Everything You Need to Know About Vector Databases for AI Agents

[@WreckAnEyeZe]

Would love to see how you are loading email data into the db and if you are doing any categorization, tagging, etc on it

[@customaistudio]

No categorization or tagging right now. It's going in raw.

[@soppliger]

I am curious, with respect to emails as the source doc in particular, if you ran your "sent" emails through the pipeline, would the method you outlined enable the selected language model to create replies in your style/voice?

[@customaistudio]

Hey just got done with 4 projects where I did this exact thing. Works like a charm

[@antonpictures]

i was hoping you pick text + image embedding - nobody shows this example. if you find one guy like you explaining multimodal embeding, locally, that works please let me know - thanks George

[@customaistudio]

I'll dig in on that. I've never used it before.

[@antonpictures]

@customaistudio Upon clicking, the app captures an image from the webcam and takes a user-provided prompt (a question or request). It then utilizes an LLaVA LLM to process both the image and the prompt, potentially employing loops for reasoning. i used 'clip-ViT-B-32' model inputs >77 tokens. The system is designed to be context-aware, incorporating a short-term memory of recent interactions and searching a long-term multimodal RAG database for relevant memory information. The captured JPGs primarily serve to store information about the user and their environment (typically a room or office setting) in the RAG database. The AI's objective is to interact with the user in a personalized manner, recognizing their face and using their name, while also comprehending and reacting to their surroundings and actions. I'm not a coder myself, but I've built most of the app. However, I'm currently facing challenges with the multimodal memory storage aspect, as this memory is crucial for true intelligence. While models can be swapped or upgraded through the vision LLM selection, the rest of the code primarily involves looping responses and comparing them within the database to create a goal-oriented response, which is then stored again. Please let me know if the MMRAG memory functionality is working correctly. Thank you for your assistance! :) George.

[@soppliger]

Agree. Multi-modal is becoming table stakes. Not necessarily for the video tutorials but for the applications themselves.

[@willrodgers7974]

Have been playing with the different agent frameworks via code for a little while now, with mixed success. Definitely going to look into this n8n way of doing it. I usually don't go for no code but this looks like an extremely useful tool to have in the toolbox for times where good enough but fast development trumps better but slow to write. Looking forward to the series you mentioned. 🙏

[@customaistudio]

Awesome man, I definitely like the UI with N8N. It makes it a lot easier to design the flows and structure the tools.

[@jackmermigas9465]

If something works in n8n how or why is it better to go to the effort to write it all yourself? Honest question BTW :)

[@rcardenete]

Brilliant! When are you dropping the video about feeding Pinecone with past company data sir? I keep refreshing your channel.

[@customaistudio]

Been busy man, but working on it

[@freethinkerca]

Great work Devin. Thanks for creating these.

[@sai_ai-z1c]

Automation, agents, and AI tools are revolutionary! I've been able to coordinate cooperative AI workflows that greatly increase productivity with SmythOS. Intelligent automations are incredibly simple to create and implement thanks to the no-code interface.

[@Affstuffcom]

does this works using community edition of n8n

[@customaistudio]

Yes it does

[@ConorDean]

Curious if you ever experimented or considered using the OpenAI Assistants API for defining your agents which I see is in option in n8n. I think the key benefits are that it manages the conversion threads and provides you with a VectorStore.

[@customaistudio]

I've tried it multiple times but couldn't get it to work. For some reason, the assistant doesn't call the function consistently. Have you done it before with success?

[@ConorDean]

@@customaistudio not with n8n - I’ve written code that does this and toyed with Agency Swarm (youtube.com/@vrsen?si=m5r1gxIQVi_CPHGL) but I wind up spending too much time debugging instead of building out useful workflow.s. As a SWE, I typically shrug off the no code solutions, but your presentations and shared experience has changed my view of things. Thank you for sharing and I look forward to more videos!

[@customaistudio]

Thanks man! I've been a huge fan vrsen for a while and have used Agency Swarm as well. It's a great framework but I agree, it takes more time to debug than I'd like. The tinkerer in me wants to keep playing with it but I'll do that at the expense of more important things.

[@webdancer]

​@@customaistudio very true.

[@jackmermigas9465]

@@ConorDean I'm curious why SWE's shrug off no code solutions when they are so fast and easy to use? Is it mostly a legal and privacy thing when working for larger companies?

[@jackmermigas9465]

Brilliant this helped me immensely and also looking for clear easy solutions that make businesses money. My new fav channel!

[@Fortheexceptionalmen]

Would it be possible for you to create a video showing us example of preparing the data? Even if it just 1 day of data so we can have a starting point in mind. Would be greatly appreciated. Thanks!

[@customaistudio]

Yep, I'll post one around that

[@kuromakhang6202]

Sir i meet one problem, how can i get specific data of the whole knowledge basement to another tool in same Agent. For example, i put the list of name and contacts information like : email, phone number,... . how can i extract these to other tool. after vectorize them?

[@customaistudio]

Hmm, can you outline the step-by-step here?

[@yewtree7]

Really appreciate the videos mate, thank you for sharing your knowledge and experience

[@needspersonality]

16:23 how do I upload data from the google doc to Pinecone vector store? I added an "edit fields" node but I'm not sure what the formatting should be. Thanks for this tutorial!

[@customaistudio]

So I misspoke on this part. I added it to the google doc because I wanted to make manual edits to the data before adding it to pinecone. You don't need to add it to a google doc if the data is already structured how you want. If you are using a google doc, you'll need to create a separate workflow that converts the doc to a binary file and uploads it to Pinecone.

[@needspersonality]

@@customaistudio got it! Thanks for the quick reply.

[@christiansard5247]

Thank you for sharing your experience in this area ;-) can't wait your next videos ..;-) to debug all those ai workflow i use langfuse that like n8n can be self hosted, what you use ?

[@customaistudio]

Thanks! I don't use anything honestly. I just put error messages in ChatGPT

[@sirishkumar-m5z]

Great overview of vector databases! SmythOS can help you effectively manage and utilize vector databases in AI agent development. #VectorDatabases #SmythOS

[@customaistudio]

Cool, thanks

[@Sasha_sementsev]

🙏

[@customaistudio]

🫡