Hey everyone! I've been diving into vector search for a personal project, and I keep hearing Pinecone is the gold standard. So I spun up a small test to index about 500,000 embeddings (384 dimensions each, generated from an open-source model).
When I looked at the projected cost for a single index of that size on their standard pod, my eyes kinda popped out 😅. It seems like it would run over $100/month just to keep it running, not even counting the queries.
I get that managed services have a premium, but coming from a background where I'm used to running PostgreSQL or even Elasticsearch on a modest VM, this feels like a huge jump. My project is just a prototype, so that cost is a non-starter.
My naive thinking was: it's just storing vectors and doing similarity math, right? So my questions for the more experienced folks here are:
1. What exactly am I paying for at that scale? Is it mostly the RAM for holding everything in memory for fast search?
2. For a side-project scale (500k vectors, maybe a few thousand queries a day), are there practical, more cost-effective alternatives in the current stack? I've heard names like Weaviate, Qdrant, and pgvector thrown around, but I'm not sure where to start evaluating.
3. Does the cost equation change dramatically if I'm willing to handle more of the infrastructure myself (like using a library like FAISS and orchestrating updates myself with Airflow)?
I'm trying to understand the trade-offs between "fully-managed, zero ops" and "some ops, but my wallet isn't crying." Any insights or personal experiences would be super helpful!
-- rookie
rookie