Process - How customers go live

Three steps: provision the deployment you need, ingest your documents, point your AI client at the MCP server that comes out the other side. An afternoon for the first one, usually less for the next.

Onboard

You pick a deployment mode — managed cloud, self-hosted, or air-gapped — and the stack comes up accordingly. The cloud path provisions a per-tenant project. The self-hosted path is a Docker Compose file you run on your own infrastructure. Both ship the same code.

A short call covers your document landscape — file types, volumes, who touches what — and the access model you need. Most teams go live in an afternoon, not a quarter.

Compliance review fits in here too. If your legal team needs to see the encryption story, the deployment topology, or the data flow diagram before approving anything, that conversation happens first — not as an awkward follow-up after you've already started.

Included in this phase

  • Deployment mode selection
  • Tenant provisioning
  • Document landscape review
  • Access model setup
  • Compliance walk-through

Ingest

You create a bucket for each knowledge domain you want to expose — a policy library, a claims handbook, a product catalogue, a year of board minutes — and drop in the source documents. Today we support PDF, DOCX, CSV, JSON, plain text, Markdown, and Confluence exports.

Behind the scenes Laminae handles chunking, embedding, sparse indexing, and storage. Each chunk keeps a hard reference back to its source document, page, and position, because provenance isn't something you can bolt on later.

Connectors handle the documents that don't live as files — Confluence, Github and Gitlab today, more on the way. There's no streaming pipeline; ingestion is batch and explicit, which is the right trade-off for a knowledge layer that has to be auditable.

We build retrieval the boring, careful way — explicit ingestion, trackable provenance, real control. The result is a system you can trust, and put in front of a legal team without sweating.

David Bunting, Founder, Laminae

Connect

Now we expose as many MCP servers as you need on an open standard. Point Claude Desktop, Cursor, Continue, or your internal Copilot at them and the AI client you already use becomes smarter on your own data.

When the client asks a question, Laminae runs hybrid retrieval — dense plus sparse plus re-rank — and returns the matching chunks with their source attribution attached. The client generates the answer; the human reading it can always verify it against the original document.

Once it's wired up, the day-to-day for the user looks like their existing AI client, just suddenly informed about the company they actually work at.

Included in this phase

  • MCP endpoints. Each bucket gets a stable MCP URL with per-tenant auth. Any MCP-compatible client connects.
  • Hybrid retrieval. Dense vectors plus sparse search plus a re-ranker. Source-attributed chunks come back; the AI client generates the answer.
  • Observability. Per-bucket query logs, retrieval traces, and re-ranker scores — enough to debug a bad answer back to the chunk that produced it.

Deployment - Managed cloud or self-hosted — same product, your call

The cloud version is the same Docker Compose stack we ship for self-hosted. Air-gapped is the same again.

Managed cloud

Per-tenant Hetzner project, provisioned by us, billed as a bundled fee. No shared infrastructure between customers. Right for teams who want the product without the operational overhead.

  • Per-tenant isolation; no shared databases or indexes
  • EU hosting by default
  • Encryption at rest, BYOK available on request
  • Cloud-hosted model endpoints (OpenAI, Anthropic, etc.)
  • Bundled Ollama

Self-hosted & air-gapped

The same Docker Compose stack, run on your own infrastructure. Air-gapped variant ships with bundled Ollama, so nothing leaves your network. Right for teams with strict data residency or no-egress requirements.

  • Your hardware, your keys, your network
  • Bundled Ollama
  • No telemetry from the product
  • Same code path as managed cloud

Constraints - What this product won’t do

Knowing what we don't build is more useful than another list of features. If your problem lives in one of these boxes, Laminae isn't the right tool.

  • Not a chat UI. Laminae returns evidence; your AI client generates the answer.
  • Not a RAG wrapper. The wedge is hybrid retrieval plus MCP plus self-host, not a thin layer over `top-k` vector search.
  • Not fine-tuning. We don't train models and we don't use your data to train ours. Retrieval over your documents, full stop.
  • Not real-time streaming. Ingestion is batch and explicit. File uploads and connector sync, on your schedule.
  • Not enterprise-only. Our product is there to scale the capability of each employee, whether there are 5 or 5000.
  • Not generic. One product, three deployment modes. We don't customise the wedge per customer; we customise the deployment.

Want to see Laminae on your own documents?

Based and hosted in

  • Frankfurt
    Frankfurt am Main
    Germany