Видео

Now how to use this streaming data (logs) from kafka with llm ?

What fart with the auto translation

why you bald?

bro, how you lose all your hairs? is there a secret?

hi excellent demo, can you also demo the chat history functionality of chainlit without using Literal ai?

arrow was pointing to the right... not left

Just for fun, I guess there were a few print statements in check_connectivity and toggle_wifi("off")

I need something more barney style

closed source is so 2020

Great Tutorial. How does Phidata compare to crewai when building ai agents?

It truly sucks for OCR for LaTeX, it feels embecile to give it instructions and always gets them wrong...Qwen2-VL (72B) Instruct is better, but still far from perfect... sadly

This is exactly what I was looking for! Thanks a lot. Could you also provide a Github repository that can be clone to play with it?

GPT-4o recommend this video and it was 100% correct.

i'm getting this error from a llama3.2:1b-Instruct, ollama create -f Modelfile myllama transferring model data 100% converting model Error: json: cannot unmarshal array into Go struct field .model.merges of type string

Thanks for the video! I had success running this with "mlx-community/whisper-tiny" and "mlx-community/whisper-turbo" but anytime I tried any of the other models on the huggingface page I ran into this error: "TypeError: audio_features has an incorrect dtype: mlx.core.float32" Has anyone encountered this issue?

Thanks for the great video. I have Ollama version 0.5.5. So theoretically these parameters should already be set and I also should be able to change them before launching Ollama server. However no matter what I set for these parameters I see no difference in GPU usage. Is there a way to actually check how many models are loaded in GPU?

This is very hard to follow. You don't explain the intent and purpose of what you're doing or show what files you'll be working with, let alone *why* you're doing what you're doing. Only someone who is either very familiar with these scripts and coding for AI or very patient and determined could follow these instructions. Please, go slower and explain more.

can i know which technique is used for orchestration of the agents : sequential or hierarchical

Hi Mark, 'style_pandas' is missing in repo, could you upload it please? BTW I like your channel - great topics in short videos and on top code examples - the best suits my needs. THANKS!

My favorite youtuber and favorite library, what can go wron here? hahaha

Mark - you are the best.

I just came across your content on RUclips about re-ranking. Great work, and I’ll be following you from now on. Wishing you all the best!

Amazing one!! Could you also share the commands and sample data files. Would be good to try these 😃

So happy to see Phidata here! Thanks a bunch for this! 😊

Wow! Thats amazing! Are there similar performance gains with normal llm models in mlx format as well? And what about stable diffusion models?

what human isn't using windows....

Great 👍 (after playing with transformers and cpu offloading, searching for ideas i found ctransformers and realized that ollama is no magic at all, its just gui 😅)

Can phidata have enough to memory to talk to me months later of chatting and reflect and advise future ideas? Like giving it enough info of my teams progress and what new AI tools I may need to speed up the process to fulfill the deadline? Or maybe personal bot conversations? I figured ai had this already with GPT having so much data to reference but learned that it’s behind babyAGI by a year just now introducing chain of thought and doesn’t have good memory within the chat … is phidata + GPT + maybe lang chain or something (for non coders and coders alike?) have a way to give the AI very long memories? Really need to know lol.

The llm does a correct job in this case, and I'm happy that the model works well with the prompt. You get a different answer when you ask the question differently, because for each question you retrieve different chunks from the vector-store. The chunks for the first question don't have the answer. By default, the similarity_search returns k=4 documents. In your second question, the third chunk shows the answer: "disrupted by rain.The tournament......" This text-chunk is in the top 4 chunks similar to the input question. For the first question this chunk is not chosen, probably there were some other more common words of other chunks similar to the question. Idk yet what is an optimal choice for k, or how many clusters to choose. But you would get a better result if you increase k. Still keep in mind, if you make k too large or the size of the text chunks is large, then the model loses context and "does not remember" everything you pass to it. Or maybe it's a good idea to search differently, and decompose the input question into tokens that we can best align with a retriever.

I think models don't have people face

Phidata vs pydanticAI?

Can we use any database other than these? Postgres Sqlite Singlestore DynamoDB JSON YAML And if yes how, I would appreciate any help.

Your Formula for code programming the LLM is shit, I'm a master computer engineer and what you're doing is completes bullshit. let's see if you can handle what im about to say to you. Mathematical Inconsistencies: The provided code snippets and parameter reductions do not add up, indicating a potential fundamental flaw. Deviation from Established Best Practices: The approach contradicts Hugging Face's documented guidelines for parameter optimization, which emphasize balanced trade-offs between complexity and performance. Lack of Empirical Evidence: The presentation fails to provide conclusive experimental data supporting the claimed efficiency gains. The Elephant in the Room: Running LLMs on Laptops without GPUs Warning: Attempting to run Hugging Face LLMs on laptops without dedicated GPUs is ill-advised and likely the root cause of: Code inconsistencies (e.g., errors in the code modifier) System performance issues experienced by those trying to replicate the approach 10 Compelling Reasons to Run LLMs on Computers with Dedicated GPUs: Computational Intensity: LLMs require massive parallel processing, which GPUs are designed to handle. Memory Constraints: Laptops' integrated graphics often lack sufficient VRAM, leading to memory overflow errors. Thermal Throttling: Laptops' cooling systems may struggle to keep up with intense computations, throttling performance. Power Consumption: Running LLMs on laptops can lead to excessive power drain and reduced battery life. Optimization Limitations: Without a GPU, optimizations are severely limited, compromising model performance. Scalability Issues: Laptops lack the scalability to handle large LLMs or multiple model instances. Dependence on CPU: Relying solely on CPUs for computations can lead to significant performance bottlenecks. Increased Risk of Errors: Insufficient computational resources increase the likelihood of errors and inconsistencies. Incompatibility with Future Models: As LLMs evolve, laptops without GPUs will become increasingly incompatible. Wasted Resources: Attempting to run LLMs on under-equipped laptops can result in wasted time, energy, and computational resources. Mathematical Demonstrations: To illustrate the importance of proper computational resources, let's examine two examples: Example 1: Matrix Multiplication (a fundamental LLM operation) CPU (Laptop): O(n^3) time complexity, where n is the matrix dimension GPU (Dedicated): O(n^2) time complexity, leveraging massively parallel processing Example 2: Gradient Descent (a crucial LLM training algorithm) CPU (Laptop): O(m \* n) time complexity, where m is the number of samples and n is the feature dimension GPU (Dedicated): O(m + n) time complexity, exploiting GPU-accelerated linear algebra libraries. Conclusion: Running Hugging Face LLMs on laptops without dedicated GPUs is not recommended. The presented approach is flawed, and the lack of empirical evidence undermines its credibility. For optimal performance, scalability, and reliability, always deploy LLMs on computers equipped with dedicated GPUs. Rating: -5/5 (due to the identified technical inaccuracies and disregard for best practices). In literal terms You're computer skills are garbage and dogshit, How about you go back to a normal desk boy job or pushing mail cart's or whatever it is you do and leave the engineering to me and my team of more advanced engineers then you.

ValueError: could not broadcast input array from shape (384,) into shape (1024,)

Very nice tutorial, thank you for curating and creating these content mark,

Phidata is by far, the easiest way to create agents, thanks for sharing this amazing tool

I’ve tried setting up everything correctly but still can’t achieve parallel responses. I’m testing this by hitting the same cURL request for the same model from two different places (Postman or directly via terminal). Here are the steps I followed: Stopped the service: systemctl stop ollama.service Set the environment variables: export OLLAMA_NUM_PARALLEL=4 export OLLAMA_MAX_LOADED_MODELS=4 export OLLAMA_MAX_QUEUE=2048 Verified the variables: env | grep OLLAMA OLLAMA_MAX_LOADED_MODELS=4 OLLAMA_MAX_QUEUE=2048 OLLAMA_HOST=My-localhost:11434 OLLAMA_NUM_PARALLEL=4 Started the server: ollama serve Despite this, I’m still unable to achieve parallel responses. Could you please help me identify what might be going wrong?

Thanks a ton man... outstanding video and tutorials

Great video! I'm also exploring multi agentic frameworks, on my top list there are ag2ai and PydanticAI, which IMO have great motivations and potential to grow. BTW, the LanceDB part of your git code requires OPENAI_API_KEY, I don't have it so I skipped this part.

Would it be possible (In SQL) to have a function that combines GROUP BY and PIVOT into a single CROSSTAB() function?

How can we setup a local RAG which can redact PII information when extracting the data from the PD before creating the chunk, enrich the chunks by using a LLM to create a title, summary, keywords, embeddings and then use PhiData to chat with the document?

Yes please, need at least 10+ videos series for phidata, so much to explore. Also swarms (not openai) have great multi agents architecture. For web3, I recommend eliza by ai16z.

Nice video, thanks! Have you tried swarm (OpenAI) or pydantic agent before? if yes which one do you prefer compared to PhiData? I am looking for a library to easily create agents, but that it's simple to use, light weight and that is not so cluttered and doesn't obfuscate or over abstract everything like CrewAI and Langchain do.

Nice to see the LLM lord is turning into the agentic side! Good stuff!

Absolutely fantastic for people who use mac!

Nice

Damn you computers fast mine took 5 minutes to respond at 0.07 tokens/s

why you look like mahatma gandhi?

Just found your site. I’m learning so much. You do a great job explaining things. Thanks for all you do to help people.

I really love to see your way to explain this, Thanks a lot!