G2

Akshay D.

Talent Sourcing Coordinator - MEAAA

Mid-Market (51-1000 emp.)

10/25/2026

"Cost savings and simpler hybrid search for multi tenant RAG"

4.5/5

What do you like best about Turbopuffer?

Honestly the cost model. We were running a multi tenant RAG setup where most of our namespaces sit idle the majority of the day (customer knowledge bases, support docs, etc.). With our previous vector DB we were still paying for the full capacity even when nothing was being queried. Switching to Turbopuffer felt almost too good to be true at first. Data lives on object storage so idle namespaces cost almost nothing and the hot ones stay fast once they are cached. The first real oh shit this actually works moment was when we onboarded a big new customer over a weekend. We dumped a few hundred thousand docs into a fresh namespace and the bill barely moved. Same workload on the old system would have made me open a spreadsheet and start sweating. Hybrid search (vector + BM25) in the same API is also really nice. We do not have to run a separate keyword engine anymore. But the cost savings and the fact that it just scales without me having to think about clusters is what I like best. Review collected by and hosted on G2.com.

What do you dislike about Turbopuffer?

Nothing major, but the cold start latency can be noticeable if a namespace has not been queried for a while. In our case most of the important namespaces stay warm because they get regular traffic, so it rarely shows up. The first time we hit a completely cold one during testing though the response took a few hundred milliseconds longer than we expected. Once it is cached everything is fine again. It is a fair tradeoff for the cost savings and the way the architecture works, and it has not been a problem in production for us. Still, if your workload has a lot of rarely accessed namespaces and you need consistently low latency every single time, it is something to keep in mind. Review collected by and hosted on G2.com.

What problems is Turbopuffer solving and how is that benefiting you?

The biggest problem it solved for us was the cost of keeping large amounts of mostly idle vector data available. We had a multi tenant setup with hundreds of customer namespaces. Most of them only get queried a few times a day or less. On our previous vector database we were still paying for capacity as if everything was hot all the time. The bill kept climbing even though actual query volume was not that high. Turbopuffer changed that because the data lives on object storage and only the namespaces that are actively used stay warm. Idle ones cost almost nothing. When we switched, our monthly spend dropped noticeably while we were still able to keep every customer’s data searchable. It also simplified our stack. We used to run a separate keyword search system next to the vector one. Now hybrid search (vector + full text) lives in the same place, so we have one less service to maintain and one less place where relevance can drift. Overall it let us scale the number of customers and the amount of data we index without the infrastructure cost growing at the same rate. That is the part that has made the biggest difference for us. Review collected by and hosted on G2.com.