Key facts
| Workflow fit | Review, extraction, summarisation and chronologies; legal judgement stays with lawyers |
| Grounding | Retrieval with /v1/embeddings and plugsky-embed; cite sources, never answer from memory alone |
| Matter isolation | Separate workspaces and scoped keys per matter or client |
| Residency | Region-locked planes: EU (Frankfurt), GCC (UAE), APAC (Singapore), US (Virginia) |
| Deployment | Cloud, VPC private endpoint, on-prem and air-gapped for privileged workloads |
| Retention | Configurable prompt retention; shorten windows for privileged material |
| Audit | Key, admin and inference events exportable to SIEM with per-matter attribution |
| Endpoint status | Chat completions, streaming, JSON mode and function calling are live; assistants and responses are coming soon |
TL;DR
- One workspace and key set per matter keeps privilege walls technical, not procedural.
- Ground every answer in retrieved sources and require lawyer verification of citations.
- Cut retention windows for privileged material and log every access.
- Read-only tools first; document edits flow through tracked changes.
- Build on live chat completions; assistants-style endpoints are roadmap items.
How it works, step by step
- Choose one review-heavy workflow — clause extraction, diligence summarisation or chronology building — and define the required output format.
- Create a workspace and scoped keys per matter or client so retrieval and logs cannot cross boundaries.
- Index the matter documents with /v1/embeddings using plugsky-embed and store vectors inside the matter workspace.
- Build the loop to answer only from retrieved passages, returning citations alongside every claim.
- Add verification steps: a lawyer reviews flagged clauses and confirms each citation resolves to the source document.
- Shorten prompt retention for privileged material and stream access and inference events to your SIEM.
- Track precision and missed-issue rates against a manually reviewed sample before expanding to new matter types.
Try it yourself
Open the RAG architecture builder →
What agents do well in legal work
Legal agents shine where the task is exhaustive and mechanical: extracting clauses from hundreds of contracts, building chronologies from correspondence, summarising diligence documents, preparing redaction candidates and running first-pass research over a closed corpus. These tasks reward consistency and tire humans, which is exactly where a tool loop helps.
In Plugsky terms, the agent calls /v1/chat/completions with retrieval tools and returns structured output. It does not opine on strategy, and it never answers from model memory when a citation is expected. Every substantive claim resolves to a passage a lawyer can open.
Privilege, isolation and retention
Privilege is an architectural property once AI is involved: if a workspace is shared, a wall is only procedural. Use a workspace and key set per matter or client, keep vector indexes inside the same boundary, and align log residency with document residency.
- Least data: index only what the matter needs; avoid central indexes containing everything.
- Short retention: configure prompt retention per workload and purge when the matter closes.
- Traceability: log who queried which matter, which documents were retrieved and which model answered.
Where client contracts or regulators require it, run the workload in a VPC, on-prem or air-gapped deployment so documents never leave your perimeter.
Accuracy controls that survive scrutiny
The failure mode that damages credibility is a confident wrong citation. Defend against it with retrieval discipline: chunk documents carefully, require the model to quote supporting text, and add a verification pass that checks each citation against the source. Flag any answer with no retrievable support.
Measure on a reviewed sample — precision of extraction, missed issues, false citations — and re-run the suite whenever prompts or models change. With 30+ models on one compatible endpoint you can test a cheaper model for extraction and reserve frontier models for synthesis, adjusting per matter type without rebuilding the workflow.
Honest comparison
| Concern | Plugsky agents | Typical hosted platform | Building in-house |
|---|---|---|---|
| Matter isolation | Workspace and keys per matter or client | Shared tenancy by default | You design tenancy |
| Grounding | Retrieval with plugsky-embed plus citation checks | Varies; often memory-based | You build retrieval and evals |
| Confidentiality | Region-locked planes plus VPC, on-prem, air-gapped | Limited deployment control | Wherever you run |
| Retention | Configurable per workload; purge at matter close | Usually fixed | Your implementation |
| Audit | Per-matter access and inference events to SIEM | Vendor dashboards | Your pipelines |
| Model choice | 30+ models behind one endpoint | Vendor catalogue | You serve each model |
Frequently asked questions
Will our documents train a public model?
No. Prompts and completions are not used to train public models. Retention is configurable, and enterprise deployments support customer-managed keys and perimeter-only operation.
How do we stop hallucinated citations?
Answer only from retrieved passages, require quoted support, and add a verification pass that resolves every citation to a source document before a lawyer reviews it.
Should each matter have its own workspace?
Yes. Per-matter workspaces and keys make privilege walls technical rather than procedural, and keep retrieval indexes and logs inside the same boundary.
Can we run legal AI air-gapped?
Yes. On-prem and air-gapped tiers use open-weight models with the same OpenAI-compatible API, so privileged material never leaves your perimeter.
What should we log for legal work?
Matter identifier, user, documents retrieved, model and prompt version, and the final output. Export access and inference events to your SIEM with per-matter attribution.
How do we evaluate quality?
Build a manually reviewed sample with known answers, then measure extraction precision, missed issues and citation accuracy. Re-run it on every prompt or model change.
How do we start small?
Run a single matter on the free plan with plugsky-micro and plugsky-lite, or use the 14-day full-access trial for a heavier workload. Plan details are on the live pricing page.