3 ms·
OP is incorrect. Embeddings are still needed since (1) context windows can't contain all data and (2) data memorization and continuous retraining is not yet via
by lazzlazzlazz 3y ago
OP is incorrect. Embeddings are still needed since (1) context windows can't contain all data and (2) data memorization and continuous retraining is not yet viable.
- nextworddev 3y ago"yet"
- coding123 3y agoIt's also much slower. LLMs are generating text token at a time. That's not very good for search. Pre-search tokenization however, probably a good fit for LLMs.
- zwily 3y agoBut the common use case of using a vector DB to pull in augmentation appears to now be handled by the Assistants API. I haven't dug into the details yet but it appears you can upload files and the contents will be used (likely with some sort of vector searching happening behind the scenes).