We're using Pinecone to build our RAG pipeline. We need a vector database, and we have a lot of options in the market. RAG is the biggest use case for us.

The first thing is that we've always been using AWS. AWS provides OpenSearch serverless out of the box, but OpenSearch happens to be pretty expensive because you have to pay per hour of use if you want to have an OpenSearch server alive. It's billed as the number of OCUs. Pinecone, on the other hand, is pay-as-you-go on the number of queries. You only pay for the queries that you hit.

Pinecone's integration with AWS was seamless. All we had to do was take one of the API keys and upload it to AWS's Key Management Service, and then configure that through it, and then it starts working seamlessly. When you're building a production system for RAG, Pinecone gives you the vector search, but you still have a lot of pieces that have to come with it, including embeddings, chunking, pre-processing the query, and security. Pinecone doesn't provide that out of the box. AWS has the infrastructure for it. When you're using Bedrock with Pinecone, it becomes a good combination because Bedrock itself is free. They only ask you to pay for the model invocations.

Pinecone is flexible. They give you a bunch of options. One of the good features is that they also provide embeddings within Pinecone, which is a neat feature. You can essentially choose your embedding sizes and things like that. So you do have some control over it. It's easy to set up, and we felt like it's not that expensive for us in comparison to serverless. That's why we took it.

If Pinecone gave us RAG as a service, we'd be more than happy to use that. Then we wouldn't have to go to something like AWS again.

We've been using Pinecone for a little over four months.

So far we haven't scaled it to that extent. We're just building a beta version of it. For the beta version, at least so far, it's been good. We're demoing this to a few people, and then we'll possibly scale up if needed. But so far, it's looking good.

We've rolled out the early version as a beta access to a few, maybe twenty to thirty customers. So far, there haven't been that many complaints, but also it hasn't been really stress-tested for say, ten thousand requests per minute or something like that. We haven't really put it to the test. But for these demos for our clients to use, it's working fine so far.

I have not personally engaged with customer service, as there are people above me who are making those decisions. I work as a developer and am just integrating everything. I haven't needed support because the documentation is good enough to help developers get up to speed.

The documentation is great. Plus, they have a chatbot that can help you answer all the questions about documentation, which I find helpful. I would say it's even better than AWS's documentation because AWS's SDK documentation is just not as helpful.

We weren't really sure about Pinecone security, and that's why we're using AWS for it. AWS is going to handle that whole pipeline of security and making sure that everything is passing through correctly. Pinecone comes in at just one of the stages, where it has to either at inference give you the most similar vectors or store your embedded chunks into a vector database. It's just one small piece in this. Most of the heavy lifting is done by our back-end plus AWS.

We were also using S3 Vectors, but it's still in preview. They haven't released it for all regions. It works in the US East, but in Europe West, it's not live yet. So we weren't able to go ahead with S3 Vectors. Pinecone was available though, and that's what we're using right now.

We're using Pinecone as a vector database over OpenSearch.

As a standalone vector database, I think Pinecone gets the job done. I would give it an eight out of ten. Overall, I rate this product an eight.