Core, Pro and Enterprise
All three tiers run the same extraction engine: the same formats, the same languages, the same code. None is a cut-down build. What differs is what runs around the engine: queueing, storage, tenancy and scale.
All three are self-hosted. You run them on your own infrastructure, and documents do not pass through Xberg-operated systems.
At a glance
Section titled “At a glance”| Core | Pro | Enterprise | |
|---|---|---|---|
| License | MIT | Proprietary | Proprietary |
| Availability | Released | Released | By arrangement, with private images and chart |
| Runs as | Library, CLI, Docker, Helm | Docker Compose or Helm | Helm, on your Kubernetes cluster |
| Scale ceiling | One process | One node by default | A cluster |
| Price | Free | Per license, by contract | Per license, by contract |
| Support | GitHub, Discord | Email, dedicated Slack channel | Private channel, contractual SLA |
Core embeds the engine in your application, or runs it as a CLI or single-service API. You own scheduling, storage and retries. Documentation is at docs.xberg.io.
Pro is one container plus PostgreSQL. It provides what the platform does for a single team: durable jobs, a web console, crawling, search and retrieval, Auto-Tune, storage backends, source connectors and an audit log. One environment, no message bus.
Enterprise is the distributed platform. It includes everything in Pro, plus multi-user teams, SAML and SCIM, elastic workers, managed webhooks, document versioning and data rooms. It installs from a private Helm chart and private images, by arrangement.
Pro is one product, not a set of plans. A valid license turns on everything Pro does, and nothing is sold as an add-on. See Pricing and licensing.
Engine
Section titled “Engine”Identical in every tier.
| Capability | Core | Pro | Enterprise |
|---|---|---|---|
| 107 file formats, 141 extensions | Yes | Yes | Yes |
| 371 programming languages (tree-sitter) 1 | Yes | Yes | Yes |
| OCR: Tesseract, PaddleOCR, EasyOCR, vision models | Yes | Yes | Yes |
| Layout and table models, formula recognition | Yes | Yes | Yes |
| Tables, metadata, schema-driven structured extraction | Yes | Yes | Yes |
| Entities and PII, keywords, language detection | Yes | Yes | Yes |
| Output: Markdown, HTML, JSON, plain text, Djot, DocTags | Yes | Yes | Yes |
| Chunking and embeddings | Yes | Yes | Yes |
| Audio and video transcription | Yes | No | No |
1 Pro downloads grammars on first use, so code-file ingestion is unavailable on an air-gapped Pro install. Enterprise worker images include them. See Air-gapped installation.
Pro and Enterprise reject audio and video input when you submit it, before any bytes upload. Run Core directly when you need transcription.
Deployment
Section titled “Deployment”| Capability | Core | Pro | Enterprise |
|---|---|---|---|
| In-process library, 15 languages | Yes | Via Core | Via Core |
| CLI and MCP server | Yes | Via Core | Via Core |
| REST API | Yes | Yes | Yes |
| Client SDKs: Python, TypeScript, Go | No | Yes | Yes |
| Docker image | Yes | Yes | Yes |
| Helm chart | Single service | Single instance by default | Distributed |
| Web console: playground, job history, settings | No | Yes | Yes |
The SDKs target either a Pro or an Enterprise deployment with the same client library. Core is MIT-licensed, so Pro and Enterprise customers can use the library and CLI alongside their deployment at no cost. See SDKs.
Jobs, storage and operations
Section titled “Jobs, storage and operations”| Capability | Core | Pro | Enterprise |
|---|---|---|---|
| Durable job queue and history that survive restart | No | Yes | Yes |
| Object storage: S3, GCS, Azure Blob, filesystem | No | Yes | Yes |
| Presigned uploads: presign, upload, confirm | No | Yes | Yes |
| Result cache | On disk | Disk, plus object store | Shared, cluster-wide |
| API keys and multiple projects | No | Yes | Yes |
| Curated extraction presets | No | Yes | Yes |
| Saved presets | No | Yes | Yes |
| Extraction settings stored on a saved preset | No | No | Yes |
| Project extraction defaults | No | No | Yes |
| Auto-Tune: parameter search against your corpus | No | Yes | Yes |
| Append-only, hash-chained audit log | No | Yes | Yes |
| Prometheus metrics, OpenTelemetry traces and logs | Opt-in build | Yes | Yes |
| Grafana dashboards, delivered through support | No | Yes | Yes |
| High-availability profiles: red/green generations, fenced promotion | No | Yes | Yes |
| Data-policy controls: disable RAG and extraction caching | No | Yes | Yes |
| Sandbox extraction | No | No | Yes |
Auto-Tune searches your own corpus for the settings that extract it best, then promotes the winner to a reusable profile applied at extraction time. See Auto-Tune and Presets.
Pro enforces monthly job and page quotas per project. A request over quota gets 429 quota_exceeded with a Retry-After header until the next UTC month, and quotas are unlimited when unset. Replicated deployments need a customer-managed load balancer. See Deployment operator.
Search and retrieval
Section titled “Search and retrieval”| Capability | Core | Pro | Enterprise |
|---|---|---|---|
| Searchable document collections, ingest pipelines | No | Yes | Yes |
| Chunking strategies: text, Markdown, YAML, semantic | No | Yes | Yes |
| Hybrid search: vector plus full-text, rank fusion | No | Yes | Yes |
| Reranking: local cross-encoder, or an LLM you choose | No | Yes | Yes |
| Embedding-model migration as a tracked job | No | Yes | Yes |
| Private MCP listener, off by default | No | Yes | Yes |
Retrieval runs on PostgreSQL with pgvector, defaults to hybrid mode and fuses the two rankings by reciprocal rank. Retrieval is the whole surface: there is no chat or answer endpoint. See RAG.
The private MCP listener is off by default and binds to loopback when enabled. It serves one mode per listener: rag (list collections, list documents, retrieve) or extraction (submit a document, poll status, read the result; on Pro, chart 0.7.5 or later). Pro and Enterprise ship it with the same configuration surface, and neither routes it through public REST or ingress. See MCP.
Content sources and delivery
Section titled “Content sources and delivery”| Capability | Core | Pro | Enterprise |
|---|---|---|---|
| Source connectors | No | Yes | Yes |
| Web crawling: robots.txt-aware, per-domain rate limits | Via Core | Over HTTP | Adds browser rendering |
| Content enrichment: keywords, named entities, labels | No | Yes | Yes |
| Per-request callback URL, optionally HMAC-SHA256 signed | No | Yes | Yes |
| Registered webhook endpoints: retry, delivery history, replay, test | No | No | Yes |
| Document versions and diff | No | No | Yes |
| Multi-document splitting of combined PDFs | No | No | Yes |
| Data rooms: members, roles, invitations, publication | No | No | Yes |
Source connectors cover Google Drive (including native Docs, Sheets and Slides export), Dropbox, OneDrive, SharePoint, Notion, Confluence, S3, GCS, Azure Blob and SFTP. See Source connectors.
Pro crawls over plain HTTP. A crawl that needs a rendered page, such as a browser-only mode, stealth or a named browser backend, is refused with a clear error rather than answered with a page that was not fetched as asked. Enterprise runs those. See Crawling.
Access and tenancy
Section titled “Access and tenancy”| Capability | Core | Pro | Enterprise |
|---|---|---|---|
| Single sign-on: any OIDC provider, email-domain allowlisting | No | Yes | Yes |
| SAML 2.0 single sign-on | No | No | Yes |
| SCIM provisioning | No | No | Yes |
| Per-project isolation, enforced in the database on every query | No | Yes | Yes |
| Usage reporting and monthly job and page quotas | No | Yes | Yes |
| Team members, invitations, owner and member roles | No | No | Yes |
| Cross-project usage analytics | No | No | Yes |
| API key regeneration | No | No | Yes |
| Editable profile display name | No | No | Yes |
See Authentication, SAML single sign-on, SCIM provisioning and Teams and roles.
Processing scale and concurrency
Section titled “Processing scale and concurrency”| Core | Pro | Enterprise | |
|---|---|---|---|
| Documents processed in parallel | min(cores, 8) |
min(cores, 8) |
min(cores, 8) per worker, across all workers |
| Documents per request | Multiple | 10 | 10, fanned out |
| Queued work | 100 jobs in flight, held in memory | Unbounded, durable in PostgreSQL | Unbounded, durable in the message bus |
| Time budget per document | 10 minutes | 10 minutes | 10 minutes |
| Autoscaling | None | Vertical: give the node more cores | On queue depth, with workers scaling to zero |
| Cost when idle | None | One container | Workers scale to zero |
- Throughput depends on the workload. A text-native PDF and a 300-page scan that needs OCR differ by orders of magnitude. Treat the table as concurrency capacity, not a page rate. Xberg will size a deployment against a sample of your own documents.
- PDFs that need layout inference process one at a time per node. Raising the thread budget adds no parallel documents for that workload, so add nodes instead. This affects all three tiers.
- Pro scales vertically, Enterprise horizontally. Pro’s default ceiling is the largest node you are willing to run. A replicated extraction-only profile (
deployment.mode=replicated, 2 to 32 replicas) also exists; see Deployment operator. Enterprise adds workers on queue depth and returns them when the queue drains. - Pro and Enterprise accept the same request. The batch size, parallelism and time budget match. The difference is how far each scales out behind the request.
- Every limit above is a default, and operators can configure it. The shipped Pro container sets parallelism to 4.
Related pages
Section titled “Related pages”- What Pro replaces lists the work you stop doing when you move up from Core.
- Moving between tiers explains how to change tiers without rewriting your integration.
- Roadmap lists what is not available yet.