RAM

New release. Shared from the Koher page:

Tool release · RAM 0.5 · 15 October 2026 · open source

RAM (It’s RAM; not Ram) reads a short piece of office writing, such as a file noting, and finds each thing somebody has to do in it. For each act it says who does it, as “ karega”, Hindi for “ will do it”, or “RAM bharose” when the writing leaves the act to nobody. “The Registrar approved the proposal” gets “Registrar karega”. “Necessary action may be taken” gets “RAM bharose”, and so does “The competent authority has approved it”, because the competent authority names nobody.

The model inside RAM does not write. A parser finds the acts and code asks one fixed question about each. An extractive reader of 33 million parameters marks the words in the writing that do the act, or marks nothing. Code then refuses doers that name nobody, from short lists anyone can read, and gives the answer. Long writing is read act by act. The answer is for the person who wrote the sentence, on their own draft, and RAM is not for judging anybody else’s writing.

On five test sets RAM reads right all 178 acts it finds; a parser on its own gets 162. All five sets were written by models, and all have been seen. Notings written by people are not yet measured. The README sets out the limits.

The reader is fine-tuned from deepset’s minilm-uncased-squad2, which is built on Microsoft’s MiniLM, and the acts are found with spaCy. It runs on a laptop CPU.

Code, data and model card:
https://github.com/koherarchitecture/ram

The note:
https://koher.app/notes/ram

Code AGPL-3.0 · weights and training data CC-BY-4.0

← Notes