mmm
Read a whole archive with language models
mmm (mini multi modal) turns a folder of mixed files into one structured description per file. Texts, office documents, PDFs, images, sound and video are read by a language model, local through Ollama or Claude, with your own prompts and your own answer structure. Its studio connects the corpus to Arvest without uploading it.
Ready-to-install releases for macOS, Windows and Linux are on their way. Until then, the code and its instructions are on GitHub.
Every file, read and described
mmm works out what each file is, prepares it for a model (extracted text, PDF pages, video frames with a transcription of the speech), asks your questions and checks every answer against your structure. Long runs can be stopped and picked up again.
Your corpus in Arvest, without uploading it
While the studio runs, it serves the corpus to Arvest as IIIF documents from your own computer. They appear in your documents, viewers and workspaces like any other, with mmm's answers as their data; the files never leave your machine.
Answers become data
The answers become dimensions to sort, filter and plot: the corpus as points, in clusters, each opening its file. Send them to Arvest as metadata, or bring documents from your Arvest account into the corpus.
Tutorial
The tutorial video is on its way.