Sergio Castro

Pressure Testing GPT-4-128K With Long Context Recall 128K tokens of context is awesome - but what's performance like? I wanted to find out so I did a “needle in a haystack” analysis Some expected (and unexpected) results Here's what I found: Findings: * GPT-4’s recall performance started to degrade above 73K tokens * Low recall performance was correlated when the fact to be recalled was placed between at 7%-50% document depth * If the fact was at the beginning of the document, it was recalled regardless of context length So what: * No Guarantees - Your facts are not guaranteed to be retrieved. Don’t bake the assumption they will into your applications * Less context = more accuracy - This is well know, but when possible reduce the amount of context you send to GPT-4 to increase its ability to recall * Position matters - Also well know, but facts placed at the very beginning and 2nd half of the document seem to be recalled better Overview of the process: * Use Paul Graham essays as ‘background’ tokens. With 218 essays it’s easy to get up to 128K tokens * Place a random statement within the document at various depths. Fact used: “The best thing to do in San Francisco is eat a sandwich and sit in Dolores Park on a sunny day.” * Ask GPT-4 to answer this question only using the context provided * Evaluate GPT-4s answer with another model (gpt-4 again) using @langchain evals * Rinse and repeat for 15x document depths between 0% (top of document) and 100% (bottom of document) and 15x context lengths (1K Tokens > 128K Tokens) Next Steps To Take This Further: * Iterations of this analysis were evenly distributed, it’s been suggested that doing a sigmoid distribution would be better (it would tease out more nuanced at the start and end of the document) * For rigor, one should do a key:value retrieval step. However for relatability I did a San Francisco line within PGs essays. Notes: * While I think this will be directionally correct, more testing is needed to get a firmer grip on GPT4s abilities * Switching up prompt with vary results * 2x tests were run at large context lengths to tease out more performance * This test cost ~$200 for API calls (a single call at 128K input tokens costs $1.28) * Thank you to @charles_irl for being a sounding board and providing great next steps