Key facts
| Vision | Image and document input in chat completions (live) |
| Embeddings and RAG | Semantic search and built-in retrieval collections (live) |
| Deployment | Region-pinned cloud, VPC, on-prem and air-gapped bundles |
| Coming soon | Image generation, audio transcription, speech and moderation endpoints |
| PII handling | No-PII mode auto-redacts, detect-only tags, passthrough for your controls |
| Audit | Model, strategy, tokens, latency, status, key, project and region logged; exportable to SIEM |
| Residency | Data never leaves the pinned region; sa-central-1 available on Enterprise |
| Review | Human officers remain accountable for citizen-affecting decisions |
TL;DR
- Use live vision and RAG for document-heavy administrative work.
- Keep audio and image generation on the roadmap, not in citizen-facing promises.
- Redact before inference and run inside the approved boundary.
- Log every request and export the trail to your SIEM.
- Keep human officers accountable for decisions that affect citizens.
How it works, step by step
- Choose an administrative process — permit intake, correspondence triage or policy search — with clear success measures.
- Confirm the deployment topology and classification level for that process before building.
- Send form or letter images through vision extraction with strict JSON schemas.
- Index policy and procedure corpora for retrieval so answers cite official documents.
- Redact personal identifiers and pin inference, embeddings and logs to the required region.
- Keep officers reviewing outputs and export request logs for oversight.
Try it yourself
Open the RAG architecture builder →
Document processing inside the boundary
Much of government back-office work is document handling: forms, letters, permits, case notes. Vision input in chat completions reads those documents directly, extraction can be forced into structured JSON, and retrieval over policy corpora lets staff ask questions that resolve to official sources rather than model recollection.
All of it runs inside the deployment boundary you choose — region-pinned cloud, VPC, on-prem or fully air-gapped. In an air-gapped installation the models are served locally with no outbound calls, which is usually the deciding requirement for citizen data.
Planning audio and image endpoints honestly
Audio transcription, speech output and image generation are coming soon and should be documented as roadmap items. The temptation is to prototype around them and discover later that a launch depends on an unshipped capability. Instead, keep modality-specific code behind one interface and feature flags so enabling them is additive.
- Status discipline: record endpoint availability in your capability register and re-check before commitments.
- Language: benchmark model quality for the languages your service covers with your own evaluation set.
- Limits: the 16 MB request body means multi-page scans are processed per page.
Oversight, evidence and accountability
Automation in public administration must be explainable. Every request through Plugsky logs the model, tokens, latency, status, key ID, project and region, and those events can be exported to your SIEM. Pair that with no-PII redaction and a named officer reviewing anything that affects a citizen's case.
Procurement-wise, enterprise agreements include a DPA with options for in-region routing, and sovereign packages are scoped to deployment, security and volume. Start with an internal or low-sensitivity process, document the evaluation, and let the evidence — not a vendor demo — decide the expansion path.
Honest comparison
| Concern | Plugsky | Public cloud AI | Agency-built stack |
|---|---|---|---|
| Deployment | Cloud, VPC, on-prem, air-gapped | Vendor regions | Own data centre |
| Live modalities | Vision, embeddings, RAG | Varies | Whatever is served |
| PII handling | No-PII, detect-only, passthrough | Varies | You build it |
| Audit | Per-request metadata to SIEM | Vendor logs | Own pipeline |
| Roadmap clarity | Endpoints labelled coming soon | Marketing-led | Under your control |
Frequently asked questions
Can vision run in an air-gapped environment?
Yes. Air-gapped bundles run locally with no external connectivity, so document processing stays inside the boundary.
Are citizen documents used for training?
No. Prompts and completions are not used to train models and request bodies are not retained; no-PII mode adds automatic redaction.
When will audio transcription arrive?
The audio endpoints are coming soon per the docs. Plan them as roadmap items and keep modality code behind a flag until they are live.
How are decisions audited?
Each request logs model, tokens, latency, status, key, project and region, exportable to your SIEM, and officers remain accountable for outcomes.
Does it support our official languages?
The platform serves multilingual workloads; validate quality for each language on your own documents before committing to a service level.
What is the document size limit?
The maximum request body is 16 MB, so multi-page scans should be split and processed per page.
How should an agency start?
Pick a low-sensitivity administrative process, document the evaluation criteria, run a bounded pilot with officers in the loop, and review the logs before expanding.