2 ms·
not an expert, but i think usually one of two ways - either use openai’s finetune api (the right way) or embed all the docs, retrieve them using embedding of th
by swyx 4y ago
not an expert, but i think usually one of two ways - either use openai’s finetune api (the right way) or embed all the docs, retrieve them using embedding of the questions, and add just enough of them to the prompt to generate a response (gpt-index or langchain help automate some of this).
downside of former is lack of sources, downside of latter is context constraint
- firasd 4y agoI'm not sure fine tuning helps with this kinda thing. Because the fine tuning mechanism involves sending the API a bunch of data like {"prompt": "<prompt text>", "completion": "<ideal generated text>"} So I'm not sure how you would feed it PG's essays in this mechanism