Source map and access matrix
An inventory of the sources in scope, real volumes, formats, duplicates, and a reading of your existing permissions. This document is yours whatever you decide next.
AI-assisted document search
Your company’s ChatGPT, hosted in Switzerland, connected to your documents
Clairdex indexes your specifications, test reports, internal standards and site minutes, then answers a business question by citing the document and the page. The corpus stays on Swiss infrastructure and is never used to train a model.
±0.2 mm, adopted in revision C after the load test of 14 March. Revision B specified ±0.5 mm.
Cited sources
Select a corpus and carry over your access rights
Ask the question in your own working language
Check the citation, then accept or flag it
Who it is forEngineering firms, industrial companies and technical teams of 30 to 500 people.
In 30 seconds
You ask a question the way you would ask a colleague. Clairdex finds the useful passages in your documents and answers with the source cited.
File server, document system, mailboxes: they stay where they are.
Indexes meaning, respects your access rights, hosted in Switzerland.
The document, the page and the passage, to verify in one click.
A model running on infrastructure you control. Your documents are never used to train it.
Every statement points to the document and the page. Without a citation, an answer is only an opinion.
What a user cannot open, they cannot find through search either.
Six to eight weeks on a narrow scope, with written criteria, before any rollout.
An engineering practice accumulates twenty years of files: contracts, amendments, calculation notes, test certificates, supplier datasheets, successive revisions of the same standard. The information exists. It is simply spread across a file server, a document system bought in 2014, the mailboxes of three project managers, and a shared drive nobody dares reorganise.
The result: the question “what tolerance did we settle on for this type of joint, and on which project” gets answered by asking a colleague rather than a system. When that colleague leaves, the answer leaves with them.
Full-text search does not close the gap, because the business question almost never uses the document's exact words. Clairdex works on meaning, then always points back to the source so the answer stays verifiable.
What you receive
An inventory of the sources in scope, real volumes, formats, duplicates, and a reading of your existing permissions. This document is yours whatever you decide next.
Connection to the approved sources, text extraction including scanned PDFs, chunking and indexing, with scheduled incremental reindexing.
One question, one answer, and under every statement the document, the page and the exact passage. Filters by project, by date and by document type.
Thirty to fifty real business questions with their expected answers, replayable on every change. Two hands-on sessions per user group.
What gets measured
Time to complete a comparable set of tasks, measured before and after, with the same people.
The share of correct answers, and the share of citations that genuinely point at the useful passage.
The share of questions where the system admits it does not know. A zero rate is a bad sign, not a good one.
Estimate
Three numbers are enough to size the issue and tell whether a pilot makes sense for you.
Indicative estimate from your inputs, without commitment. The assessment measures reality on your documents.
Pricing
Indicative ranges, excluding applicable taxes. They are not an order: scope and exclusions are set in the quotation.
01
3'000 – 5'000 CHF
Two to three weeks. Source map, reference questions, an honest estimate.
Request an assessment02
8'000 – 15'000 CHF
Six to eight weeks on a narrow corpus, with real users and written criteria.
Request an assessment03
20'000 – 60'000 CHF
Then per month 2'000 – 6'000 CHF
Rollout across the contractual scope, monitoring and quarterly review.
Request an assessmentAn assessment may conclude that a well-tuned conventional search is enough. That is a planned outcome, and it leaves you the source map and the question set.
Answers may contain errors and remain subject to human validation. Connectors, volumes, models and compute costs are fixed in the quotation. No training on your data is planned without a separate written agreement.
FAQ
No. The pilot starts from your existing sources and leaves them in place. The connectors needed are chosen after reviewing available APIs, licences and permissions.
The system must report an insufficient source rather than produce a plausible answer. The acceptable threshold is agreed before rollout and verified by the evaluation set.
No. The target architecture keeps the corpus on the agreed infrastructure and sends the model only the passages needed for the answer in progress. Any other use would require a separate written agreement.
A useful pilot starts around 5,000 documents and stays comfortable up to 50,000. Below that, a well-configured conventional search is often enough, and we will say so.
Source permissions are carried over at indexing time: a user cannot obtain a citation from a document they are not allowed to open. This is part of the acceptance tests.
Request an assessment
Describe your document volumes, the systems they sit in, and the business question that costs you the most time. We reply within two working days, in English, French or German.