Language:DeutschEnglish Last updated: 2026-08-19

AI-assisted cataloguing¶

Since v0.63.0, Anton can generate suggestions for titles, text fields, dating, actors, places and keywords at the push of a button — based on the attached media files (images, PDFs, audio transcripts, video frames) and the existing context of a record.

By default the function is switched off. It is activated separately per archive. The default provider is Infomaniak (hosted in Switzerland, FADP/GDPR-compliant), so that archival data does not leave Switzerland.

Activation per tenant¶

Three switches all have to be set for AI-assisted cataloguing to be available:

Env variable AI_ENABLED=true in the tenant's .env file (global kill switch via deployment)
Provider keys in the env:
INFOMANIAK_API_KEY + INFOMANIAK_TENANT_ID (default profile)
optionally ANTHROPIC_API_KEY, GEMINI_API_KEY
Setting ai_enabled = true in the tenant settings UI

Without these three steps: no button, no dashboard links, no risk of provider errors. Tenants without an AI contract notice nothing.

The operator workflow¶

When editing a record, a button "🤖 Generate AI suggestions" appears at the top of the action bar (since v0.65.0 directly next to the upload area, visible as soon as a file is attached).

After clicking:

Anton uploads the attached media to the configured provider (PDFs are rendered as page images; audio is transcribed via Whisper; videos are split into 5 evenly distributed stills)
Within a few seconds, the provider delivers a structured suggestion
Anton shows a hint chip with the AI suggestion for each field
Three actions per suggestion: apply, append (text fields only), ignore

Saving takes place as usual via the Anton save button. No hidden database writes — the AI only makes suggestions, the human decides.

What the AI receives as input¶

Images (photographs, scans) directly as vision input
PDFs — page images; additionally the OCR text layer if present (since v0.65.0, for better quality with typed documents)
Audio — Whisper V3 transcript via the Infomaniak async batch API. The transcript can optionally end up in the full-text index as a text field.
Video — 5 ffmpeg stills, evenly spaced across the running time
The existing title of the record as context

Actors, places, keywords: two patterns¶

The AI distinguishes between creator relationships and content relationships:

Pattern	Example	Where it is stored in Anton
Creator / place of recording	Photographer of a shot, filming location	New `AntonEvent`
Named or depicted in the material	People in a photograph, places mentioned in the text	Content descriptor (keyword/person/place list)

For each suggestion it is apparent whether an entry already exists (in which case it is linked) or would have to be newly created (a click opens the prefilled creation form). Keywords without further details can be created and linked directly with one click.

Admin dashboard¶

Under Admin → AI cataloguing there are five tabs:

Tab	Content
Consumption	Monthly budget, cost status, last 50 calls per record with token counts and CHF costs
Models	Available providers/models with description, vision capability, active status, prices per 1k tokens
Budget	Monthly caps per calendar month (since v0.65.0 also editable via the UI, previously SQL only)
Profiles	Defined AI profiles with create/edit/delete
Audit trail	All calls with prompt, AI response and operator decision (applied/changed/ignored)

Since v0.65.0 the audit tab shows not only "suggestion generated" but also whether it was applied, changed or ignored — important for quality control and for later prompt tuning.

AI profiles¶

A separate profile with an adapted prompt block can be created per tenant or object form. Example for personal files:

"You are analysing personal files of Swiss public figures. Pay particular attention to dates of birth and death, professions, family relationships and places of residence."

The AI receives this block on every call in addition to the standard schema instruction.

The profile creation form offers:

Provider selection — Infomaniak (Switzerland), Anthropic (USA), Google Gemini (USA/EU)
Model selection — suitable models per provider, with description
Scope — global, per tenant, per object form, per fonds
Language — German by default
Audit threshold — 100 = every call is sampled, 0 = never

Profiles without an API key in the env are visible in the creation form (with a warning) but do not appear in the operator profile selection — users only see profiles that work.

Since v0.65.0 a default profile per tenant is preconfigured, so that the profile selection in the operator workflow only unfolds on request.

Costs and budgets¶

Default models: Infomaniak Gemma 4 31B-it (multimodal, 256k context) + Mistral Ministral 3 14B
Optional fallback: Anthropic Claude Sonnet 4.6 (vision), Google Gemini 2.5 — with a data protection warning, because data is processed outside Switzerland
The cost per call is recorded in the table ai_usage_log and totalled in the dashboard
Default monthly budget for the pilot: 50 CHF. If the cap is exceeded, Anton blocks further calls until the turn of the month.
The margin on the provider costs (default 1.30, that is, a 30 % surcharge) is set installation-wide in config/ai.php and can be overridden in the env via AI_BILLING_MARGIN

Data protection¶

By default the system uses models hosted in Switzerland (Infomaniak). FADP/GDPR-compliant. If a US/EU provider (Anthropic, Gemini) is explicitly selected, a data protection warning appears in the operator UI pointing out the data flow outside Switzerland.

Technical key facts¶

6 new database tables: ai_profiles, ai_profile_versions, ai_usage_log, ai_pricing, ai_budgets, ai_audit_samples
Cross-repo package: kraenzle-ritter/ai-cataloging (framework-free)
Profile versioning: every edit of a profile creates a snapshot, so that audit trails remain reproducible
JS chip decorator instead of a modal stack: suggestions appear directly at the form fields rather than in a separate panel
Auto-save on upload (since v0.65.0): drag-and-drop uploads save the record automatically and keep the editing page and scroll position