Add the summary / conclusion?

[r4881t]

It would be great if you could add a conclusion on your observations. I landed here hoping to see some comparison. The notebooks are great, but to run all of them and conclude for myself is a bit too much of effort for initial judgement.

[IuriiD]

Hi, @r4881t . Thanks for the question. I did some comparison and summary in the accompanying video - please see for pinecone, faiss and pgvector, correspondingly. If to sum up in a few words and judging from my modest experience so far:

• Pinecone is good for quick prototyping, startups etc, where one needs quick development and also would like to distantiate from the devops/support side, but is Ok to pay for that;
• pgvector allows to save money, possibly avoid adding a new integration (if a Postgresql DB is already in the stack), also allows to keep all the data internally (though Pinecone probably offers security in this sense too). I'd expect to see it in commercial conversational projects;
• FAISS seems to be a quite efficient but a rather specialised tool, might need additional effort to backup the index (which also might increase the costs).

[steve-liang]

@IuriiD This is awesome, thanks for running the comparison and your summary will save me days or even weeks.
FAISS is truly a specialised tool like you said, I am planning an architecture that aims to achieve highest level of efficiency AND accuracy that is to use pgvector as the first level coarse search, then after narrowing down candidates, use faiss for secondary fine search, on the fly.