Key facts
| Vision | Scanned and photographed documents in chat completions (live) |
| RAG | Matter-scoped retrieval collections with citations (live) |
| Embeddings | /v1/embeddings for semantic search across document sets (live) |
| Audio | Transcription and speech endpoints with whisper-plugsky — coming soon |
| Isolation | Project-scoped keys and collections keep matters separate |
| Audit | Model, tokens, latency, status, key, project and region logged per request |
| Retention | Matter documents stay until deleted; request bodies are not stored |
| Limits | Maximum request body of 16 MB; split long documents per section |
TL;DR
- Read scanned contracts and exhibits with live vision input.
- Search matter files with RAG and require citations to source documents.
- Keep per-matter keys and collections so privilege boundaries hold.
- Plan audio transcription as roadmap, not as a current workflow.
- Retain request logs with the matter record for review.
How it works, step by step
- Create a project and scoped keys per matter, with a dedicated retrieval collection.
- Send scanned or photographed documents through vision and extract clauses into structured fields.
- Index the matter corpus for retrieval and require answers to cite the source passage.
- Set retention for matter documents and confirm redaction for identifiers not needed for the task.
- Keep lawyer review in front of anything used in advice, negotiation or filing.
- Retain the request log with the matter file and review routing or extraction accuracy after large reviews.
Try it yourself
Document review with vision and RAG
Legal work arrives as much in scans and exhibits as in clean text. Vision input in chat completions reads photographed contracts, stamped exhibits and correspondence, extracting clauses and metadata into structured fields. In parallel, embeddings and retrieval let reviewers ask questions across the matter corpus and receive answers anchored to source passages.
The combination is what makes it defensible: extraction produces fields a reviewer can verify, and retrieval produces citations rather than ungrounded assertions. Because the endpoint is OpenAI-compatible, existing document pipelines can add vision without a new vendor.
Privilege, isolation and retention
Privilege is a boundary problem, and boundaries are per matter. Give each matter its own project, scoped keys and retrieval collection so no assistant retrieves across matters. Keys can be revoked when a matter closes without downtime, and audit logs record key actions with actor, timestamp and request hash.
- Retention: matter documents remain until you delete them; request and response bodies are not stored.
- Deployment: VPC, on-prem and air-gapped options for client-mandated work.
- Redaction: remove identifiers that are not relevant to the task before inference.
Audio and the roadmap
Deposition and hearing transcription is the capability most legal teams ask about next. The audio endpoints — transcription via /v1/audio/transcriptions and speech via /v1/audio/speech, with whisper-plugsky planned — are coming soon. Treat them as roadmap: do not promise a client a transcript pipeline built on an unshipped endpoint.
Prepare the workflow now. Define who may record, how consent is captured, where transcripts are stored and how long they are kept. Keep the transcription call behind one interface so the swap is additive. When the endpoint ships, the process and the controls are already in place, and only the input layer changes.
Honest comparison
| Capability | Plugsky status | Review requirement | Privilege note |
|---|---|---|---|
| Scanned contract reading | Live via vision | Lawyer verifies fields | Matter-scoped collection |
| Matter-wide search | Live via embeddings and RAG | Citations to source | Per-matter keys |
| Clause extraction | Live via chat completions | Structured validation | No cross-matter retrieval |
| Deposition transcription | Coming soon | Transcript review | Consent and retention plan |
| Speech output | Coming soon | Not for client advice | Out of scope for now |
Frequently asked questions
Can it read scanned contracts?
Yes. Vision input in chat completions accepts scans and photographs, and extraction can be forced into structured fields your reviewers validate.
How do citations stay reliable?
Use retrieval so the model cites passages from the matter corpus, and require answers to reference a source rather than relying on model memory.
Is privilege protected?
Each matter gets its own project, scoped keys and collection, so retrieval cannot cross matters. Keys are revocable when a matter closes.
Can we transcribe depositions?
Not yet. Audio endpoints are coming soon; define consent, storage and retention now, and keep transcription behind an abstraction.
What happens to uploaded documents?
Matter documents remain until you delete them, and request bodies are not stored. Configure retention to match the engagement.
How large can a document be?
Request bodies cap at 16 MB; split long documents by section or page and combine the results.
How should a firm pilot this?
One matter, one document type, lawyer review on every output, and a measured error rate before expanding to the wider practice.