Skip to content

Data handling

This page is for security reviewers. It states where data lives in each deployment model and which outbound connections a deployment makes. Each statement names the setting or component that makes it true.

Xberg Pro and Xberg Enterprise are both self-hosted. Pro runs as one container next to your PostgreSQL and your object storage. Enterprise runs on your Kubernetes cluster with your PostgreSQL, your NATS and your object storage. In both cases the documents, results and metadata stay in infrastructure you operate. Nothing enters Xberg-operated infrastructure.

Data residency is a deployment choice, not a setting. You choose the cloud, region and storage bucket by choosing where to deploy.

Data Where Setting
Uploaded documents and extraction results Your object storage or local disk STORAGE_PROVIDER (fs, s3, gcs, azure) with STORAGE_BUCKET or STORAGE_FS_ROOT
Job metadata, audit log, projects, API keys Your PostgreSQL DATABASE_URL
RAG chunks and vectors Your PostgreSQL, with pgvector DATABASE_URL
Extraction result cache (Enterprise) NATS JetStream key-value store, 24-hour TTL NATS runs in your cluster
Local result cache (Pro) Local disk, and your object storage when the object tier is on CACHE_ENABLED, CACHE_DIR, CACHE_MAX_BYTES, CACHE_TTL_SECS, CACHE_OBJECT_STORE_ENABLED, CACHE_OBJECT_PREFIX
Integration credentials and OAuth tokens Your PostgreSQL, encrypted with AES-256-GCM INTEGRATION_KEK
Model weights Local disk or your object storage, rooted at XBERG_CACHE_DIR HF_HOME (Pro), worker.modelCache (Enterprise)

Project data access uses PostgreSQL row-level security and scoped capability functions. Deployment coordination uses separate installation-scoped capabilities. See Security.

Set dataPolicy: extraction_only in the Pro or Enterprise Helm chart. Outside Helm, set XBERG_DATA_POLICY=extraction_only in every participating service. The policy blocks RAG payload reads and writes and disables extraction-result caching. Status, cancellation and erasure remain available for existing RAG data. Pro also exposes configMap.PRO_RAG_ENABLED to disable the RAG feature separately.

An accepted extraction-only job keeps that restriction if you later change the operator policy to standard. Changing the policy does not erase historical rows, objects, caches or backups, so apply your erasure and retention procedures separately.

PostgreSQL remains required. Operational records include identifiers, ownership, filenames, MIME types, configuration, lifecycle and recovery information, and usage counters. Extraction-only mode suppresses document-derived submission metadata, bounds persisted extraction failures and skips Pro structured extraction recording. Input and result objects still use your configured storage provider and retention rules.

The private MCP listener follows the data policy. The rag mode requires the standard policy, and the extraction mode (Pro chart 0.7.5 or later) requires extraction_only. The listener is disabled by default, binds to loopback when enabled, and is not exposed through public ingress.

The system cannot satisfy a requirement that no customer information enters PostgreSQL. XBERG_DATA_POLICY=strict_sql is rejected because no separate strict SQL policy exists.

Outbound connections depend on your configuration.

  • Your configured infrastructure. Services connect to PostgreSQL, object storage and, on Enterprise, the NATS broker. Workload identity can also contact the cloud provider’s credential endpoints, such as AWS STS or the EKS Pod Identity agent.
  • Model downloads. On first use, models come from Hugging Face Hub and, for Tesseract language files that the image does not ship, from GitHub. A deployment without internet egress pre-seeds the cache instead. See Air-gapped installation.
  • LLM providers, only when you configure one. Structured extraction, vision OCR and LLM reranking send document content to the provider named in XBERG_MANAGED_LLM_PROVIDER, XBERG_VLM_OCR_MODEL or PRO_VISION_BASE_URL. On Enterprise, XBERG_VLM_OCR_BASE_URL (Helm: worker.env.XBERG_VLM_OCR_BASE_URL) sends vision OCR to an OpenAI-compatible model server you host, such as vLLM or Ollama. The endpoint allowlist (VLM_ENDPOINT_ALLOWLIST on Enterprise, PRO_VLM_ENDPOINT_ALLOWLIST on Pro) must name any private host. With no provider configured, no document content leaves the deployment.
  • Webhooks. Deliveries go to the URLs you register. The outbound policy refuses private and link-local addresses. See Security.
  • Your identity provider, only when OIDC login is configured. Token verification fetches the issuer’s JWKS. Set OIDC_JWKS_URI to skip discovery, or leave it unset and Xberg asks the issuer at OIDC_ISSUER_URL.
  • OAuth providers, only when the integration hub is configured. Access goes to the provider’s APIs with the OAuth client you registered.
  • Crawling. A crawl job fetches the URLs you submit.

License verification is local. Pro checks the Ed25519 signature of XBERG_LICENSE_KEY against a public key inside the image at start and re-checks it hourly. A license is required to run. See Licensing.

Telemetry (traces, metrics and logs) goes to the OTLP collector you name in OTEL_EXPORTER_OTLP_ENDPOINT. With the variable unset, logs stay on stdout and metrics stay on the local /metrics endpoint.

For certifications, see Security.