Enterprise RAG in Switzerland is rarely blocked by model quality — it is blocked by jurisdiction, access control and the unglamorous work of getting documents in cleanly. This guide covers the four decisions that determine whether a deployment succeeds: build or buy, deployment model, scope of the first corpus, and how you will know it works.
Build or buy
A demo is a weekend. A production system is chunking that respects document structure, OCR for the scanned half of the archive, multilingual retrieval, per-document access control, deletion that cascades, audit logs, evaluation, and someone on call. Teams that build in-house consistently underestimate the last six.
Building makes sense when retrieval is your product, or when a genuinely unusual corpus needs custom handling. Buying makes sense when the corpus is ordinary enterprise material and the differentiator is what your people do with the answers. The middle path — buy the platform, own the corpus, the ontology and the evaluation set — is where most Swiss deployments land.
Deployment models
| Model | Fits | Trade-off |
|---|---|---|
| Public AI assistant (no RAG) | Public material only | No corpus, no citations, no access control |
| Foreign cloud RAG | Non-sensitive corpora | Cross-border disclosure to assess and defend |
| Swiss multi-tenant SaaS | Most Swiss enterprises | Shared infrastructure; requires per-tenant encryption and strict isolation |
| Dedicated single-tenant Swiss | Regulated institutions, supervisor pressure | Higher cost, still fully managed |
| On-premise | Air-gapped or classified environments | You own GPUs, upgrades, availability and the AI roadmap |
For most Swiss organisations the honest answer is a Swiss RAG platform as managed SaaS with per-tenant encryption, moving to dedicated infrastructure only where a supervisor or a client contract requires physical separation. On-premise is a real answer for a small number of environments and an expensive habit everywhere else.
What it costs
Budget in three lines, not one. Licence scales with users and corpus size and is the predictable part. Ingestion is a one-off that scales with document count, scan quality and format variety — a clean DMS export is cheap, twenty years of scanned PDFs on a shared drive is not. Internal effort is the line people forget: someone must decide what goes in, who may see it, and what “a good answer” means. Reserve one informed part-time owner for the first quarter.
As a reference point, Swiss platform pricing for a managed sovereign deployment starts in the low hundreds of francs per month for a team and moves into four figures for organisation-wide rollouts with dedicated infrastructure.
A realistic timeline
- Weeks 1–2 — scope. One department, one question type, a corpus you can name. Resist “index everything”: it inflates cost, dilutes retrieval quality and turns access control into a project of its own.
- Weeks 2–4 — ingest and access model. Load the corpus, mirror the existing permission model per document, fix what OCR mangles.
- Weeks 4–6 — evaluation. Write 30–50 real questions with known answers, from the people who will use it. This set, not a vendor benchmark, is what tells you whether the system works on your documents.
- Weeks 6–10 — pilot with real users. Watch the query log for the questions that fail; most fixes are corpus or permission fixes, not model fixes.
- Then scale by corpus, not by headcount. Each new document set gets the same scope-ingest-evaluate cycle.
Where Swiss deployments fail
- No evaluation set. Without one, “it feels worse since last week” is unanswerable and every change is a guess.
- Access control retrofitted. Permissions must be part of ingestion. Adding them after go-live means re-ingesting, and until then any user can surface anything.
- The scanned half of the archive. If OCR quality is poor, retrieval quality is poor, and no model compensates for text that was never extracted.
- Language assumptions. Swiss corpora are mixed German, French, Italian and English. A pipeline tested only in English degrades quietly — see multilingual RAG for Swiss documents.
- Sovereignty decided last. Discovering in the security review that embeddings go to a foreign API restarts the procurement. Settle the data path in week one with the twelve-question checklist.
What to ask for in a pilot
Your own documents, your own questions, on the production data path — not a sandbox in a different jurisdiction. A pilot that cannot use real documents because of where it runs has already told you the answer to the sovereignty question. See how this maps to Swiss industries, or read RAG vs. fine-tuning if the internal debate is still about training a model on your archive.