Hey everyone! I'm exploring **Amazon Bedrock** for deploying GenAI applications at a production scale by 2025. I've read through documentation and marketing materials, but I'd love insights from real users. I'm curious about several points: 1. Are you using Bedrock in production, and for what types of applications (like chatbots or content generation)? 2. How does it compare to running models on SageMaker or using APIs from OpenAI or Anthropic? 3. Have you faced issues with latency, costs, model performance, or vendor lock-in? 4. How's the integration with LangChain, RAG, or vector databases such as Kendra or OpenSearch? Is it smooth or complex? 5. Do you feel it's enterprise-ready or still in the works?
5 Answers
Many big brands are already using Bedrock for their production workloads, and it’s reported to be as mature as many other AI service platforms. The main issue seems to be capacity management, but cross-region inference could help with that.
We’re currently using Bedrock, but tracking costs has been a real hassle. You can’t tag model invocations per call, which complicates things if you’re managing multiple clients. Setting up separate agents for each client leads to a lot of overhead. It should be simpler to just have a tagging system during invocation to track output and consumed tokens efficiently.
I think Bedrock is pretty cool as it acts like a wrapper for a bunch of models, giving you a standardized way to experiment and find the right fit for your applications. The access to different models is great, and it's evolving constantly. It would be even better with integrations to OpenAI or Gemini, but I understand why they haven’t included those.
But isn’t it more than just a wrapper? They actually host the models, so it’s definitely a more robust solution.
Exactly! It’s a full GenAI platform with features like agents and knowledge bases that really enhance the development experience.
Yes, Bedrock is definitely being used by larger enterprises for significant inference workloads. There are default quotas, but those can be increased based on your needs. Plus, Bedrock has evolved a lot with features beyond just basic FM inference; it seems ready for enterprise use.
True, but have you personally experimented with Bedrock’s capabilities, or just read the docs? Real-world experience can be a game changer.
From what I’ve seen, running Claude against AWS Bedrock tends to have lower latency and better uptime compared to using the Anthropic API. Just keep in mind that you might face some throttling regardless of your implementation. Also, it's worth mentioning that we had a 2-day outage recently with Bedrock, and we noticed occasional timeouts even with minimal traffic. After switching to OpenRouter, we haven’t had any issues since.
That’s interesting! I’ve heard some people praise Bedrock for uptime, but outages can definitely throw a wrench in things.
Have you looked into application inference profiles? They might help streamline your cost tracking.