Requirements
PostgreSQL 17, optional pgvector, Redis, S3-compatible storage, the container images, and what to plan for.
Services
| Service | Required | Notes |
|---|---|---|
| PostgreSQL 17 | Yes | Everything durable. DATABASE_URL must be a direct connection or a session-mode pooler: background jobs use session-level advisory locks, which transaction-mode pooling breaks. |
| pgvector | No | Only for meaning-based search of project files. OCI detects it but never creates it. The bundled postgres:17-alpine image does not include it; see Upgrades → pgvector. |
| Redis | Recommended; needed for several replicas | Resumable replies, shared rate limits, cancellation. The bundled stack uses redis:7-alpine. Without it, live replies still work but cannot be resumed after a disconnect, and rate limits are per replica. |
| File storage | Yes | A local volume for one API replica, or S3-compatible storage (AWS S3, MinIO and similar) for more than one. Configured in the dashboard. |
| SMTP | For invitations, password resets, verification and reports | Configured in the dashboard. |
| A TLS reverse proxy | For production | In front of the web container. See Reverse proxy. |
Images
| Image | Runs | Port |
|---|---|---|
ghcr.io/ncecere/open-chat-interface/web:<version> | Caddy with the browser app; forwards /api/* to the API | 8080 |
ghcr.io/ncecere/open-chat-interface/api:<version> | Node.js 22 API, as the non-root oci user; includes the PostgreSQL 17 client tools for backups and about 10 MB of Noto fonts for PDF export | 3000 |
Each release is tagged vX.Y.Z (such as v0.10.2), and latest follows the newest stable release. Pin both images to the same version; latest is for evaluation only. Published images are linux/amd64 only.
Database connections
Budget, per API replica, the application pool (ten connections by default) plus one private connection per background job running at once. Include that headroom in PostgreSQL's connection limit.
Sizing
OCI has not been measured at large scale yet; test your own install. Points to plan for:
- File exports (Word, PDF, PowerPoint, Excel) run in a worker thread with a 512 MB heap limit and a 60-second time limit. A replica runs at most two at once. Budget up to two CPU cores and about 1.2 GB of extra memory per API replica for exports at peak.
- Imports of ChatGPT and Claude exports accept uploads up to 512 MB by default (
IMPORT_MAX_UPLOAD_BYTES), processed in the background. - The search index on message text takes disk space, a sizeable fraction of the stored message text. Project-file passages take about as much space as the files' extracted text.
Outbound network
The API makes outbound requests to: your model providers, your web search provider, connector MCP servers and their authorization servers, rerank and embedding endpoints, SMTP, S3 (files, backups and compliance export), webhook endpoints, your OIDC/SAML identity provider, and your OpenTelemetry collector. Allow that egress where it is filtered.
Browsers
The app is a single-page application for current versions of the major browsers, on desktop and phone. It targets WCAG 2.2 AA.