How do legal teams use Model Fusion with Plugsky?

Key facts

Router model	model="plugsky-fusion" selects across tiers per request (live)
Long context	Long-context models handle full agreements; routing can target them by token count
Fan-out	Parallel calls let several models review the same passage for high-stakes drafts
Isolation	Project-scoped keys and collections keep matters separate
Strategies	cost_saver for discovery triage, balanced by default, max_quality for drafting
Audit	Model, strategy and rule logged per request; exportable to SIEM
Retention	Matter documents stay until deleted; request bodies are not stored
Roadmap	Classifier routing /v1/plugsky/route (model=auto) is coming soon

TL;DR

• Use long-context routing for full contracts instead of chunking by habit.
• Fan out high-stakes drafts across models and compare before a human review.
• Keep one project and key set per matter for privilege isolation.
• Run discovery triage on cheap tiers where volume is highest.
• Document every model decision for client and regulator questions.

How it works, step by step

1. Map the matter workflow: intake classification, metadata extraction, clause retrieval, analysis, drafting, review.
2. Create a project per matter with scoped keys and a dedicated retrieval collection.
3. Set a balanced default and add rules that send long documents to a long-context model.
4. Reserve max_quality and fan-out for drafts and negotiations where quality dominates cost.
5. Log every routed request and review the trail before any output leaves the firm.
6. Rebenchmark after model or rule changes, using a fixed set of representative clauses.

Try it yourself

Open the AI model selector →

Long documents and high-stakes drafts

Legal work strains a single model in two directions at once: contracts are long, and the output must be precise. Fusion addresses both. Add a rule that routes inputs above a token threshold to a long-context model, so a full master services agreement is analysed whole rather than shredded into chunks that lose cross-references.

For drafting, use max_quality or fan out the same passage to several models and compare. Parallel fan-out is valuable here precisely because reviewers can see disagreement; where models converge, confidence is higher, and where they diverge, a lawyer's attention is warranted anyway.

Matter isolation and privilege

Privilege is a data-boundary problem. Give each matter its own project, scoped keys and retrieval collection so no assistant can retrieve across matters. Keys can be revoked or rotated when a matter closes without downtime, and audit logs record key actions with actor, timestamp and request hash.

• Redaction: no-PII mode removes identifiers that are not needed for the task.
• Retention: matter documents remain until you delete them; request bodies are not stored.
• Deployment: VPC, on-prem and air-gapped options for clients who require it.

Cost control across a discovery set

Discovery volumes punish blanket frontier use. Route first-pass relevance and metadata extraction to cost_saver, keep balanced for general analysis, and reserve max_quality or fan-out for the small percentage of documents that drive the case.

Because self-serve plans are flat with unlimited fair-use usage, high-volume multi-model patterns do not translate into unpredictable token invoices. The decision log then shows how much of the corpus actually needed a premium model — useful evidence when a client asks how the bill was kept down.

Honest comparison

Concern	Plugsky Model Fusion	Legal AI suite	In-house ensemble
Long documents	Routing to long-context models	Vendor-chosen model	You wire each model
Draft quality	Fan-out and max_quality tiers	Vendor model only	Custom comparison logic
Privilege isolation	Project-scoped keys per matter	Workspace controls	You build isolation
Decision logging	Model, strategy and rule per request	Suite-level logs	Custom instrumentation
Cost shape	Flat plans, fair-use RPM	Seat or usage pricing	Per-token to each vendor

Frequently asked questions

Can fusion handle entire contracts?

Yes. Add a token-threshold rule that routes long inputs to a long-context model, so full agreements are analysed without chunking that breaks cross-references.

How does this protect privilege?

Each matter gets its own project, scoped keys and retrieval collection, and audit logs record every key action. No assistant can retrieve across matters unless you allow it.

Will outputs cite clauses accurately?

Pair fusion with retrieval tools that return clause IDs and text, and require the model to cite only tool results. That keeps citations verifiable.

Does this replace lawyer review?

No. Fusion selects models; lawyers remain responsible for advice. Use it for classification, extraction, analysis support and first drafts under review.

Can high-stakes drafting use several models?

Yes. Fan-out lets multiple models process the same passage in parallel, and the request log records which model produced which result.

Can it run in our own environment?

Yes — VPC, on-prem and air-gapped deployments are available for client-mandated workloads.

How do we justify the cost internally?

Use the decision log to show how much volume ran on cheap tiers and how little needed premium models, then compare that against your prior per-token spend.