| Residency | Pin workspaces to a single region; VPC, on-prem and air-gapped available |
| Audit | Model, strategy, rule and request metadata logged and exportable to SIEM |
| Retention | Request and response bodies are not stored; prompts are not used for training |
| Roadmap | Classifier routing /v1/plugsky/route (model=auto) is coming soon |
TL;DR
- Route forms, coding and extraction to cheap tiers; escalate reasoning.
- Redact PHI before inference and keep the workspace region-pinned.
- Use fusion's decision log as part of your clinical audit evidence.
- Keep clinicians in the loop for anything that influences care.
- Benchmark fusion on your own case set before widening scope.
How it works, step by step
- Split workflows by clinical weight: administrative extraction, coding support, summarisation, decision support drafts.
- Set strategies per workspace or key so routine work defaults to cheaper tiers.
- Enable redaction and confirm which PII mode your policy requires.
- Build an evaluation set from historical cases and compare fusion against your current model on accuracy and escalation.
- Pin the workspace to one region and configure retention.
- Review the routing log weekly with clinical informatics, then widen scope when precision holds.
Try it yourself
Open the LLM benchmark explorer →
Matching models to clinical tasks
Clinical work is not one workload. Reading an intake form, normalising a medication list and flagging a care gap are extraction tasks that cheap models handle well. Longitudinal summarisation, differential-style reasoning support and policy interpretation benefit from stronger tiers. Fusion expresses that split as strategy rather than code.
Set balanced as the default, cost_saver for administrative pipelines, and max_quality for summarisation that a clinician will act on. The router escalates on hard prompts within the default chain, so a harder-than-expected extraction does not silently fail on a weak model.
PHI, residency and review
Redact before you prompt. No-PII mode filters and auto-redacts identifiers; detect-only tags them without modification; passthrough leaves controls entirely to your application. Most clinical deployments choose no-PII plus a second redaction pass in their own gateway, because model routing should never be the only thing standing between a patient identifier and a provider.
- Residency: pin inference, embeddings and logs to one region so data stays in-jurisdiction.
- Retention: request bodies are not stored and prompts are not used for training.
- Review: anything that influences care stays behind a named clinician's judgement.
Evidence before scale
Before widening access, benchmark. Build a case set with known outcomes, run it through fusion and your incumbent configuration, and compare tool selection, escalation behaviour and output quality. Fusion's per-request log lets you attribute errors to the model that produced them, which is far more useful than a single aggregate score.
Pilot one department, keep humans in the loop, and document the evaluation for your clinical governance committee. The free plan with plugsky-micro and plugsky-lite is enough to build and measure the pipeline; move to paid tiers or the 14-day full-access trial when you scale. Enterprise contracts add a BAA and customer-managed keys.
Honest comparison
| Concern | Plugsky Model Fusion | Single strong model | Manual tiering |
|---|---|---|---|
| Routine extraction | Cheap tiers by default | Overpaying per call | Engineer chooses each time |
| Escalation | Automatic on hard prompts | None | Manual retry |
| PHI controls | Redaction modes plus your own gateway | Whatever provider offers | You build redaction |
| Residency | Region pinning plus on-prem and air-gapped | Provider regions | Whatever you deploy |
| Audit | Model choice logged per request | Single model, less detail | Custom logging |
Frequently asked questions
Is fusion safe for clinical use?
Fusion is an infrastructure feature, not a clinical decision-maker. Use it for extraction, summarisation and drafting support, keep humans accountable for care decisions, and validate on your own case set.
Which models will handle patient-adjacent text?
Routing selects within your tier across 30+ models, from plugsky-micro and plugsky-lite to plugsky-pro and frontier tiers. You can pin a fixed model wherever policy demands it.
How is PHI protected?
Use no-PII mode for automatic redaction, add your own gateway pass, pin the workspace to one region, and keep request bodies unstored.
Does fusion help with evaluation?
Yes. Per-request logs show which model answered and why, so you can attribute quality and error rates by tier instead of guessing.
Can it run on-prem?
Yes. The same API runs in your VPC, on-prem or air-gapped where patient data cannot leave the environment.
Will routing add latency to clinician-facing tools?
Fan-out can add overhead. Use cost_saver or balanced for interactive tools, reserve max_quality for asynchronous summarisation, and measure before rollout.
How do we start?
Build the pipeline on the free plan, measure on a historical case set, then move to paid tiers or the 14-day full-access trial. See the live pricing page for plan details.