Mozilla a prezentat experimental MemoryCache add-on, implementing a machine learning dialogue system capable of considering the content that users interact with in their browser. Unlike other AI chats, MemoryCache allows for personalized interaction with users and uses specific data about the user when generating responses to questions. The project code se răspândește is licensed under MPL. The installation in Firefox currently susținută is only manual through the ‘about:debugging’ mode and requires a patch.
All MemoryCache components run on the user's device and do not access external resources. The model GPT4All-J, trained on English-language collections Pile (825 GB of data), ShareGPT (13 MB) and Dolly (660 MB), is used as the knowledge base in MemoryCache. The model is available in versions with 3, 7, and 13 billion parameters. To interact with the model, the toolkit privateGPT, optimized for working with large language models on local computers.
Memory Cache saves viewed pages in the directory ‘~/Downloads/MemoryCache’. A periodic script extracts pages from this directory and passes them to the privateGPT handler (ingest.py) to update the current language model. Thus, the model adapts to the user’s interests and can be used to discuss the viewed content in a text chat with the bot. The user can naturally ask the bot questions about the saved pages, for example, to clarify facts, and the bot will generate a response using the local model, without accessing external services.
Sursa: linux.org.ru
