[kankerlijer]

For those building applications with Langfuse and Clickhouse - do you like these products? I get the odd request to do an AI thing, and my previous experience with LLM wrappers convinced me to stay away from them (Langchain, Llamaindex, Autogen, others). In some cases they were poorly written, and in other ways the march of progress rendered their tooling irrelevant fairly quickly. Are these better?

[deaux]

Ive used Langfuse. It's completely unrelated to tools like Langchain and Autogen. It's just logging/tracing for LLMs. Sure they added stuff like "prompt management" and "epxeriments" etc. probably to keep investors happy but those are entirely optional sidedishes.

The tools you mentioned are indeed to be avoided. I trialed them early on and quickly realized in 99.9% they do nothing but bog you down. Pretty sure they'll be dead sooner rather than later.

[embedding-shape]

The observability stuff can be nice for deployments but really, these libraries/frameworks don't actually do much more than provide some structure, which unless you're expecting a team with high turnover to maintain it, doesn't really matter all that much, especially if you're an experienced developer, you'll find better design/architectures fitting for your use case without them.

[st3fan]

Hm I find this very much a "please reinvent the wheel" take.

These frameworks provide structure for established patterns,but they also actually do a lot that you don't have to do anymore. If you are for example building an agentic application then these kind of frameworks make it very simple to create the workflows, do the chat with the model providers, provide structure for agentic skills, decision making and the human in the loop, etc. etc.

All stuff that I would consider "low level". All things you don't have to build.

If you have an aversion to frameworks then sure - by all means. But if you like to move faster and using good building blocks then these frameworks really help.

One thing to keep in mind - many of these AI frameworks are open source and work really well without needing backend services. Or you can self host them where needed. But for many that is also the premium model, please use and pay for our backend services. But that is also a choice of course.

[embedding-shape]

> All stuff that I would consider "low level". All things you don't have to build.

But those are also very trivial to build, and you end up having to customize them for your need, and if the framework don't have those levers, better be prepared to either fork the framework, or spend time contributing upstream.

Or, start simple yourself with what you need, use libraries for the hairy parts you don't want to be responsible for the implementation of, then pipe these things together. You'll get a less compromised experience, and you'll understand 100% how everything works, which is the part people generally try to avoid and that's why they're reaching for frameworks.

> But if you like to move faster and using good building blocks then these frameworks really help.

I find that they help a lot with the "move faster" part in the beginning, but after that period, they slow you down instead. But I'm also a person that favors "slow software design and development" where you take your time to nail down a good design/architecture before you run. Slow is fast, and avoiding hairballs is the most important part if you're aiming for "move fast for longer" rather than "a sprint of fast".

[agsqwe]

We like Langfuse for observability via OpenTelemetry. Prompt management is too basic for our needs.

[marcklingen]

Let me know what features are missing in prompt management