Never mind. I figured it out. I had to call a call the print function since I was not using the Jupyter plugin Dave was using. Silly me. E.g., # Step 3: Parse the response # -------------------------------------------------------------- event = completion.choices[0].message.parsed event.name event.date event.participants print(event)

Hi Dave. Thanks for making this forum. I appreciate ya.

@Grant Fox Hi Grant. Nice to meet you.

Hi Carlos. I share your aversion of dll hell- and all those other bloated MS namespaces. I get cold sweats just hearing the word dll or Visual Studio. As to your question, Is langchain really good for production?, my sense is the answer is could be. Some thirty years ago or so, My brother and I was in a bar playing pool. I was an okay at pool. Well, this really drunk guy who could hardly stand up comes over to the table, lays down a $5 bet, goes over to the rack of pool sticks, grabs one without any tip, and it was twisted as a tree branch. I broke the rack; his turn came up and he ran the table on me. No kidding- true story. LOL. The take away is that the outcome wasn't about the pool table, the pool stick or any of that, it was about his and my level of ability to put the billiard balls in the pockets. I feel a lot of the praise and criticism of various frameworks, and languages, and cloud services and such are merely people trying to get you upset or all jazzed in order to get your attention so that they can get ad revenues or sell you something. It's like building MLM's with the Python ecology- which a lot of people say is slower than C, if you don't know anything about data, data wrangling, statistics, how to test the assumptions of the various statistical methods, and understand the numbers from the measures of precision and accuracy, for example, and don't even really know how to write production ready code, then you're not going to be able to use the Python ecology for create production ready MLM's. My sense is that both LangChain and crewAI are "good" for production, only for those who are experienced LLM DevOps engineers, in which case they are probably not going to be using either of those- like the gods of Linux who do hard-core data analysis with emacs or Vim. For the rest of us, they are great for standing up a MVP, and maybe even putting a front-end on it with Streamlit or Gradio. And, always its a good idea when trying to learn a new development niche to write as much code from scratch as you can. Both LangChanin and crewAI have great documentation and quick start tutorials.

@Carlos Caetano My background is in Organizational Development, specializing in tech integrations. LLM agents for ERP should give you lots of wiggle room to do AI integrations. A really quick way to get your foot in the door would be with using a platform like Zapier. Then, while you are working on that side of the AI development, you could be learning LLM agent integrations. You might want to start with figuring out which of the many, many Enterprise resources you want to optimize- unless you are just generating Enterprise resources utilization process plans. If you are only creating process plans, as opposed to, using AI to optimize resource allocation and acquisition outcomes, then AI agents would be pretty strait forward for you to code and implement. However, if you are aiming at using LLM agents to model and execute streamlined data architectures, then I would suggest starting with doing a process improvement plan for a low-level data integration problem of your choice. Once, you've gone through a few process improvement plans for common to the ERP ecology, then you could begin to figure out how LLMs can do their magic to optimize one or more of those issues you analyzed. Nice chatting with you, Carlos. If you have any other questions, please let me know. Talk soon.

In most cases fine tuning is the process of refining one or more of the prompts being passed to the LLM model. The prompt is the "instruction set" for how the LLM will generate results; so, if the prompt is not getting you the exact results you want, you would fine tuned the "instructions" to see if you get better results. In addition to fine tuning, within the prompt, examples can be given to the LLM. For the more sophisticated models like chatgpt usually just one example, called a one shot, is all that it takes to improve performance; however, a few shot is the process of passing 3 to 5 examples to the model. The only use case I can think of with a fully trained LLM to uploading 100's or 1,000's of documents would be for Retrieval Augmented Generation (RAG), which has nothing to do with fine tuning. The way I remember this is that for any LLM that you need an API for, that LLM is fully trained. What the LLM developer has to do with the LLM is craft the prompt to get the LLM to generate the developer's desired results.

@Mo Ezderman Nope. None of them are in production. They are just MVP proof of concepts. As I was saying the open source versions of those wrappers are not production ready; however, LangChain (langsmith enterprise) and CrewAI (Enterprise) have production ready platforms and those platforms have a pretty impressive list of top-tier orgs on board. The teams at both LangChain and CrewAI have some pretty good coders busting out impressive libraries. At the end-of-the-day, someone has to write the code to get the Agent into dev ops; so, my thought is it might as well be a team of top-notch coders who specialize in LLM wrappers. That may not be true for the tier one enterprises, but for small medium enterprises (SME's) that might be as good as it gets. Take for example other Python libs: can you imagine yourself coding Pandas on top of Numpy? Or, writing scikit-learn from scratch. Good chatting with you. What do you see as the up-side and down-side to my point of view?

@Mo Ezderman I think you nailed the tension in the emerging LLM Agent app engineering niche- complex abstractions vs. production ready code. And, there is kind of an irony in that you mentioned Pytorch. Not only is it "out of the question for most!" folks to code; it's out of the question for most people to use because (almost) no one can train a better deep learning model at a lower cost than is available for "free" at huggingface or elsewhere. So, Pytorch may be an example of a framework that is already obsolete for most use cases- except for those who want to take an educational deeper dive into deep learning about deep learning- which I would recommend. I started coding full stack web apps back in the early 2000's and the JavaScript, then PHP landscapes are a junk yard of rusted out frameworks. But, frameworks are still around. My sense is they serve two main purposes: 1. they standardize a set of complex processes, 2. they are as much a social club and support network, maybe more, than a tool. The other reason frameworks (libraries, packages, etc.) are not going to go away is because of the necessity of top-tier Enterprises to adopt standardized code bases because standardization ensures that there is a robust and ready supply of coders to fill the demand for workers, and no one, but no one, this late in the tech revolution wants to get stuck with consultant ware! The next reason for frameworks, take LangChain for example, is that their "complex abstractions", like what Pandas does to Numpy or Jupyter does to Python compilers, reduces the learning curve to a single (few) libraries. The complex abstractions of say LangChain essentially enables a developer to use a unified set of modules to manage the integration of a potentially large number of API's. Even a simple RAG agent requires an LLM (or two or three...) a vector database, document processing, and this and that... and the other. So, even a base chat bot can present a decent to seriously front-loaded learning curve. And all of that even before you get to fine tuning. Not to mention, getting it deployed to the cloud, which is another story in itself.