Key facts
| Router model | model="plugsky-fusion" applies rules per request (live) |
| Long-document rule | Prompts above a token threshold route to a long-context model |
| Strategies | cost_saver for discovery volume, balanced default, max_quality for drafting |
| Isolation | Project-scoped keys and collections keep matters separate |
| Overrides | Explicit model names bypass routing for reproducible review |
| Decision log | Chosen model, strategy and rule recorded per request |
| Retention | Matter documents stay until deleted; request bodies are not stored |
| Roadmap | Classifier routing /v1/plugsky/route (model=auto) is coming soon |
TL;DR
- Add a token-threshold rule so full agreements reach long-context models.
- Run discovery triage on cheap tiers where volume is highest.
- Keep per-matter keys so routing and retrieval never cross boundaries.
- Pin a model for review workflows that must reproduce.
- Document routing decisions alongside the work product.
How it works, step by step
- Classify documents and tasks: intake, metadata extraction, clause analysis, drafting, review.
- Create one project and key set per matter, each with its own strategy.
- Add a token-threshold rule routing long documents to a long-context model.
- Set cost_saver for discovery triage and max_quality for drafting or negotiation work.
- Pin fixed models where a reviewer must reproduce an earlier result exactly.
- Review the decision log before outputs leave the firm and retain it with the matter record.
Try it yourself
Open the LLM cost calculator →
Rules built around document length
Legal workloads split cleanly by document size. Emails and short memos fit any model; a full master services agreement does not. A token-threshold rule — send inputs above roughly 4,000 tokens to a long-context model — is the single highest-value routing rule a legal team can add, because it prevents silent quality loss from truncation or over-aggressive chunking.
Combine it with a strategy per task: cost_saver for first-pass relevance and metadata extraction, balanced for general analysis, and max_quality for drafting that a partner will revise.
Matter isolation and cost control
Each matter gets a project, scoped keys and its own retrieval collection, and the strategy travels with that scope. Discovery-heavy matters can run cheap-first without affecting a bet-the-company matter that needs premium tiers, even though both sit on the same platform.
- Volume control: route relevance screening away from premium tiers; only escalate documents that matter.
- Predictable spend: flat self-serve plans carry no per-token charges, so high-volume discovery does not create runaway invoices.
- Privilege: keys are revocable on matter closure, and audit logs record key actions.
Documented decisions for the file
When a client or regulator asks how a document was analysed, the decision log answers concretely: which model handled which request, under which strategy, triggered by which rule. That record can be retained with the matter alongside the customary review notes.
Where reproducibility is required — a second-pass review of a production, for example — pin the model explicitly and bypass routing. Otherwise, review routing distributions after major matters to see whether thresholds should move, and use the cost calculator to sanity-check the economics before proposing changes to the partners.
Honest comparison
| Concern | Plugsky routing | One model for all work | Manual model choice |
|---|---|---|---|
| Long contracts | Token-threshold rule to long-context models | Truncation or chunking risk | Engineer decides |
| Discovery volume | cost_saver strategies | Premium on every document | Ad hoc |
| Drafting | max_quality strategies | Same model as triage | Manual escalation |
| Matter isolation | Project-scoped keys per matter | Shared keys | Shared keys |
| Decision record | Model, strategy and rule logged | Single-model logs | None |
Frequently asked questions
How do we stop long contracts from being truncated?
Add a token-threshold rule that routes inputs above roughly 4,000 tokens to a long-context model, and verify with your own longest documents.
Can each matter have its own routing?
Yes. Strategies are scoped per workspace or API key, so each matter project can run its own policy.
Does routing protect privilege?
Projects, scoped keys and dedicated collections separate matters, and keys can be revoked when a matter closes. Routing operates inside that boundary.
Can we reproduce an earlier analysis?
Yes. Pin an explicit model on that workflow and routing is bypassed, so the same request takes the same path.
How does this control discovery costs?
Route relevance screening to cheap tiers and reserve premium models for documents that matter. Flat plans mean high volume does not create per-token cost.
What appears in the decision log?
The chosen model, strategy and rule for every request, alongside token, latency, status, key, project and region metadata.
Is classifier routing available?
Not yet — /v1/plugsky/route with model=auto is coming soon. Explicit strategies and custom rules are live today.