Відео

If you could choose one AI agent framework that you believe will become the industry standard, which one would it be?

Thank you sir

Great video! I would definitely be interested in seeing some evaluation methods. Cheers!

I really love your videos <3

very good , i was not actually sure how to compile up a docker : !!

Dude is an ai video himself

How dare you compare this to perplexity that a click away it literally took you an hour to explain how to use your tool

Ive been away for a while, not by choice but excited to catch up on my favorite channel, this looks great so far. You are doing exactlt what i wanted to do am advanced long form output research agent so this is great wish i could sponser but im hoping i can in 2 to 3 months when things are going better

WOW. glad you popped on my feed ! Question : is it possible to change the system prompt ?

Great work

Very cool, subscribed

Adopting the Chainlit task list definitely enhances the responsive feel but given the slow speed of o1 I’d recommend generating a unique project id that you and the agent work on and use the webhook session restore to report status messages. Each message can have preset actions that replace your slash commands. That will enable you to offer feedback on the fly, instead of aggregating all the feedback into a single message. That slightly increases the complexity, more or less given the construction of your langgraph state classes. I cloned the repo last night, hopefully I’ll have a chance to look at the code and see if there’s any PR I could do.

It is awesome, I am going to contact you once I redo my project with the new concepts. I think I can use your help.

Awesome work here! 👏 For those of us that are not Blessed with having OpenAI Tier-5 access, is it possible to offer an OpenRouter setting in this project so we can give this a try? 😁

Anyone can get access to the o1-preview model via open router

can we run this with local llm olama models?

Thank you for your breakdown. I found especially useful to understand which model is better for what task in an agentic framework. I'm guessing o1 will also be quite expensive on top of the slower speed, so it should be reserved for mostly complex tasks where they need to follow steps to the T

Lovely as usual : I think you need to make the decision on the embedding database and rag setup ! , is it over use ? are you using a hammer when a screwdriver is required ? As you know rag is an augmentation to creating addtional content for the model submission of a query : But it can also just be a simple knowledge base also : hece the rag itself can be a tool : ie a knowledge base for the model to search for infroamtion before doing a web search : So when a task is placed it can first search its knowledge base then search the web ! if required : So we should be thinking that we may need an additional service ! this is lets call it the web crawler service : SO as researcher you will know your topics that your interested in : so you should be able to put the model to work searching the internet for topics and creatig a mini collection of documents to be processed into the rag for use off line ! this will enable for local research !( SHORT TERM MEMORY) so when youo ask a question regarding topics you have rresearched they will be already held locally ! ( also you can add a date -Marker ! this will tell the model to seek updated knowledge if the current data is more then a month or two old ! ) ... hence blocking thr web search and keeping the work local ! this gives you the advantage of faster quering! , as you have a rag and it should be used as a knowledg base ie coolecting your browsing historys and merging them to the rag ! so also perhaps a mini browser in app to display pages ( perhaps converted to markdown ) ( aha a markdown browser automatically convverting webpages to markdown !!) ... this ienables you to save your pages after searching as markdown pages and even to a search dictionary (json ) and later fed to the rag or Fine tuned in to the model ! SO now the app becomes a true research tools ! So you should also add more rag tools to retrive various formatted documents to your UI ! SO ow you can have a seties of good tools to acess and manage your knowledge base and other for searching ! ie local or online ! or for knowledge building ! Or Report creation tools to display your rih information in the ui , or even DataBuilding ( corpus building Feature ) to build corpusses based on various searches in your reseach stack to produce docuemtns or database for fine tuninng into your model ! ( hence being able to utilize the rag as a Trasactional database as well as alocal knowldgebase enhancement ) we would exxpect that trained data can still be held in the rag but should be marked for archiving : as the modle should have the knowledge internally ! So on Rag optimization the model can remove these entrtyu from the rag as they are now in the pretrained stack ! IE THE LONG TERM MEMORY.... Obvioulsy your chat historys should always be optimizes by sumarization and also uploaded into the rag , before closing ! hence your WORKING MEMEORY become Uploaded in to your SHORT TERM MEMORY ! .... } So he we implement the full Memeory Stack Also ! this also enables for personalization of the model ! As local history and commuication and local knowledge is being optimized ito the rag : as well as being outputted in to a datset for training into the pretrain Stack: full complete pathway ! I specify again the most importsant thing is to train a model local model for Planning ! this really changed the model for me even after a year of training all these methodologes the modle seemed comparitive to larger models despite being 7b ( only can be run local for me ) < SO i think it was sucessfull ! but when i trained the model onn the react process and its clones as well as PLANNING ! the model Changed ! : So now when i used it for Mastro and aider etc as all the models i found now its planning really augmented its coding and tasks ! as well as its responses for normal nformation al querys !... I also implemented the repl tool ( runs notebook cells ) ... which also made the model always echeck the output before returning the code ! , Every thing changed .......Now its totally on point ! I think these were the missing components : As well as i changed my strategy : all of these think, action, observation etc these stages are all just prompts ! , SO a single self query tool is all you need and the pre wrtten prompts : these entry points are used to create agents ( or preloaded tools ( self-Query ) ) so the model can use the prompt as a tool ! LOL !!! << KEY ! So if i explain a type of execution strategy it can create the neccasaryu prompts ( for these tools ) ... Now a final Response can also be a tool , as the tool will format the response and even execute all code snippets ! As i added this to my input and output loops : so before i send code to the model , ( i used the markdown tag ) it will execute these snippits and pass the executions to the model ! so it will see the output from these codes and any output it sends to me will also be executed ! This interception also changed the model behavior ! to really being great with coding ! (python) ... Obvioulsly i also spend a lot of time collecting DOCSS from librarys ! and feeding these docs to the model ! <<< , So i do not use rag ! , but i understand the methodolgy and how it should be utilized for the full working memory, short term memory , and long term memory as i am also a Business inteligence Data Scientist ! (BSC/MA) ...( another day i will go over the data mining strategy ( web mining ) - ) I hope this hellp you to percive new directions for th app ! as you did enlighten me and force me to make the graph ( from scratch ) and design the chain! and begin to use state for every thing ! Which also Raised the BAR ! ... you did not convert me to RAG ! ... LOL !

Hi @Data Centric. Do you plan to add a video on how you can make two (or more) agents talk to each other to fulfill a task from scratch? :)

EIC not applicable to UK companies.

I've been away for a while, and I'm thrilled to return and see the amazing progress and new features that have been implemented. It all makes perfect sense, and I’m excited to try this out. Thank you for your hard work!

It's available on openrouter...

Nice love your work.

If you want access to openai's GPT o1 model via api and dont have the right tier you can go through openrouter and access it now.

That sounds really interesting! I'll spin up a copy and give it a try. I've been meaning to explore integrating Neo4j with OpenAI, hould be fun to see how it goes. Thanks!

Thank you for sharing this and explaining it in such detail. I get really excited when I see a new video of yours pop up on my subs tab. Curious, how big of an improvement do you think this is compared to just using gpt4o for all the agents?

Love what you are up to with Jar3d.

can i use this with ollama?

Great video! You did an awesome job explaining the concepts. I just wanted to clarify one point regarding LangChain. It already includes libraries for GROQ and local models. For GROQ, we can use langchain_groq, and for any other local model, we can use: ChatOpenAI(api_key="ollama", model="llama3", base_url="localhost:11434/v1") Given this, I'm curious why you opted to create custom code under /models. Just want to make sure I'm fully understanding your approach.

Question what Hardware do you use ?

Building AI agents from scratch involves designing algorithms that can learn and make decisions. Start with defining the problem, select appropriate machine learning models, and train them using relevant data. Simplify the process by breaking it into manageable steps and using frameworks that streamline development.

Excellent explanation!

Great video! Thanks for the well-structured code and clear explanations

knowing what you know now about Langgraph, does your position change re: your suggestion/preference to build custom workflows from scratch? I recall you mentioning that preference in your CrewAI tutorial & you don't mention that at the end of your video here. Has Langgraph converted you to frameworks?

Great Video!! I wanted to check how can we show these tool calls and other things as Chainlit steps?

very informative video!

re you planning to use Neo4j exclusively for your project, or do you still require another database for semantic search? I asked ChatGPT, and it seems that while Neo4j can handle semantic search, it may not be optimized for quick searches if the database becomes too large.

I think it would be great if we/community could implement the ability to upload PDF documents for chunking and adding to the knowledge graph. I have not had any time to thinker on adding features but I've been playing with the system prompt and graph structure.

Good Day Sir, I am interested in using a 70B model and at Timestamp 58:50 is where you discuss other models to your project , so is it possible to add new models without breaking the code? Appreciate any feedback.

You don't reply to issues on GitHub. That +is+ necessary ... please :)

very good ! ( but ... Think Should be a tool ) ( observe should be the return tool call ) ( action : should be the selected tool or action ) ( think : should be a tool to self query for a plan or methodology or how to porceed to the next step of a plan ? ) .... This way we can create Any methodology or chain of thoughts : if we make the components Tools themselfs .. such that : create a plan ( should be a tool ( which is a prompt , to query itself with a task , so given a task as input it will return plan ) < the prompt in the chain manages this ( tool ) .... Refine Code ( this can also be a tool : As it will query itself again, with the code for refinement ( the only chage is the prompt ... Tool or chain , or Graph this will always be the question ) ... So we can have a refiner prompt : the rifiner prompt could have a set of feirce requirements to folow and use for correcting code : or indsutrialising code etc :; So when we design processes etc for our agent we can also have template plans in the planner ! < the user would like to repair some code : so the planner would offer the tool choices and even disribute a boilerplate , from a set of boiler plates : ie a code planner could have many basic template to begin a project : Supplied to the model as a list of docstrings to select a template it would call the tool to return the template from file or disk storage : ( ie tools have acess to the required system and files and do not go outp=of bounds ) ( open code execution and open syystem command executions should always be human intervention as it can destruct your system : Hence Tools istead : they have boundrys : A planner could also have acces to tool collections , which can be loaded to the agent which will perform the task : ie pass a toolbox to the model ! Hence your central controller bot will access all these tools ! given a menu of tasks , which return tool collections and plans and Start nodes : allowing the model to pass a custom state to a chain and retun its output : ie create an app which does : it would ask its general planner bot : who would tell it the correct planner to begin : then that palnner would give the agent the tools and requirements and the start node ! alowing for the agent to execute the correct Graph ( tool:) with a state and expectation ! I built this process the same way as you : But after i realised the whole above discussion : and then i realized even higher perspective : that our front end should be the same as dialog flow : as this add the personality layer ( it may not be intelligent but its the constrainment we need on top) we are currently working at a very low level and should realize that we have many layers on top of this : not a UI ! between the react process (or thought process and selected methodology for the task : as the react is very good for long and indepth tasks with a graph based structiure ) but not good for general chat ! hemmce we need the front end on top of this layer , a very simple , keyword and detect/Response method : a dialog manager : as eve the rag should be plugged in at this higher level : as it is not reuired unless its for indepth tasks : as basic chat history can keep your model personalised and up to date : the rag is a short /long term memory: whihc after optimization can be finetuned into the main llm : but it still does form a part of the reseach and content process when required ( it should be a tool ( only used when required ) ! ..as a State is enough to use to perform a task and chunking and embedding for simularity should oyu be for indepth querys ! not for dowloading and sumarizing ! <the model does ot need that to sumarize data ) ( a sumarizer is only a prompt chain away : only a prompt is required ! ( so a self query tool again !) ... SO sub agents are only tools which self query the model , with clean chathistory only for the execution of the tool : only the exchange between agent and tool is saved : hence internal processing in a tool is not returned to the agent only the final output : ( it ca be shared using verbose ) ie create two reponses a verbose ( lots of lovely logging , and custom trip wire exceptions to incrase loggig potential ) or just a simple output : peserving the minimun token exchange between agents : , ie oly query , state and output... Tools ca be conditional nodes also ! so that given a input to a Tool which functions as a condotional node or routerr it will be able to have branching inside the tool ! alowing for a refiner to call a coder tool : to return the refined output to the calling agent ! ( so it branched to the coder then back to itself as the coder node always returns data to the refiner tool ! Unless specfically (boolean) told to return the output without refinement ! or we would be hiting odes whic cannot return diredct output : hence many nodes in a graph can be executed as tools: Really i hope you understand this post is an enhancement to your video and perhaps your next thought pathway: You may even find it is very less intensive on memeory as the tool unload after execution ! and your actual agent can be a very smal 1b model as the tools are the intelligence : so if you will use the saem model a moderate b is required as you will be sub calling the same model perhaps twice ( in the case of the refiner , r planner which may call a sub planner ( carfully prompted ) ) ... SO aGod 7-14b is the best model to use ! ( the newest model have been trained on function calling and tool use and planning ) !! << Concept Overview: When creating graphs or decision trees in programming, the key components are nodes and edges. Here's how they function: Nodes: Represent states, decisions, or actions. Edges: Represent the paths or transitions between these nodes. The structure typically starts with a root node and ends with a final node, forming a tree of nodes. The edges are crucial for drawing routes, checking if paths exist, or determining if they are complete. However, the actual execution of the tree doesn't depend on the edges but rather on how each node points to its subsequent node, including conditional nodes. Key Points: Execution Flow: The tree executes by starting from the root node. Each node, possibly through recursion, leads to the next, ultimately reaching the final node. Between nodes, you can pass a State-a data packet that each node processes or modifies. The final node then returns this modified state as the output. Graph Structure: Edge Map/Matrix: This is a list of lists or a dictionary that tracks all possible edges (paths) connecting nodes. For example, the edges connecting node A, B, C, etc., are mapped to allow traversal and application of algorithms like breadth-first search (BFS), depth-first search (DFS), or shortest path. For visualization, this edge list can be used to draw the graph, but the actual logic lies in how nodes are connected and executed. Graph Construction and Execution: The primary node is often an Intent Detector node, acting as a router that determines the path based on the user's intent. Each task, whether it's web research, essay writing, or coding, would have its own planner that the graph follows based on the route chosen by the intent detector. Conditional Nodes: These are special nodes that determine which path or sub-tree to follow based on specific conditions. Additional Considerations: Start Nodes: These nodes initiate the graph and can serve as entry points for different UI elements or triggers. Multiple start nodes allow for flexibility in how and where the graph begins, accommodating different types of user inputs or scenarios. Chatbot Integration: The chatbot front-end can enhance input by detecting keywords or context, thereby prompting the LLM (Language Model) appropriately. The front-end should dynamically load and unload tools based on the conversation's needs, ensuring the LLM remains responsive without being overloaded with unnecessary tools. The chatbot can also handle certain tasks directly, without needing to involve the LLM, to maintain efficiency. Final Thoughts: This approach blends traditional chatbot systems with modern LLM-based tools, creating a robust framework that combines structured dialog management with the flexibility and power of LLMs. For long-running models (e.g., 24/7 services), it's important that the graph returns to a start or reset node after completing a task, ensuring that the system is always ready for the next input. This integration of old and new methods will allow the creation of intelligent, personality-driven agents that can handle both structured tasks and dynamic problem-solving, all within a responsive and adaptive system. the point is that you dont need a langGraph or a Neo4J as well as they being the exact same thing but slightly customized specifcally for its purpose ! If you deploy jared instead as a Space in Huggingface or Runpod : then you can design a mini UI which hits your Space ! Like a chat gpt .. hence the model woul not need to tranfer so much data and you would have the speed of the space! your only processing will be in your endpoint ! so you will be using it like a perplexity and can now compare its speed and results side by side ! browser to broswer or endpoint query to query from the ui ! Sorry for the long post but i have been thinking about it indepth as you see !

Workplaces can change when AI and analytics are combined. For this, have you looked into how SmythOS and other platforms improve AI agent collaboration? #SmythOS #Aitools

What is your thoughts on integrating GraphRag or HybridRag?

If you don’t mind, I’m wondering about the use of LayoutPDFReader with a URL. Isn’t it designed for PDF files?

You named it after me! Thanks man, it's even pronounces the same ❤

Thank you. Certainly explained quite a few things that the online documentation is missing.

Love your videos. Are you a professor?

Cool dude. :-)

thanks! i hope this works in gpt, groq and ollama! :D thanks!

Fantastic work