Hello. I want to use the Assistant API to create an assistant whose knowledge base will consist of a JSON file (array of 1000 entries for example). Next, the user will ask him questions and he will answer them based on his knowledge base. but, I encountered a problem that the assistant (file_search or retrieval) does not perceive the file to the end, that is, he does not see some part of the records. I know about 800 tokens into which the AI breaks the contents of the file, but how can I make the AI understand the entire file and give accurate answers to questions? Thanks
If all information in your document, from the start to the end, are important to your success result, all you can do is to put all necessary text content in your thread message to tell assistant to answer based on them thoughtfully, and ensure you run a model that has enough context window size.
RAG and it’s similarity nature is to buy speed and convinency with accuracy, to overcome the context window size limitation, especially in former 4k context size period.