Setting up Seba for a Gallery

A short guide for museum directors and curatorial staff on using Seba for a new gallery. All exhibits are added and updated using the Exhibit Management tool, described in the Exhibit Management User Guide. The demo tool linked here works on a separate sandbox copy that is reset every day at 2 a.m. Eastern time, so feel free to try anything; changes won’t affect the live app.

1. The Content Seba Uses

Everything Seba says about the collection comes from a handful of plain-text descriptions, with no coding involved. Seba isn’t “trained” on this material; it reads it each time it answers, so an edit takes effect immediately. This is exactly the kind of text curatorial staff already write, review, and update, and it is the single biggest factor in a good visitor experience. Each exhibit also needs a few reference photos so Seba can recognize the exhibit a visitor is looking at.

• Exhibit Description — one plain-text description per exhibit, pasted into the Exhibit Management tool. It can be written in English, German, French, or Spanish; either way, Seba speaks to each visitor in their preferred language.
• Exhibit Photos — add catalog images, or take photos with the Exhibit Management tool on a mobile device. Photos taken in the gallery itself, from the angles and in the lighting visitors will see, give the most reliable recognition. A few from different angles is better than one. The Analytics section below explains how to spot exhibits that need more.
• Venue Theme — an overview of the gallery or collection as a whole, roughly 200–300 words. This is the general context Seba keeps in mind for every conversation, whichever exhibit a visitor is looking at: the institution’s character, the scope and strengths of the collection, and any historical background worth knowing. It is written once per venue and rarely needs to change.
• Venue Header — the first thing a visitor sees when they open the app, about 10–25 words, in any language. A brief welcome plus a reminder to photograph an exhibit to get started.
• Venue Options — a few venue-wide settings, also set in the Exhibit Management tool:
  • Knowledge Scope: LIMITED or GLOBAL — LIMITED (the default) restricts answers to the museum’s own exhibit descriptions. GLOBAL also allows generally known information beyond them. Use GLOBAL with care, since answers then go beyond text the museum has reviewed.
  • Duration: 15, 25, or 40 — the default length, in seconds, of Seba’s spoken answers. Visitors can still adjust this themselves in the app.
  • Silent Mode — turns Silent Mode on for every visitor, for galleries where quiet is expected. It hides the voice button, so visitors can’t ask questions aloud; they can still tap the suggested questions or type their own. Answers are still spoken, so visitors should use headphones.
  • Matching Thresholds — how confident Seba must be before it names an exhibit. The defaults suit most galleries, and a “Restore values” button returns to them after any experimenting.

2. Mobile Devices, Tablets & Headphones

Seba runs in a web browser, so most visitors can simply use their own mobile device, with no app to install. A few practical choices still make a real difference on the gallery floor.

Most mobile devices and browsers are supported. A full breakdown, including how visitors allow camera, microphone, and sound on their own device, is in the companion Visitor Guide, which can be handed to visitors directly. In general, Android devices running Chrome give the smoothest experience. Apple devices work well but need a few extra taps.

Headphones are worth encouraging; even a simple, low-cost pair makes a noticeable difference. They keep the gallery quiet instead of filling it with overlapping speakers, give clearer audio in echoey, high-ceilinged, or busy rooms, and make the experience feel more personal, closer to a private docent than a public announcement. Headphones also pair naturally with Silent Mode: a visitor can listen privately in a gallery where speaking aloud isn’t appropriate, and ask follow-up questions by tapping or typing.

Consider lending tablets and headphones. If budget allows, a small fleet of reasonably priced Android tablets running Chrome, set up in advance with camera and microphone access already allowed, removes all setup for visitors who would rather not use their own device, or whose device doesn’t support every feature. Older, low-powered devices tend to be slow and can spoil the experience, so choose reasonably current hardware, even if it doesn’t need to be top of the line.

3. Operating Costs

The Seba software itself is free. Its running costs come from the two AI services that power it, both billed by usage.

1. Anthropic’s Claude writes Seba’s introductions and answers, at a small cost each time. Two Claude models, Sonnet and Haiku, are combined to keep both cost and response time down. As an example, consider a visitor who views 15 exhibits and asks 3 follow-up questions about each. Based on measured usage during development, a Sonnet response costs about $0.0069 and a Haiku response about one-fifth of that, so this visitor’s whole visit costs about $0.17. For 100 such visitors a day, that is about $16.50 a day, or roughly $6,000 a year. Actual costs depend on how many exhibits visitors view and how many questions they ask.
2. Voyage AI matches each visitor’s photo to the exhibit’s reference photos, at roughly $0.00003–$0.0012 per photo. With the same 100 visitors a day each photographing 15 exhibits, that is under $700 a year even at the top of that range.

There is always a trade-off between price and performance. Several free AI services are available but have not yet been tested; their reliability, restrictions, and answer quality would need checking first. Speech recognition, which turns a visitor’s spoken question into text, could also be improved with a paid service.

4. Analytics

Seba keeps simple usage statistics so staff can see how the guide is being used and where it needs attention. It does this without collecting personal information: there are no visitor accounts or sign-ins. Each visitor’s browser is given an anonymous random ID, which the “New Visitor” button clears, and every request is recorded on the museum’s own server.

From these records, the analytics report shows:

• Visitors and visits — how many people used Seba and how many separate visits they made. A gap of more than six hours starts a new visit.
• Exhibit views — which exhibits visitors look at most and least.
• Time spent — the average time visitors spend with each exhibit, and the average length of a visit. Long idle gaps, over 15 minutes, aren’t counted, so a visitor who wanders off doesn’t inflate the figures.
• Recognition problems — each time Seba couldn’t identify a photo, and why. The two main reasons point to different fixes:
  • Not confident enough — the photo didn’t closely match any exhibit. This often means the exhibit needs more reference photos, especially from the angles and lighting visitors actually see.
  • Too close to call — the photo matched two exhibits almost equally well. This is common with similar objects displayed together; reference photos that show what makes each one distinct usually help.

Together these turn upkeep into a short routine: check which exhibits cause recognition problems and add reference photos for them; see which exhibits draw the most interest and time, and give those descriptions extra care. Popular exhibits with short viewing times may simply need a more engaging description.

One limitation for now: staff testing the app is counted along with real visitors. Separating staff activity is planned. In the meantime, keep testing to a minimum during periods you want to measure, or note when it happened.

5. An AI-Powered Museum Website

One more thing worth considering, once Seba is running well on the gallery floor: the same approach could extend naturally to the museum’s public website.

Many museum websites fall into one of two unsatisfying extremes for exhibit information: a single thin paragraph that leaves a curious visitor wanting more, or a dense scholarly PDF that only a specialist would read from start to finish. Either way, the page is static. It says the same thing to every visitor, and there is no way to ask it the one question you actually have.

Imagine instead a visitor browsing the collection from home, tapping a photo of an exhibit on their screen, and simply starting a conversation about it: the same guide, with the same depth adjusted for a curious teenager or a specialist, available to anyone researching the collection at any hour, not only to people standing in the gallery.

The appealing part is that this wouldn’t mean starting over. The content built for Section 1 — the venue overview and the per-exhibit descriptions — is exactly what would power it. The same investment would do double duty.