Discussion

You Don't Need a Vector Database

sirfz: You don't need a vector db, you just need np.dot

kenforthewin: If I understand correctly, the vecstore product is an api that wraps a vector database (among other features). So the pitch here is something like “you don’t need a vector database, you need an api wrapper to a vector database that we manage for you”.

EspadaV9: If you need to store vector, surely the easiest solution is to run `CREATE EXTENSION vector`. When, or if, you need more, then look at alternatives, but I'm sure most people will have a much easier time just adding it to their existing database.

SirHumphrey: I was asking myself the whole article “what does this mysterious semantic search api actually do?” and was a bit underwhelmed when the result came out to be - managed vector database.

_puk: So the classic build vs buy question

canpan: Pgvector is definitely the easiest way to get started. I did not really get the problem the article was trying to sell.Setting up an ingestion pipeline to your existing db, vs ingesting into yet another db seems to not solve a problem I have.If there was one thing I wish pgvector was better at, it would be to allow composite indexes (ie find vector where category). But it's a minor point.

stephantul: This is a thinly veiled commercial, not really useful.

cpursley: Or you could just use the new BM25 extension and if thats not enough, bring in the vector extension, which you can run as hybrid and not have to bolt yet another paid 3rd party thing:https://postgresisenough.dev/tools?category=search

Raro: If this were on Mad Men:"You don't need a hamburger... you need McDonald's".

exhost: > Please don't use HN primarily for promotion. It's ok to post your own stuff part of the time, but the primary use of the site should be for curiosity.

alansaber: Exactly this, hybrid search (weighted in favor of a good sparse retrieval strategy) is universally the best way to go.

wat10000: When they generated this post, they set their LLM to “long and obfuscated” instead of “concise and clear.”

kencho: you're completely wrong. we also added "make no mistakes" in the end

noemit: The Vector Database obsession came from RAG, which came from a marketing idea to calm down enterprise fears about hallucination with RAG. Will save this article because I feel like I have this conversation weekly when people think they need a vector database for something they definitely do not.

alansaber: That RAG is marketing and doesn't significantly affect performance is incorrect. As to whether retrieval really benefits from vector DBs is another question.

kencho: the article was aimed more at teams that don't have an existing postgres setup and are evaluating standalone vector databases from scratch. if you're already running postgres with pgvector, you're in a good spot

kencho: pgvector is great if you're already on postgres and only need text search. funny enough, Neon Postgres actually featured us in a case study about this exact topic. we replaced pinecone and rds with neon under the hood: https://neon.com/blog/vecstore-replacing-pinecone-and-rds-wi...the gap shows up when you need image search, face search, or content moderation on top of text search. that's where a dedicated api makes more sense than rolling your own on postgres

kencho: Exactly. We talk to teams every week who spent a month setting up pinecone or qdrant and then realize they just needed search that worked. The vector database became the default answer to every search problem because of the RAG hype cycle, even when the actual need is way simpler

simonjgreen: > Spent a month setting up Pinecone? Really?

kencho: there's a lot more in "setting up" than creating an account and a collection on pinecone or any other service

EspadaV9: I was genuinely surprised just how easy it was to get a fully working RAG set up with Postgres. It was a few hours over a weekend to get something "working" and then probably a bit less time a following weekend to have a nicer database structure and rebuild it learning from the mistakes during the first attempt. The harder part comes next, because that involves multiple tables of user provided data, multi tenancy with a shared core vector schema, and all the actual business logic, so I've put it all on hold for a real breakdown now, but I wouldn't expect it to be much of a problem with what I've found so far with pgvector, and Postgres in general.