Mozilla published The experimental addition MemoryCache with the implementation of the dialogue system of machine learning, taking into account the content to which the user turns in the browser. Unlike other MemoryCache AI chats, communication with the user and use the data important for a particular user when forming answers to questions. The project code is distributed under the mpl license. Installation in Firefox so far is supported only manually in the “ABOUT: Debuging” mode and required Applications Patcha.
All components of MemoryCache are performed on the user system and do not turn to external resources. As a knowledge base in MemoryCache, the model gpt4all-j , trained on the collections of English-language content pile (825 GB of data), sharegpt (13 MB) and dolls (660 MB), and accessible in the options from 3, 7 and 13 billion parameters. To work with the model, the instrumentation is involved printtegpt , optimized for the use of large language models on local computers.
The Memory Cache additions to the viewed pages in the ~/downloads/MemoryCache catalog, from which they are then extracted by a periodically launched script and transfers them to Privategpt (INGEST.PY) processor to supplement the existing language model. Thus, the model adapts to the interests of the user and can be used to discuss the viewed content in a text chat with a bot. The user can in the natural language ask the bot questions about the preserved pages, for example, to clarify some facts, and the bot will form an answer using a local model without addressing external services.