From brochure to dialogue: how AI learns your museum in five minutes

01The problem RAG solves

Conversational artificial intelligences, however capable, do not know your museum. They know general things about the Louvre or MoMA because that information is abundant online. But they know nothing about your pricing policy, your special hours during school holidays, the exact composition of your temporary exhibition or the booking terms for school groups. This information is specific to your institution. It lives in your brochures, your catalogues, your internal documents.

RAG solves this problem by giving the AI direct access to your documents. Instead of guessing or fabricating a response, the agent first searches your knowledge base, finds the relevant passages, then formulates an answer grounded in verifiable information. It is the difference between an agent that improvises and an agent that knows.

02The four-step process

The first step is upload. You access the AI ARTEDUSA dashboard, with no technical skill required, and upload your documents. Accepted formats are PDF, Word and plain text. The maximum size is ten megabytes per file. You can upload your institutional brochure, your exhibition catalogue, your pricing schedule, your internal FAQ, your visitor information sheet.

The second step is chunking. The system does not treat the document as a single block. It breaks it into intelligent segments of four hundred tokens each, with an overlap of fifty tokens between consecutive segments. This overlap ensures no information is lost at the boundary between two segments. If a sentence begins in one segment and ends in the next, it appears in both. The system also performs automatic deduplication to eliminate duplicates.

The third step is vectorisation. Each segment is transformed into a mathematical vector of seven hundred and sixty-eight dimensions using our vectorisation engine. This vector captures the meaning of the text, not just the words. The segment "open Tuesday to Sunday" and the question "which days are you closed?" produce similar vectors even though the words differ, because the meaning is related. These vectors are stored in a vector database dedicated to your institution, completely isolated from those of other institutions through the multi-tenant architecture.

The fourth step is retrieval. When a visitor asks a question over the phone, the agent transforms that question into a vector and searches for the five most relevant segments in your database. The minimum relevance score is set at 0.8 out of 1.0. If a segment does not reach that threshold, it is not included. The agent only responds with sufficiently reliable information.

03What this looks like in practice

Consider four concrete examples drawn from real use cases. A museum uploads its brochure in PDF format. A visitor calls and asks about opening hours. The agent searches the segments, finds the relevant passage and answers: "We are open Tuesday to Sunday, from 10 a.m. to 6 p.m. Late opening on Thursdays until 9 p.m."

The same museum uploads the catalogue of its temporary exhibition. A visitor asks about the Monet exhibition. The agent searches the catalogue segments and answers: "The Monet and Light exhibition features 47 works," adding details found in the document.

A third document, an internal FAQ in Word format, is uploaded. A visitor asks whether there is a car park. The agent finds the corresponding segment and answers: "Yes, an underground car park with 200 spaces is accessible at 12 rue..." giving the exact address from the document.

A pricing schedule in text format is uploaded. A visitor asks about the rate for a school group. The agent finds the segment and answers: "The school rate is 5 euros per pupil, free for accompanying adults."

04Automatic contextual injection

The system goes beyond on-demand search. At the start of every call, the three most relevant documents are automatically injected into the AI's context. This means the agent "already knows" before the visitor even asks their first question. It knows the day's opening hours, the current rates, the exhibition on display. This contextual injection reduces response time and improves the relevance of early answers in the conversation.

05Why five minutes and not five hours

Setup time is a legitimate concern for a museum director. The answer is that registration takes two minutes and the agent is operational within five minutes of the document upload. There is no training to programme, no scripts to write, no developer to hire. The dashboard is no-code: you upload and you are done. Advanced configuration, which includes custom schedules, integration with an external booking system via webhook, and pricing parameters, can take a few hours. But the agent answers questions correctly from the very first minutes.

06Semantic search versus keyword search

What sets AI ARTEDUSA's RAG system apart from a simple search engine is its understanding of meaning. A visitor who asks "what time do you close on Sundays?" will not find an answer in a keyword system if the document says "open Tuesday to Sunday, from 10 a.m. to 6 p.m." The word "close" does not appear in the text. But the semantic system understands that "what time do you close" and "open until 6 p.m." are related. It retrieves the relevant segment, and the agent formulates the answer in natural language.

This semantic capability relies on the our vectorisation engine model, which transforms each text into a vector of seven hundred and sixty-eight dimensions. The distance between two vectors measures the semantic proximity between two texts. The smaller the distance, the closer the texts are in meaning.

07Document data security

Uploaded documents belong to each institution. The multi-tenant architecture guarantees complete isolation at the database level through our secure database per-institution data isolation. Vector collections are separated by institution in our vector database. No institution can access another's documents. Data is hosted in the European Union and the system is GDPR-compliant.

Deduplication by our verification algorithm hash prevents accidental upload of the same document twice. If you upload an updated version of your brochure, the system detects identical segments and retains only the new ones.

08Keeping the agent's knowledge up to date

Museum information changes regularly. A temporary exhibition closes and a new one opens. Hours vary by season. Prices are revised. To update the agent's knowledge, you simply upload the new document. The system processes the new segments, vectorises them and makes them available for retrieval. The agent automatically uses the most recent information.

This simplicity of updating is a decisive advantage over a human receptionist, who must be informed of every change, remember the information and relay it correctly. The agent always consults the most recent source.

AI ARTEDUSA was built precisely to meet this need. Find out how at ai.artedusa.com.