I'm sure it's so very tempting for him to put up a corrected or version 2 run without his learning curve. I am so glad he resists this urge. Yes it would be much less tedious. But we are coders and users that need to see the thought process on how to correct our errors when we encounter them. Seeing his MO on this helps me understand how to attack errors I will be getting.
@zakariaabderrahmanesadelao3048 May I sugest that you try Yacana ? Search for "Yacana github" on google. It's the fun without the pain. No silly errors like in crewAI. ^^
It would be more encouraging if we had somebody who knew what they were doing. Everything he is doing he has learned as he goes, just like many of us who use ChatGPT and Cursor.
I really love that you showed all the struggle working with unknown libraries and codes. Instead of performing the 'chef's trick' you show all the issues you run into. This is a reassurance that I am not stupid and even the best of us face the same problems.
For those of you who will try to follow this guide. The problem is that we need a venv to install crewai and then crewai installs it's own venv... And Python gets upset. Here is what I did. Once your project stops working, which is when trying to use langtrace. At that point, I copied all the source for my crewai project up one folder - into the folder where my original venv is located- everything except the .venv folder. I deleted the rest (the whole crewai project folder) Then I just re-ran crewai install in the folder where my src and .venv (the one currently activated) are located. And Voila! Working! Hope this helps.
can you elaborate a little more on this please. I am new to coding, and have been to find my way through with gpt, but I still cant solve the langtrace problem :(
TBH, I am not a long time viewer and think your AI content gets a bit redundant and boring, though that is how I found your channel. I get AI is all the rage and embrace profiting from it... Then I watch this and learn that you are actually a skilled coder. This half hour video earned my respect. I hope you put out more content like this. Scrolling through the comments, it seems you once did. The AI testing is cool, but this kind of content actually brings value. An at least once a week version of this would be very nice to see. It would really make the channel stand out from the pack and give you a better position to be in when full saturation occurs on RUclips with this content. As a coding hobbyist myself but much better engineer, I'm glad that a seasoned programmer has the same issues I run into, because I'm constantly thinking I simply don't know what I'm doing! It seems that turns out to only be half true. I actually like python until dependencies break everything. Why isn't there a simple method of making it so that every unique python script runs in it's own environment by default, relying on the dependencies file to pull both local and new dependencies, that get stored to a local pool. It's a pain to have to keep track of running virtual environments each time a script is run. Conda sucks! Isolate each script automatically and only attach the dependencies that are relevant. 1 day of writing code and 2 more of chasing dependency issues is not cool, especially when dealing with incompatible versions, etc.. There has to be a better solution for handling this chaos!
In real life, most of the work lies in the analysis, which then leads to an efficient code structure. One hour of analysis can save ten hours of coding. So, I usually spend a fair amount of time exchanging messages with o1-preview just to brainstorm and think the project through. After that, I follow the class diagrams, project file structure, and so on. That could make for a great idea for an upcoming video 🙂 Love your content, man! I’ve been following you almost since the beginning of your channel.
Great video Matt! This one feels so authentic and hits home for me personally. I enjoy (don't take this wrong) seeing you making the same mistakes we all make, and working through issues, while working on real problems that aren't just Snake in Python! Cheers.
Hey Matt, thanks for doing near realtime AI builds with cursor and AI copilots. I honestly think that this content is some of the most valuable that you do, and yes, that means if it took paying to subscribe to it, I for one would pay to see one of these a week? Using different platforms and tools on a variety of use case scopes. Excellent content!
When you are using cursor. I would highly recommend using cmd + k or cmd + i for ai help. cmd+k is inline quick fixing. will propose an inline diff of changes for fix. cmd+i is composer mode which is essentially the ai developer. you have detailed control over context and can specify files, folders, internet usage and more. also be sure to setup the stuff in "Cursor Settings" like indexed docs sites and cursor rules it will make your experience so much better
Conda is sort of a necessary evil. You have bare metal, a VM, a docker container and then conda. Sometimes I don't want to create a new docker or VM for a clean python install, and conda does the trick. It's pretty handy for running a bunch of different environments or inference engines locally.
@@chrisstephens9460 I do not see Conda as being necessary for anything. Why would I want to use it for anything? I have managed to do everything I need and want to do without Conda.
Super-helpful to see your live tuning thoughts and troubleshooting to learn from you. Thanks! Sometimes that's more helpful than something you know will work.
🎯 Key points for quick navigation: 00:00 *📘 Project Overview: Educational Portal using Crew AI* - Introducing an educational portal project to teach AI concepts from basics to advanced. - Automating content creation for educational resources, including text, images, and step-by-step guides. 01:08 *🔧 Setting up Python Environment for Crew AI* - Creating a new Python environment with Conda for project setup. - Installing Crew AI and LangTrace SDK within this environment to manage dependencies. 02:18 *🚀 Initiating Crew AI Project and Configuring Files* - Crew AI project structure setup, creating necessary files and folders automatically. - Selecting OpenAI’s GPT-4.0-mini model for initial testing with low-cost, lightweight configuration. 05:02 *💡 Troubleshooting Python Environment and File Structure* - Challenges with managing Python environments across different editors and configurations. - Overcoming issues with LangTrace SDK installations and correct environment selection. 08:42 *🛠️ Integrating Perplexity API and LLM Model Selection* - Testing the Perplexity API for automated research without web scraping. - Experimenting with LLM models to improve research capabilities; troubleshooting 404 errors with API setup. 14:12 *🔍 Setting Up Real-Time Web Search with Serper Dev* - Installing and configuring Serper Dev tool within Crew AI to access real-time web data. - Testing Serper Dev API key integration and confirming working real-time search results. 18:04 *🌐 Testing and Refining Search and Reporting Outputs* - Running search queries and generating updated, detailed reports using current AI trends and models. - Transitioning from Perplexity to Serper for improved, up-to-date results in generated reports. 20:24 *🔒 API Key Configuration and Environment Security* - Moving API keys to environment variables and ensuring they’re excluded from main files for security. - Adding `.env` file configurations and confirming `gitignore` setup. 21:18 *📊 Initial Test Run and Model Performance Analysis* - Testing setup with GPT-4 mini, analyzing initial cost-effectiveness, and identifying model limitations. - Observing results on broad topics and considering more specific prompts for in-depth insights. 22:14 *🎓 Refining Agent Roles for Educational Content Creation* - Adjusting agent roles: defining a senior researcher and content creator with tasks aimed at producing comprehensive educational materials. - Updating agents' purpose and instructions to prioritize clarity and educational value. 24:17 *✍️ Customizing Output Requirements for Educational Content* - Setting precise output expectations, like detailed, structured educational content over a simple report. - Implementing Markdown formatting for topic breakdown and clear organization. 25:26 *📈 Report Testing with Real-World Content Examples* - Generating reports on complex topics like retrieval-augmented generation, assessing quality and coverage. - Ensuring content includes clear, structured sections like definitions, phases, and benefits. 26:07 *⚙️ Switching to Advanced Models for Improved Depth* - Testing 01 Mini model for enhanced content generation, observing higher quality and cost tradeoffs. - Reviewing performance on LangTrace, noting token usage and cost increases with the 01 model. 27:26 *🖼️ Future Plans: Reviewer Agents and Visual Enhancements* - Planning additions like a reviewer agent to improve content accuracy and coherence. - Exploring potential for dynamic graphic generation to supplement educational content. 28:06 *💰 Cost and Model Selection Insights* - Analyzing cost implications of advanced models like 01 Mini, recognizing improved quality at a higher expense. - Comparing model efficiency and value in the context of generating comprehensive reports for educational use. Made with HARPA AI
Hey man, always a great content!! Instead of "conda" use "uv" or "poetry" which are easier and modern way of managing environments/dependencies! Hope you use in your next journeys
I use VS + Cline + Aider but I do see that CURSOR has many benefits, especially all these "auto complete" features ! It was an absolute pleasure watching this video ! More like this ! 🤤Looking forward to part 2 !
There should be a name for these kind of videos that show the pain and struggle of actually using a software library. I'd watch the crap out of those for libraries I'm trying to use.
In your code editor if you use cmd + shift + p and type ‘select Python interpreter’ to set your editor up to use the environment you previously defined, you’ll avoid a whole lot of issues around environment management
Seems promising... I guess... But still, CrewAi feels so random and doesn't work well with local LLMs. I prefer using Yacana to build agents. It can also do brainstorming even though the LangGraph like workflows feels way better.
Yeah CrewAI, even with the monitoring stack is one hell of a mess. I would love Yacana to add the monitoring though. Alhough it's less of a headache than any other frameworks out there, monitoring is always useful...
Thanks for the interesting content! Cursor is great but becomes perfect when you give the right context to every question. You would then not have to go to an external source to fix your issues.
I recommend using Poetry instead of Conda for creating virtual environments in Python, as it makes the process much easier and more efficient. In complex ML/AI projects with many dependencies, Poetry often provides a smoother and more reliable experience due to its robust dependency management and streamlined workflow.
Python environment management can be made a little easier with tools like Miniconda, as you can easily choose which versions of Python to use for each of your environments. Some projects like to follow the standard Python methodology and make a .venv folder within the project, in which case the environment is activated using “source ./.venv/bin/activate” and deactivated with “deactivate.” Sometimes it’s a good idea to make a conda environment even if the project will contain a .venv folder, so the version of Python used is not just the system version. If it turns out the conda environment is redundant you can just delete it, but some .venv contents might just be symbolic links to the conda environment, so check that first.
FYI - You can set the temperature and the context length in the openai APIin your Python code. Gpt40-mini works great for most things just lower the temp and give it the models max context length, your get more accurate and longer responses.
I believe that your issue is related to the IDE used. That's true I don't use VS Source, I use Pycharm that understand external environment created with conda/pip/etc ....
Virtual Environment Issue: The virtual environment issue exists because we create a venv with conda, then when we run crewai install, it creates another virtual environment. A simple fix to this is: - within your project folder, edit the file .venv/pyvenv.cfg - Change the line that says include-system-site-packages = false to include-system-site-packages = true This will cause the .venv created by crewai to reference the conda environment packages, which is where pip is installing the libraries.
Thank you for sharing. I was recently working on a crewai project with vs code. I was amazed how you see these traces and costs analysis? Is it the IDE you are using or any plugin? It will be very helpful for me to use it.
I seem to be one of the few that did not find my own solution to the langtrace error. Could you lend any advice for fixing that issue? Love these videos. Super excited to get CrewAi running myself.
Awesome video Matthew, really liked this format! As for the python env situation I also had a bunch of problems with cursor so I stuck with VSCode, but eveytime I watch a cursor video I feel the urge to change back and give it another shot 😅 I'm also trying to change my python env management to uv, its much faster and simple ( crewai uses uv to manage the envs and installations ) . Keep it up Matt, looking forward to the rest of this series!
I've seen people and I've used Dev containers in VScode. I like the docker for it's shareability. It's just a little more work to get each project set up. But when you do it's really superior to other methods.
So you're using agents to build an educational platform? Is the content AI-generated or do they fetch it from somewhere? Trying to decide if this series is for me, would help to know.
Mr. Berman please learn how the composer view works and also you can just add the errors to the chat or composer you don't need to copy them etc. Additionally, Cursor has built in web validation with the agent just use the @web command. (Also there are a ton of other sick @ things you can do)
# sys.path: Insert the correct path to your module at the beginning of your script: import sys import os module_path = os.path.abspath(os.path.join(os.path.dirname(__file__), '..')) print(f"Module path: {module_path}") sys.path.insert(0, module_path)
great video, i build my crew from watching your tutorials, thanks a lot. and i'm wondering if you can make a video about how to integrate crewai with other applications, for example i have a nodejs application, and i want that to call the crewai app to do some work. how can i connect them two. i think this will be very helpful, since all the tutorial i can find are just about crewai. and i'm new to python, so really having trouble doing it. anyway thanks for all the amazing tutorials you make! love them!
Thanks Matt appreciate the great work you do! I’m wondering if this would be easier if done in flowise? I only have used it from tutorials I’ve learned from you and others but just a thought.
I am a pretty experienced python dev, but I have never had much with luck with CrewAI over the last year or so. Just gave it another go and another big fail. Just too many bugs in the tool itself and I end up writing my own agents. I really had high hopes for this.
Can someone explain what things like crewAI have over simply defining your agent in the system prompt of gpt4o for example and using structured output? Looks like an un-necessary layer
i didnt finish the vid yet but you shouldnt have changed the env because crewai install make its own env in .venv file and you just needed to install langtrace sdk in the .venv env but overall great video
I am also driven insane by Python virtual environment management. I must have the same packages installed in like 5 different places because nothing is ever available to the code even though it totally is. Is there a better way?
I rarely have issues with environment and packages, but I also never use conda, so that might be the reason. Still can't understand why people install heavy tool like this, when all you need is builtin in to the python itself.
I wish your pricing and package details were more transparent and easy to find. For example, a user shouldn't be required to create an account and login, just to see pricing. Also, each limited item in a specified package has no explanation of what each thing is and how each limitation would affect a business trying to use this tool.
excellent video, but what i have a problem with is that it's SaaS. your not only renting the software, your paying EACH time you use it, doesn't matter how cheap it is. been saying for 2 decades, "toot many people have more bucks than brains". might as well make a UI with a vending machine face. the 2 worse things to ever happen to tech. software as a service and cloud computing. hey froggie, how's the water?
So what I got from this video is that you currently need 25 years of Python experience to set up agents...given I'm just starting to learn Python I might wait until its a little more user-friendly
Can anyone kindly point me to any resource where I can learn how to install crewai from scratch on a windows PC I have been trying but it keeps giving me an error, please 😅
I ran into the same langtrace module error. I did exactly what he said. I deactivated, then removed the venv and recreated it and pip'd crewai and langtrace and the warning went away but I'm still getting the module not installed. I wish he didn't cut the video but I understand no one wants to watch 25 minutes of pain.
I think crewai doesn't work well with python environments. I had to create the project folder edupaper, cd into that folder. Create a python environment in that folder named venv(I don't think it matters the name) 'python -m venv venv'. Activate that environment 'venv\Scripts\activate'. Then install crewai in that environment, 'pip install crewai'. Then execute the command 'crewai create crew eduresearch'. Change to that directory 'cd eduresearch'. Deactivate the current environment 'deactivate'. Activate the created crewai environment '.venv\Scripts\activate'. (notice the .venv) You should see the confirmation it's activated by (eduresearch) on the command line. Check your desired python version is correctly being used 'python --version' . (if it isn't correct, check your path to the correct python.exe). Then test your agent app 'crewai run'. It should run. Now, the new crewai created environment will not have pip so you will have to install it 'python -m ensurepip --upgrade'. Then make sure it's the newest 'python -m pip install --upgrade pip'. Now you can install langtrace, 'pip install langtrace-python-sdk'. Now have Cursor open the folder crewai created, it will be the folder inside your 'edupaper' project folder, it's called 'eduresearch'. When you open that folder in cursor you will get the alert that the environment is activated but you will not see the (eduresearch). That bugs the heck out of me so I just activate it again with '.venv\Scripts\activate' (make sure you use .venv). That will make the (eduresearch) show on the command line. Install the langtrace import and the langtrace init line in main.py like Mr. Berman does in the video. You should not see any alert about package not installed. Then run the app in the command line with 'crewai run'. It should run with langtrace installed. This worked for me. There has to be an easier way, but this work for me. I tried to select the venv with Ctrl shft P but I coudln't get that to work. I would like to see if I can. I may try a few more attempts with clean installs and that method.
I've learned a new trick. Never ever lie to AI. You don't tell it it's an NPC in Skyrim. You tell it that it's playing the NPCs in Skyrim. This makes all the difference in the world.
My Conda folder is already 70 GB in size, Python is the worst f'en language ever, horrible crap. Yet we dont have any options, just need to buy bigger SSD driver, m.2 4TB may last few weeks for Conda...
it's so reassuring to see you run into the same frustrating errors that I get, but it's encouraging to see you work through them and persist.
I'm sure it's so very tempting for him to put up a corrected or version 2 run without his learning curve. I am so glad he resists this urge. Yes it would be much less tedious. But we are coders and users that need to see the thought process on how to correct our errors when we encounter them. Seeing his MO on this helps me understand how to attack errors I will be getting.
It is fun to use LLM's to ask why your code doesn't work.
@zakariaabderrahmanesadelao3048 May I sugest that you try Yacana ? Search for "Yacana github" on google. It's the fun without the pain. No silly errors like in crewAI. ^^
came here to say exactly this! ^ its like watching myself in the mirror....except much much faster and less sweary
It would be more encouraging if we had somebody who knew what they were doing.
Everything he is doing he has learned as he goes, just like many of us who use ChatGPT and Cursor.
I really love that you showed all the struggle working with unknown libraries and codes. Instead of performing the 'chef's trick' you show all the issues you run into. This is a reassurance that I am not stupid and even the best of us face the same problems.
For those of you who will try to follow this guide.
The problem is that we need a venv to install crewai and then crewai installs it's own venv... And Python gets upset.
Here is what I did. Once your project stops working, which is when trying to use langtrace. At that point, I copied all the source for my crewai project up one folder - into the folder where my original venv is located- everything except the .venv folder. I deleted the rest (the whole crewai project folder) Then I just re-ran crewai install in the folder where my src and .venv (the one currently activated) are located. And Voila! Working!
Hope this helps.
can you elaborate a little more on this please. I am new to coding, and have been to find my way through with gpt, but I still cant solve the langtrace problem :(
I own you a beer mate. I spent 1 day trying to make that langtrace working...
Even if it makes the video longer, I like when you keep your "try and fail" attempts.
TBH, I am not a long time viewer and think your AI content gets a bit redundant and boring, though that is how I found your channel. I get AI is all the rage and embrace profiting from it...
Then I watch this and learn that you are actually a skilled coder. This half hour video earned my respect. I hope you put out more content like this. Scrolling through the comments, it seems you once did. The AI testing is cool, but this kind of content actually brings value. An at least once a week version of this would be very nice to see. It would really make the channel stand out from the pack and give you a better position to be in when full saturation occurs on RUclips with this content.
As a coding hobbyist myself but much better engineer, I'm glad that a seasoned programmer has the same issues I run into, because I'm constantly thinking I simply don't know what I'm doing! It seems that turns out to only be half true. I actually like python until dependencies break everything. Why isn't there a simple method of making it so that every unique python script runs in it's own environment by default, relying on the dependencies file to pull both local and new dependencies, that get stored to a local pool. It's a pain to have to keep track of running virtual environments each time a script is run. Conda sucks! Isolate each script automatically and only attach the dependencies that are relevant. 1 day of writing code and 2 more of chasing dependency issues is not cool, especially when dealing with incompatible versions, etc.. There has to be a better solution for handling this chaos!
In real life, most of the work lies in the analysis, which then leads to an efficient code structure. One hour of analysis can save ten hours of coding. So, I usually spend a fair amount of time exchanging messages with o1-preview just to brainstorm and think the project through. After that, I follow the class diagrams, project file structure, and so on. That could make for a great idea for an upcoming video 🙂
Love your content, man! I’ve been following you almost since the beginning of your channel.
These are the videos I Missed from you, thanks!!
100% miss these videos . happy this is not just a AI news channel
Great video Matt! This one feels so authentic and hits home for me personally. I enjoy (don't take this wrong) seeing you making the same mistakes we all make, and working through issues, while working on real problems that aren't just Snake in Python! Cheers.
Hey Matt, thanks for doing near realtime AI builds with cursor and AI copilots. I honestly think that this content is some of the most valuable that you do, and yes, that means if it took paying to subscribe to it, I for one would pay to see one of these a week? Using different platforms and tools on a variety of use case scopes. Excellent content!
Yes...finally real use case video. I'm here for this!
When you are using cursor. I would highly recommend using cmd + k or cmd + i for ai help. cmd+k is inline quick fixing. will propose an inline diff of changes for fix. cmd+i is composer mode which is essentially the ai developer. you have detailed control over context and can specify files, folders, internet usage and more. also be sure to setup the stuff in "Cursor Settings" like indexed docs sites and cursor rules it will make your experience so much better
"Get rid of conda" - is always the right move
I agree!
What else? It is breaking my neck as well somehow.
Conda is sort of a necessary evil. You have bare metal, a VM, a docker container and then conda. Sometimes I don't want to create a new docker or VM for a clean python install, and conda does the trick. It's pretty handy for running a bunch of different environments or inference engines locally.
@@chrisstephens9460 I do not see Conda as being necessary for anything. Why would I want to use it for anything? I have managed to do everything I need and want to do without Conda.
Why are we running everything in Python anyway? It would probably be better to turn this into a react JavaScript based app.
Super-helpful to see your live tuning thoughts and troubleshooting to learn from you. Thanks! Sometimes that's more helpful than something you know will work.
Thank you so so so much for creating this super informative video! ❤
Really cool Video - it is good to know that i do not do anything wrong when i work with agents. It is not that easy.
🎯 Key points for quick navigation:
00:00 *📘 Project Overview: Educational Portal using Crew AI*
- Introducing an educational portal project to teach AI concepts from basics to advanced.
- Automating content creation for educational resources, including text, images, and step-by-step guides.
01:08 *🔧 Setting up Python Environment for Crew AI*
- Creating a new Python environment with Conda for project setup.
- Installing Crew AI and LangTrace SDK within this environment to manage dependencies.
02:18 *🚀 Initiating Crew AI Project and Configuring Files*
- Crew AI project structure setup, creating necessary files and folders automatically.
- Selecting OpenAI’s GPT-4.0-mini model for initial testing with low-cost, lightweight configuration.
05:02 *💡 Troubleshooting Python Environment and File Structure*
- Challenges with managing Python environments across different editors and configurations.
- Overcoming issues with LangTrace SDK installations and correct environment selection.
08:42 *🛠️ Integrating Perplexity API and LLM Model Selection*
- Testing the Perplexity API for automated research without web scraping.
- Experimenting with LLM models to improve research capabilities; troubleshooting 404 errors with API setup.
14:12 *🔍 Setting Up Real-Time Web Search with Serper Dev*
- Installing and configuring Serper Dev tool within Crew AI to access real-time web data.
- Testing Serper Dev API key integration and confirming working real-time search results.
18:04 *🌐 Testing and Refining Search and Reporting Outputs*
- Running search queries and generating updated, detailed reports using current AI trends and models.
- Transitioning from Perplexity to Serper for improved, up-to-date results in generated reports.
20:24 *🔒 API Key Configuration and Environment Security*
- Moving API keys to environment variables and ensuring they’re excluded from main files for security.
- Adding `.env` file configurations and confirming `gitignore` setup.
21:18 *📊 Initial Test Run and Model Performance Analysis*
- Testing setup with GPT-4 mini, analyzing initial cost-effectiveness, and identifying model limitations.
- Observing results on broad topics and considering more specific prompts for in-depth insights.
22:14 *🎓 Refining Agent Roles for Educational Content Creation*
- Adjusting agent roles: defining a senior researcher and content creator with tasks aimed at producing comprehensive educational materials.
- Updating agents' purpose and instructions to prioritize clarity and educational value.
24:17 *✍️ Customizing Output Requirements for Educational Content*
- Setting precise output expectations, like detailed, structured educational content over a simple report.
- Implementing Markdown formatting for topic breakdown and clear organization.
25:26 *📈 Report Testing with Real-World Content Examples*
- Generating reports on complex topics like retrieval-augmented generation, assessing quality and coverage.
- Ensuring content includes clear, structured sections like definitions, phases, and benefits.
26:07 *⚙️ Switching to Advanced Models for Improved Depth*
- Testing 01 Mini model for enhanced content generation, observing higher quality and cost tradeoffs.
- Reviewing performance on LangTrace, noting token usage and cost increases with the 01 model.
27:26 *🖼️ Future Plans: Reviewer Agents and Visual Enhancements*
- Planning additions like a reviewer agent to improve content accuracy and coherence.
- Exploring potential for dynamic graphic generation to supplement educational content.
28:06 *💰 Cost and Model Selection Insights*
- Analyzing cost implications of advanced models like 01 Mini, recognizing improved quality at a higher expense.
- Comparing model efficiency and value in the context of generating comprehensive reports for educational use.
Made with HARPA AI
where he use ollama ?
Hey man, always a great content!! Instead of "conda" use "uv" or "poetry" which are easier and modern way of managing environments/dependencies! Hope you use in your next journeys
I use VS + Cline + Aider but I do see that CURSOR has many benefits, especially all these "auto complete" features ! It was an absolute pleasure watching this video ! More like this ! 🤤Looking forward to part 2 !
There should be a name for these kind of videos that show the pain and struggle of actually using a software library. I'd watch the crap out of those for libraries I'm trying to use.
In your code editor if you use cmd + shift + p and type ‘select Python interpreter’ to set your editor up to use the environment you previously defined, you’ll avoid a whole lot of issues around environment management
Try Composer in Cursor. With your dev experience, it may not be as helpful to you, but I find it extremely helpful as a very junior, non-pro dev.
We need an agent that can take control over our phones.
Seems promising... I guess...
But still, CrewAi feels so random and doesn't work well with local LLMs. I prefer using Yacana to build agents. It can also do brainstorming even though the LangGraph like workflows feels way better.
Yeah, Yacana's workflows are quite good!
Yeah CrewAI, even with the monitoring stack is one hell of a mess. I would love Yacana to add the monitoring though. Alhough it's less of a headache than any other frameworks out there, monitoring is always useful...
Thanks for the interesting content! Cursor is great but becomes perfect when you give the right context to every question. You would then not have to go to an external source to fix your issues.
@@Jelckeb I figured that out in part 2 :)
There's a part 2? Oh I see now, this is labeled part 1, ....*goes hunting for part 2*
@@matthew_berman what VScode exetenson are you using for chat ? and is it the same as for code completion (in-line editor) ? is it Cursor ?
I recommend using Poetry instead of Conda for creating virtual environments in Python, as it makes the process much easier and more efficient. In complex ML/AI projects with many dependencies, Poetry often provides a smoother and more reliable experience due to its robust dependency management and streamlined workflow.
Python environment management can be made a little easier with tools like Miniconda, as you can easily choose which versions of Python to use for each of your environments. Some projects like to follow the standard Python methodology and make a .venv folder within the project, in which case the environment is activated using “source ./.venv/bin/activate” and deactivated with “deactivate.” Sometimes it’s a good idea to make a conda environment even if the project will contain a .venv folder, so the version of Python used is not just the system version. If it turns out the conda environment is redundant you can just delete it, but some .venv contents might just be symbolic links to the conda environment, so check that first.
i love videos like this.. helps me understand how dev really works
FYI - You can set the temperature and the context length in the openai APIin your Python code. Gpt40-mini works great for most things just lower the temp and give it the models max context length, your get more accurate and longer responses.
this is what's up matthew !! thank you !!
I’ve been using Poetry for Python environments and I like it a lot.
When I'm testing something, I use `venv` module.
But when I want to make a "real" application, I use Poetry.
I believe that your issue is related to the IDE used. That's true I don't use VS Source, I use Pycharm that understand external environment created with conda/pip/etc ....
Virtual Environment Issue: The virtual environment issue exists because we create a venv with conda, then when we run crewai install, it creates another virtual environment. A simple fix to this is:
- within your project folder, edit the file .venv/pyvenv.cfg
- Change the line that says include-system-site-packages = false to include-system-site-packages = true
This will cause the .venv created by crewai to reference the conda environment packages, which is where pip is installing the libraries.
Great video Matt, thanks for sharing
Thank you for sharing. I was recently working on a crewai project with vs code. I was amazed how you see these traces and costs analysis? Is it the IDE you are using or any plugin? It will be very helpful for me to use it.
I seem to be one of the few that did not find my own solution to the langtrace error. Could you lend any advice for fixing that issue? Love these videos. Super excited to get CrewAi running myself.
Awesome video Matthew, really liked this format! As for the python env situation I also had a bunch of problems with cursor so I stuck with VSCode, but eveytime I watch a cursor video I feel the urge to change back and give it another shot 😅 I'm also trying to change my python env management to uv, its much faster and simple ( crewai uses uv to manage the envs and installations ) . Keep it up Matt, looking forward to the rest of this series!
We often use docker containers to isolate our environments at work.
I've seen people and I've used Dev containers in VScode. I like the docker for it's shareability. It's just a little more work to get each project set up. But when you do it's really superior to other methods.
at least you woudnt spend hours fixing problems just to test the project, that is also needs hours of fixing :)
«workspace» is likely from uv (separate packages in the directory tree that share common uv.lock (compatible dependencies version)).
So you're using agents to build an educational platform? Is the content AI-generated or do they fetch it from somewhere? Trying to decide if this series is for me, would help to know.
Copy and paste the transcript in gpt and ask it
Mr. Berman please learn how the composer view works and also you can just add the errors to the chat or composer you don't need to copy them etc. Additionally, Cursor has built in web validation with the agent just use the @web command. (Also there are a ton of other sick @ things you can do)
I've been thinking about trying some agent type things cross model. Deciding between CrewAI, langchain, or OpenAI swarm.
hahaha been down that Python rabbit hole , I was thinkin' it was me , thanks for sharing your frustration.
# sys.path: Insert the correct path to your module at the beginning of your script:
import sys
import os
module_path = os.path.abspath(os.path.join(os.path.dirname(__file__), '..'))
print(f"Module path: {module_path}")
sys.path.insert(0, module_path)
great video, i build my crew from watching your tutorials, thanks a lot. and i'm wondering if you can make a video about how to integrate crewai with other applications, for example i have a nodejs application, and i want that to call the crewai app to do some work. how can i connect them two. i think this will be very helpful, since all the tutorial i can find are just about crewai. and i'm new to python, so really having trouble doing it. anyway thanks for all the amazing tutorials you make! love them!
most or all the demos I've seen is about some research. what else can we do?
If you know about Kubernetes there's a funny project called Gno6 built using Yacana. Search for the repositories. It's real world examples.
Thanks Matt appreciate the great work you do! I’m wondering if this would be easier if done in flowise? I only have used it from tutorials I’ve learned from you and others but just a thought.
I am a pretty experienced python dev, but I have never had much with luck with CrewAI over the last year or so. Just gave it another go and another big fail. Just too many bugs in the tool itself and I end up writing my own agents. I really had high hopes for this.
Can someone explain what things like crewAI have over simply defining your agent in the system prompt of gpt4o for example and using structured output? Looks like an un-necessary layer
Matt, when are you going to teach starter/101 courses, and in slow motion ? ;-)
How do you make Python environment management easier? Don't use Python.
i didnt finish the vid yet but you shouldnt have changed the env because crewai install make its own env in .venv file and you just needed to install langtrace sdk in the .venv env but overall great video
thx matthew b.!
he didn't say he would delete the api key after the video! haha, he forgot to do that. has anyone used his key yet?
I deleted it
Matthew can you put in the notes a tutorial to set the python environment up? I am sure you did one a while ago.
I am also driven insane by Python virtual environment management. I must have the same packages installed in like 5 different places because nothing is ever available to the code even though it totally is. Is there a better way?
Its funny that after letting GPT read the ToS of CrewAI, it shows some dodgy renewel and Telemetry usage. Im not touching that with a 30ft pole.
Clearly, AI is not replacing programmers any time soon
thx mat!!
I rarely have issues with environment and packages, but I also never use conda, so that might be the reason. Still can't understand why people install heavy tool like this, when all you need is builtin in to the python itself.
install anacoda for env management , super cool
Is it possible to use Open Router API?
This is exactly what it looks like when I do anything except I'm faster at getting Ai to do it because I don't know how to code shit.
I wish your pricing and package details were more transparent and easy to find. For example, a user shouldn't be required to create an account and login, just to see pricing. Also, each limited item in a specified package has no explanation of what each thing is and how each limitation would affect a business trying to use this tool.
Video request: a personal RAG / 'second brain' based on customised input data and unlimited input tokens (or very high limit)
after all this hassle to build a agent , do you still believe ai is replacing software engineer any time soon?
4:15 Docker. Isolated environments with network connections.
Pyenv works fine
I need an agent to do what you just showed me
lol
How much would you pay for that?
Please do a show on voice assistants, none of the demos live up to the hype.
Cursor FTW.
Are you sure you are saying Mammouth correctly? xD
Yes, I triple checked with them lol
Cursor is vscode!
excellent video, but what i have a problem with is that it's SaaS. your not only renting the software, your paying EACH time you use it, doesn't matter how cheap it is. been saying for 2 decades, "toot many people have more bucks than brains". might as well make a UI with a vending machine face. the 2 worse things to ever happen to tech. software as a service and cloud computing. hey froggie, how's the water?
Different strokes for different folks - chill out.
@@mjcjrfl77 grow a skin.
So what I got from this video is that you currently need 25 years of Python experience to set up agents...given I'm just starting to learn Python I might wait until its a little more user-friendly
Hi @matthew_berman what VScode exetenson are you using for chat ? and is it the same as for code completion (in-line editor) ?
Crewai install ??? Is that a cursor thing, as you had already done a pip install of crewai into your virtual environment
Poetry is my choice for python env manager
OH MY GOD
Can anyone kindly point me to any resource where I can learn how to install crewai from scratch on a windows PC I have been trying but it keeps giving me an error, please 😅
I ran into the same langtrace module error. I did exactly what he said. I deactivated, then removed the venv and recreated it and pip'd crewai and langtrace and the warning went away but I'm still getting the module not installed. I wish he didn't cut the video but I understand no one wants to watch 25 minutes of pain.
I think crewai doesn't work well with python environments.
I had to create the project folder edupaper, cd into that folder. Create a python environment in that folder named venv(I don't think it matters the name) 'python -m venv venv'. Activate that environment 'venv\Scripts\activate'. Then install crewai in that environment, 'pip install crewai'. Then execute the command 'crewai create crew eduresearch'. Change to that directory 'cd eduresearch'. Deactivate the current environment 'deactivate'. Activate the created crewai environment '.venv\Scripts\activate'. (notice the .venv) You should see the confirmation it's activated by (eduresearch) on the command line. Check your desired python version is correctly being used 'python --version' . (if it isn't correct, check your path to the correct python.exe). Then test your agent app 'crewai run'. It should run. Now, the new crewai created environment will not have pip so you will have to install it 'python -m ensurepip --upgrade'. Then make sure it's the newest 'python -m pip install --upgrade pip'. Now you can install langtrace, 'pip install langtrace-python-sdk'. Now have Cursor open the folder crewai created, it will be the folder inside your 'edupaper' project folder, it's called 'eduresearch'. When you open that folder in cursor you will get the alert that the environment is activated but you will not see the (eduresearch). That bugs the heck out of me so I just activate it again with '.venv\Scripts\activate' (make sure you use .venv). That will make the (eduresearch) show on the command line. Install the langtrace import and the langtrace init line in main.py like Mr. Berman does in the video. You should not see any alert about package not installed. Then run the app in the command line with 'crewai run'. It should run with langtrace installed.
This worked for me. There has to be an easier way, but this work for me. I tried to select the venv with Ctrl shft P but I coudln't get that to work. I would like to see if I can. I may try a few more attempts with clean installs and that method.
@@TimothyHuey Thanks camarade, you save my day!
*comrade :)
For me it have to be open source only.....
Docker is to go tech for enviroment
No chapters?
Matthew, I would like to have had a review at the end where you walked through what you had to do to get it to work.
I would also like to see any hardware requirements when you run local.
No Timestamps? :(
you're in 2160, my guy
There needs to be an agent to setup all the crap
Instead of perplxeity you could have use zero agent.
I have never gotten conda to do anything useful. In fact, I have never gotten it to install.
The problem because press that env button in cursor
oh wow....
I'll just wait for ready-to-use agents
Poetry - use nothing Else for project setup. Not perfect but ok also for multi package repos
I've learned a new trick.
Never ever lie to AI.
You don't tell it it's an NPC in Skyrim.
You tell it that it's playing the NPCs in Skyrim.
This makes all the difference in the world.
I use Anaconda navigator to manage my environments and have never seen the env errors.
Yoooo
My Conda folder is already 70 GB in size, Python is the worst f'en language ever, horrible crap. Yet we dont have any options, just need to buy bigger SSD driver, m.2 4TB may last few weeks for Conda...
Second for the first time
This entire video is a huge waste of time. A bunch of my limited time that you sir have wasted, with nothing of value to show for it.
First comment for the first time...
not great