AI in Moliri

AI account

The AI account tells Moliri which language model it can ask. You set it up under Settings, section Data, AI account. Only administrators (status 9) can change it.

AI account settings page: interface OpenAI-compatible, endpoint, model, key, embedding model, endpoint for embeddings and model for text recognition, under Does the data leave the house? Automatic (stays in the house) for each account, below them Save and Test connection, the result of the test with the test image read as Moliri 42, the note on what goes to the model and the link to the AI overview.

Settings

SettingMeaning
InterfaceOpenAI-compatible (OpenAI, Mistral or a server of your own with Ollama or vLLM), Anthropic (Claude) or Google Gemini
EndpointWith OpenAI-compatible, the address of the interface, ending in /v1. With the others, leave it empty for the provider’s address
ModelName of the model at the provider or on your own server
KeyThe provider’s access key. A server of your own usually needs none. A saved key stays on the server; leaving the field empty keeps it
Embedding model, Endpoint for embeddingsOnly for the AI run under Similar values on the Check tab
Model for text recognitionA model that understands images, through the same interface and with the same key. It reads for text recognition by the AI service : in the OCR import, with Recognise text on the Text tab and in rules. Without this model, the AI service does not appear anywhere in text recognition
Does the data leave the house?For each account (chat model, embedding model, image model for text recognition) Automatic, Runs in the house or External. Automatic derives it from the address and gives the result in brackets; behind a proxy that is not always right. You can see it in the AI overview

Test connection and save

Test connection asks the model with the details in the form, even unsaved, and gives the response time or the provider’s message. If a model for text recognition is entered, it reads a small test image with the words “Moliri 42”; the page shows what it has read, for example Text recognition answers (4.3 s) and reads on the test image: “Moliri 42”. Save stores them. Remove key deletes the key, Use the server’s model deletes the whole account. The AI overview in the profile offers the same test for a single saved account.

Without an entry, the model that the server brings through its environment applies (MOLIRI_AI_PROVIDER, MOLIRI_AI_URL, MOLIRI_AI_MODEL, MOLIRI_AI_KEY, for text recognition MOLIRI_AI_OCR_MODEL). The page then says The server’s model applies now. An account saved here takes precedence.

Who uses the account

FunctionModelGoes to the model
Suggest fields with AI in the OCR importChat modelThe sections read, the field names and up to three existing values per field
AI run under Similar values Embedding model and chat modelThe values of the chosen fields
Dimensions from free text Chat modelThe text with the dimensions
Text recognition by the AI service Model for text recognitionThe image

The bottom of the page also says what goes to the model. If the model runs on your own server, nothing leaves the building. If you do not want to hand the images to a provider, switch off text recognition by the AI service on the Text recognition page.

The questions to the model run as jobs on the server, one after the other with transcription and text recognition; see AI jobs and nightly runs .

Further reading