This is an extension for text-generation-webui named after the Egyptian Goddess for writing, wisdom, and knowledge. The extension is designed to write scientific text based true scientific publications (and with citations to these). For the best results, the right publications should be referenced in the query, as arxiv:0000.00000 or pmid:00000000. Alternatively, you can mark keywords in your query with curly brackets, and these will be used to search arXiv and/or PubMed. Finally, you can choose to have keywords be automatically detected in the query, but this is an experimental feature that often yields poor results.
To install this extension, clone the repository with the extensions folder of your text-generation-webui installation. This will create a new directory called Seshat. For the best results, I recommend that you move the files Seshat.png and Seshat.yaml from this directory into your characters directory and activate this character when using the extension. The character card contains important instructions to try to ensure that the LLM sticks to the provided references. These instructions could have been built into the extension itself, but are instead provided as a character card to allow for prompt engineering without having to edit Python code.
The extension works well with a wide range of instruct models, including Meta Llama-3.1 8B and Mistral NeMo. My current preferred model is Qwen2.5 14B (or larger), which I find to be as good as any model at adhering to the instructions and writing as any model. Its context length of 32k is also sufficient to allow for many references.
