Currently if you drop even a single document onto Open WebUI and ask a question about that document, it appears that RAG is always used. It would be helpful to be able to switch this off.

Currently, for each attachment the document text is extracted, the text is chunked, embeddings computed and indexed, an index lookup is performed and then some of the chunks are passed to the LLM. While this is an OK solution for retrieval tasks, it abjectly fails for summarisation tasks.

For LLMs with small context windows summarisation is hard, but many LLMs now support very large context windows; GPT-4o has a 128K context windows and Anthropic Claude 3 support 200K tokens or about 400 pages of text. To make best use of these longer-context models it would be helpful to be able to pass the entire extracted text of all attachments into the LLM. Having a switch to bypass the RAG steps would achieve this.