Key facts
| Live capabilities | Vision input, embeddings and built-in RAG |
| Coming soon | Image generation, audio transcription, speech and moderation endpoints |
| Request limit | 16 MB maximum request body; split large scans |
| Model access | 30+ models behind one OpenAI-compatible endpoint |
| Residency | Region pinning plus VPC, on-prem, air-gapped and BYOC topologies |
| Observability | Model, tokens, latency, status, key, project and region logged per request |
| Governance | PII modes, scoped keys, RBAC and SIEM-exportable audit events |
| SLA | 99.9% to 99.95% single-region, 99.99% multi-region deployments |
TL;DR
- Productionise vision, embeddings and RAG now; roadmap image and audio.
- Design one internal multimodal interface so future endpoints are additive.
- Treat images and scans as sensitive data with the same controls as text.
- Pin inference by region and keep observability per request.
- Ask vendors for endpoint status in writing before promising a capability.
How it works, step by step
- Produce a capability map separating live endpoints from coming-soon ones, and circulate it with product teams.
- Define one internal interface for model calls so modality changes are provider changes.
- Design data handling for binary inputs: redaction, retention, storage location and access control.
- Choose embedding and retrieval patterns per use case, with collections scoped to domains.
- Set residency and topology per workload before the first pilot, not after.
- Add routing and audit requirements to the platform standard so multimodal inherits them for free.
Original data
Try it yourself
Open the embedding model comparison →
A capability map: live versus roadmap
The most common multimodal architecture mistake is designing for parity with a vendor's marketing page. Today, Plugsky's live stack is vision in chat completions, embeddings and retrieval-augmented generation. Image generation, audio transcription, speech and moderation endpoints are coming soon.
Publish that distinction internally. It lets product teams build document understanding and semantic search immediately, while anything requiring generated imagery or voice stays behind a roadmap dependency — and it gives you a defensible answer when a stakeholder asks why a competitor's demo includes audio.
Data handling for binary content
Images, scans and audio are data with the same classification as the text they contain, often more. Extend your existing controls rather than inventing parallel ones: redaction before inference, project-scoped keys per environment, retention rules per corpus, and residency pinning for inference, embeddings and logs together.
- Request size: the 16 MB body limit forces per-page splitting — a useful constraint because it keeps payloads reviewable.
- Storage: where you store the original artefacts is your decision; keep it consistent with the region you pinned for inference.
- PII modes: no-PII, detect-only and passthrough cover text prompts; pair them with container-level controls for binaries.
Integration roadmap and vendor risk
Architect for addition, not replacement. One internal interface that takes messages and returns completions can carry vision today and accept audio or image output later without touching product code. Keep feature flags around anything unshipped, and write endpoint status into your vendor risk register so a roadmap item cannot quietly become a launch dependency.
The platform already provides the surrounding governance: 30+ models behind one endpoint, region pinning, scoped keys, RBAC, audit export and SLAs from 99.9% up to 99.99% on multi-region deployments. Multimodal inherits all of it, which is the real argument for consolidating rather than adding a second vendor for vision alone.
Honest comparison
| Dimension | Plugsky | Single-modality vendors | Self-hosted multimodal |
|---|---|---|---|
| Live modalities | Vision, embeddings, RAG | Usually one | Whatever you serve |
| Roadmap transparency | Endpoints labelled coming soon in docs | Varies | Under your control |
| Governance | Scoped keys, RBAC, audit export | Separate per vendor | You build it |
| Residency | Region pinning, VPC, on-prem, air-gapped | Limited regions | Wherever you deploy |
| Integration surface | One OpenAI-compatible endpoint | One SDK per vendor | Custom serving |
Frequently asked questions
Which multimodal endpoints are production-ready?
Vision input in chat completions, the embeddings endpoint and built-in RAG are live. Image generation, audio and moderation endpoints are coming soon.
How do we avoid building on unshipped features?
Keep a written capability map, gate anything unshipped behind feature flags, and record endpoint status in your vendor risk register.
How should images and scans be governed?
Treat them as sensitive data: redact before inference, scope keys per environment, pin inference and logs to a region, and apply retention rules per corpus.
What is the request size limit?
Request bodies cap at 16 MB, so split multi-page documents into page-sized calls and reassemble in your application.
Which embedding setup should we standardise on?
Start with the platform's embedding model for semantic search and retrieval, and compare alternatives against your own corpus before standardising.
Can everything run in our own environment?
Yes. The same API runs in VPC, on-prem, air-gapped and bring-your-own-cloud topologies where data cannot leave your estate.
What governance comes with it?
Project-scoped keys, RBAC, PII modes, per-request logging exportable to SIEM, and SLAs from 99.9% to 99.99% depending on topology.