Is it better to have one large file or multiple small files with GPTs?
tl;dr: Using a single file seems to reduce the likelihood of a knowledge-base loading and improve response times overall.
Introduction: Custom GPTs allow you to upload files that it will reference when answering questions. In my experience, the response time with these knowledge-base GPTs is really slow. That lead me to the question:
Is there a way to improve the response time by altering the number/length of files in the knowledge base?
I expected that the single large file would result in a faster response time than many smaller files, because loading an additional file will take longer than reading the same additional content from an already opened file.
Materials and methods: I uploaded the maximum number of files with dummy
text to a GPT. One file contained a secret:
the secret is: banana bread (file
6 of 10). I then asked the GPT “what is the secret?” and timed the response with
a stopwatch. In between runs, I removed all of the files and added them back. I
ran it again with single file of approximately equal length that contained the
same secret text.
- Many files (seconds)
- Single file (seconds)
† Indicates runs where the GPT indicated knowledge-base loading
- Many files (with loading): 14.97
- Many files (without loading): 7.32
- Single file: 8.39
Discussion: This experiment was pretty hacky! Using a stopwatch was not ideal, and it was difficult to decide when to stop with a streaming message. Response times may have varied significantly due to network connections or traffic on OpenAI’s servers. We shouldn’t take away any specifics from this experiment, but we might be able to draw some general insight.
Conclusion: When the GPT loads from its knowledge base, the response times were significantly slower. But when it didn’t, or at least didn’t indicated that it did, the results were similar between multiple files and single files. Based on this, my hypothesis was wrong. Generally, the response times did not vary significantly for multiple files vs a single file.
However, the data seem to show that using multiple files is more likely to trigger knowledge-base loading. And since that takes a long time, it is preferable to use fewer files to improve response time.