Closed marcusschiesser closed 8 months ago
Thinking about this again, if we add the whole document to the context, we are quickly reaching the limit of it for the models with smaller context size. Probably better to store the documents in a vector store (see #22)
Using URL documents, after the second message, the wrong messages are sent to the LLM.
The first message is sending the correct message to summarize the content:![image](https://github.com/marcusschiesser/unc/assets/17126/554f4c7a-e133-4878-bb2e-29b16de86e06)
But the second message is sending![image](https://github.com/marcusschiesser/unc/assets/17126/932ae650-cc20-44d6-a0cf-6d05f429a69b)
urlDetail.url
instead of sending the whole document as a message (urlDetail.content
):