Standards, manuals, contracts, audit findings, product documentation: the knowledge your organisation runs on sits in documents nobody can search well. A private assistant answers questions from those documents, shows the passage it used, and runs on infrastructure you control. This page sets out how we build one, what it costs, and the questions an RFP should ask.
By Tom Joseph · Last updated: 8 September 2026
A private AI assistant is a retrieval system with a language model on top, hosted where your data already lives, that answers only from your documents and shows its source every time. The technical term is RAG, retrieval-augmented generation. What makes it private is not the model but the architecture: the documents are indexed in a vector store inside your Azure or AWS tenant or on your own server, the model sees only the passages retrieved for one question, and nothing is used to train anything. Crux Digits builds these at a fixed price: a €2,500 audit of your documents and access rules, a €20,000 Production-ready MVP on your real corpus with a test set, production from €50,000 with roles, logging and handover. Delivery is measured in weeks.
How it is built
Five parts, each of which decides whether the assistant is trusted or ignored.
PDFs, Word, SharePoint, the standards archive, the audit findings. Read, split, dated, and re-read when a version changes.
pgvector, Azure AI Search or OpenSearch, in your tenant or on your server. Nothing leaves.
Hybrid search: keyword and meaning together, filtered by document, date and audience.
The answer on the left, the exact passage it came from on the right. No passage, no answer.
Who may see what, what was asked, what was answered, and a test set that runs on every change.
The requests we receive read almost identically. A standards body with 1,400 audits a year and a library of normative documents that auditors and farms cannot navigate. An agri-tech firm with a technical blueprint and domain experts ready to supply the corpus and the test cases. A publisher whose editorial archive is the product. A manufacturer whose maintenance knowledge is in twenty years of PDFs. What they share is that the value is in the documents, the documents are too many to read, and the answer has to be checkable, because someone will act on it.
They also share three requirements that a public chatbot cannot meet: data sovereignty (the corpus and the questions stay in the EU, in their tenant), a hard citation for every answer (an auditor cannot cite "the AI said so"), and a fixed price, because a knowledge project with an open-ended budget is the one most boards have already been burned by.
The demand is measurable. Statistics Netherlands puts text mining, the reading of documents by software, at 9.8 percent of Dutch micro-enterprises in 2025, up from 3.1 percent in 2023, the fastest growth of the seven AI technologies it tracks. Reading documents is the AI use Dutch firms actually adopt; this page is about doing it with a source attached.
The same five parts every time, chosen per project. The choices below are the ones we make by default for a Dutch organisation with a few thousand to a few hundred thousand pages.
| Part | Default choice | What decides it |
|---|---|---|
| Ingestion | Connectors to SharePoint, Google Drive, a file share or an S3 bucket; parsing for PDF, Word, HTML and scanned pages (OCR); chunking by section, not by character count | Where the documents live, how often they change, whether tables and figures must survive |
| Vector store | pgvector on PostgreSQL in your tenant; Azure AI Search or OpenSearch where the platform is already standard | Existing platform, expected corpus size, who operates the database |
| Retrieval | Hybrid: BM25 keyword search plus dense vectors, re-ranked; filters on document, version, date and audience | Technical corpora with exact terms (article numbers, part codes) need keyword search; narrative corpora need semantic search; most need both |
| Model | A European-hosted frontier model through Azure OpenAI or AWS Bedrock in an EU region; an open-weight model (Mistral, Llama) on your own GPU where nothing may leave the building | Data-residency policy, language quality in Dutch, cost per question |
| Interface | Two columns: the answer with numbered citations, the cited passages beside it, a link to the source document at the page. Also an API for your own applications | Whether the assistant is used by people, by other software, or both |
| Governance | Roles inherited from the source system, a log of every question and answer, a test set of 50 to 200 question-answer pairs that runs on every change, a "no source, no answer" rule | EU AI Act classification, audit requirements, who is accountable for a wrong answer |
Three options, and the choice is a policy decision before it is a technical one.
The proof of concept is not a demo. It runs on your corpus, against a test set your domain experts write, and it produces a number.
Before we build, your experts write 50 to 200 questions with the answer and the passage that supports it. That set is the acceptance criterion. The proof of concept reports what fraction it answers correctly with the right citation, what fraction it correctly refuses, and where it fails, by document type and question type. Production is priced against that report, and the same set runs on every change afterwards, so a model upgrade or a new batch of documents cannot silently make it worse. The evaluation method is the same one we use for LLM optimisation on assistants other people built.
Fixed, per step, excluding VAT and excluding your cloud consumption (which we estimate in the audit and which is usually modest: retrieval is cheap, and the model is called once per question).
For the engineers on your side of the table:
A question-answering system that retrieves passages from your own documents and has a language model compose an answer from those passages only, with the source shown. It runs in your cloud tenant or on your servers, and neither the documents nor the questions are used to train any model.
RAG (retrieval-augmented generation) is the technique. A private AI assistant is RAG delivered as a system: ingestion, an index you own, retrieval with filters, an interface with citations, roles and logging. The technique is a week; the system is the project.
Yes, with an open-weight model on your own GPU. Dutch-language quality is lower than a frontier model and the proof of concept measures the gap on your own test set, so the decision is made on numbers rather than policy alone.
That depends on the corpus and is exactly what the proof of concept reports: the share of test questions answered correctly with the right citation, the share correctly refused, and the failure modes. Well-structured technical corpora routinely score high; scanned, inconsistent archives score lower until ingestion is tuned.
Every chunk carries its document, version and date, so a question can be answered from the current version only or from a named one. For multilingual corpora the index holds each language and the assistant answers in the language of the question, citing the original.
Mostly cloud consumption: a small database, a small application server, and model calls priced per question. For a department-scale assistant that is typically tens to a few hundred euros a month; the audit estimates it for your volume. Maintenance is a separate fixed agreement if you want us to keep it current.
Send the RFP if there is one. The €2,500 audit returns the architecture for your case, a hosting recommendation and a fixed price for a proof of concept on your real documents.
Book a conversation