1. Check out this cool chat-with-my-docs app I made with LangChain! 2. Okay, let's productize! 3. Ouch, the results aren't quite production-ready... 4. Huh,

1. Check out this cool chat-with-my-docs app I made with LangChain! 2. Okay, let's productize! 3. Ouch, the results aren't quite production-ready... 4. Huh, the bottleneck seems to be the search & retrieval process... 5. Oh, I guess I don't know anything about search, time to catch up on the field... 6. Ok, now I need to build an inverse index or knowledge graph and augment it with embeddings. 7. Hm, actually, language models (including BERT-based, Splade) could be pretty useful in building structure out of my dataset... IMO the best way to incorporate transformers into search has yet to be discovered, although whatever OpenEvidence.com is doing seems to work well. If I were still in grad school I would look into some kind of PageRank / graph NN approach to the problem, or maybe hyperbolic embeddings. Reason being that for many corpora there exists some kind of internal structure or linking that can be exploited for speed and accuracy, which you don't quite get from nearest neighbors in a Euclidean space.