Make Your RAG Agents Actually Work! (No More Hallucinations)

[@leonvanzyl]

The plan was to release this video as part of a paid course on agentic agents, but I decided to release it for FREE. You can support my work by hitting the like button and subscribing to my channel 🙏. What Flowise topics should I cover next?

[@BirdManPhil]

i need your help are you for hire?

[@piero957]

Add SQL as an option other than vector and web. Also, add a "user prompt enhancer" option using llama3.2 again. Another interesting option for non-English users would be add input and output translation optional steps, so an user can query let's say in Italian, the agents and the llm then works better in English, and at the end the user can have the reply in his/her language, Italian in this example. Doing all this using a music/movie SQL IMO may grab great viewers attention, both for personal as well as business solutions.

[@kreak100]

Great work Leon, love your work up here in Joburg. An interesting use case would be to have agents go try solve a problem for a while and when they find an answer, return the answer. e.g search through a large data base sequentially and find a trend, say large abnormal increase in value, then report that.

[@kelekcibo9172]

HI Leon, we have many Tickets in our Jira Project. I wanted to use your template for having a "JiraAI", who answers questions to tickets, etc. I am using self-hosted qdrant. I have an agent , but it always says stuff like, I dont know where to find it, etc. I am inserting the tickets via n8n. With Flowise, I only want an agent-/chatflow who answers questions in regard to tickets. Which video do you suggest me to watch to implement my scenario? Is it possible to use qdrant in the document store?

[@gianlucagilardi6200]

When the feature comes, I cannot wait to see your take on Flowise, Ollama and Graphrag integration....

[@sanjuburkule]

Subscribed! Your speed of explanation and not missing steps is important for someone like me. You could have said 'Do so on and so forth' after showing 3 to 2 LLM nodes getting added, but you went ahead and showed it. This is very valiable for my level of people. It allowed me to forward by 10s wherever needed, but that confidence in knowing each step is useful. Thanks for creating this video. Will watch the others you referred here.

[@leonvanzyl]

Thank you for the feedback 🙏

[@gmcwhinney]

This is incredible Leon. Thanks for releasing this for free so that we can all learn from your obvious expertise here.

[@leonvanzyl]

You're very welcome

[@gianlucagilardi6200]

I am definitely flabbergasted. I started fiddling with Flowise thanks to your videos but this, this is.... unbelievable. I dunno whether you have scraped your projects for paid course(s) but if you have not I am most definitely going to attend (if they have remote attendance, that is :) )

[@leonvanzyl]

Wow, thank you for the feedback 🙏.

[@DresElMagnifico]

Great tutorial Leon. Thanks for all your work. Tutorials like this are great for people trying to build real-life applications. I feel the hardest thing to achieve is to go from a solution that works 75% of the time to one that can reliably perform 95% / 98% of cases against the Wild West of high-volume user inputs. It would be great to see more content like this with slightly more advanced techniques and complex scenarios. Thanks

[@SoloJetMan]

Thanks for putting this together. I've been debating between several agent platforms. your videos helped me narrowed it to Flowise

[@leonvanzyl]

Yeah, Flowise is super poweful

[@teemunordman]

Your videos are professional quality. Thank you very much for all Flowise video content!

[@leonvanzyl]

Thank you!

[@carlosavelar1300]

Fantastic! Leon, you are the best! Thanks a lot for sharing your knowledge with us!

[@leonvanzyl]

You're welcome 🤗

[@artur50]

wow, incredible that you share that stuff with us! GREAT!

[@leonvanzyl]

More to come!

[@ibrahims5636]

i have try it, very good result for sure, the only problem right now is how do we integrate memory management, the only available option on this flow is AGENT MEMORY on state, but its sqlite based, the sqlite agent memory are not CPU friendly

[@walberc]

Great video Leon. Thank you so much for this.

[@leonvanzyl]

You're welcome

[@santiagoghione9177]

Excelentes tus videos Leon! Muchas gracias

[@leonvanzyl]

Thank you

[@AlistairWalsh]

Great stuff! This is exactly the information I needed right now.

[@leonvanzyl]

Glad to hear 💪

[@OscarTheStrategist]

Super useful. Thanks for the video! 🔥

[@leonvanzyl]

You're welcome 🤗

[@amroeser]

Wow, thank you so much for your videos! It's great that you're sharing your knowledge. Thank you so much! Is it possible to analyze Excel data with tools like the calculator and write a summary in Word of the differences between two Excel files?

[@leonvanzyl]

Absolutely! Sounds similar to this video: kzbin.info/www/bejne/jYPMh4Cfh65qn7c

[@MatichekYoutube]

very good tutorial - thank you for hard work 💙

[@leonvanzyl]

Thank you

[@photize]

Excellent as always, opens so many advances thanks .

[@leonvanzyl]

You're welcome 🤗

[@megaledafrica1080]

Awsome stuff Leon, thank you.

[@leonvanzyl]

You're welcome

[@VishwanathStudios]

Can you do a video on how to use google sheets as a data source for a chat assistant to use directly without any third party tools.

[@Bu3askoorDXB]

Thank you as always, i kept using flowise just because of you and your videos. I have noticed a video where potentially using flowise api, a user can create chat flows. It might an ok idea for a video

[@fabio-y5q]

Hey Leon, really cool and interesting video!☺️ I have a question: How would you implement an agentic workflow that answers questions about a document where the needed information is spread across the entire document? The retriever only gives a certain amount of related content, right? In an example, let's say we have a transcript of an interview, and I want the RAG system to tell me how many and which interview questions appear in this document. Here, the document would need to be completely searched and checked for this information. Would you integrate a loop that goes through all the chunks in the database and adds the information to a state value, or is there a more efficient way over self-reflection to implement this? Thank you for your absolutely amazing and interesting videos. It’s a lot of fun to recreate agentic workflows with these structured videos!🎉

[@Francotujk]

Leon, you are the best! It's incredible valuable all you are doing!! May be, it could be nice to create more tutorials not just focusing in the Flowise UI, but in creating apps (nextJS, Streamlit, etc) that use Flowise to power the AI-Backend. Because dealing with the API of flowise (for example to upsert files and the read them by filtering, while mantaining the files private is not easy). Also, I don't know if is it possible to use the Agenflows as backend, in that case it could be nice to see a full stack app. Obviously not focusing too much in the UI of the app, but in how to connect it to a flowise project. Again, thanks for everything you are doing!

[@leonvanzyl]

Excellent suggestions.

[@usiala]

great video. thanks.

[@T33KS]

Thank you! Always the best content on agentic flows and automation. I just have one question that I still can't find a solution for. How would you approach deploying/orchestrating/monitoring these flows? What would be your approach for production once you have multiple flows and duplicates for different clients (also valid question for n8n). Would love your insights on this.

[@leonvanzyl]

This is an excellent topic for a dedicated video. I'll try to put some thoughts down, but it's a bit toouch for a comment 😁. It all starts with the client's requirements. They might want to host and manage the flows themselves, so you're basically only responsible for setting things up for them. Alternatively they might want you to create, monitor and manage everything. Hosting: Flowise Cloud is a really solid option for many reasons. You don't have to worry about updating FW or scaling the platform. You can also collaborate with team members on flows. You could of course self host it on a VPS / Render, etc. This could be useful if there are some region restrictions. I would also recommend adding Langsmith to your flows. That way you can find failures and troubleshoot any issues. Health checks: I have an automation on n8n that calls the prediction API on a schedule that will inform me via email if the API is down or some other unwanted behaviour.

[@T33KS]

@leonvanzyl that will be a great video. Especially, in the case of managing multiple instances for different clients. Thanks for taking the time to reply.

[@plannedillusion]

Leon for president 👍🏾

[@leonvanzyl]

Bwahaha!

[@magmikefpv]

Thank you Leon

[@leonvanzyl]

You're welcome 🤗

[@PyJu80]

Hey, im ngl, im a n8n fan, thank f. your code is in json format. One thing im trying to do is use the bolt.forked, or even the pimped version (not sure if i can qutoe, but its on here called ottodev) as an interface and have your workflow used as an http url ollama model. Whats the chances of you looking into this for me?

[@LemurDrengene]

What is the drawing software you use in the beginning? And thanks for another great video :-)

[@leonvanzyl]

You're welcome 🤗. It's called eraser.io

[@LemurDrengene]

@@leonvanzylThx, appreciate it

[@AllanBrazute]

Nice video, thank you. It is kind of working for me. There is one question, the first one I asked to test it, that it can't answer and I don´t understand why: - question: What time is it in Australia? - websearchLLM: 12:05 AM - generateAnswer: Australia is 15 hours ahead of UTC (Coordinated Universal Time) during standard time and 16 hours ahead during daylight saving time. Since you didn't specify which region, I'll - assume it's the Eastern Standard Time zone. In this case, it would be 9:05 AM on a typical day. - hallucinationCheck: noResponse - final: I'm sorry, but I couldn't find the answer to your question. You could try rephrasing the question and try again?

[@leonvanzyl]

Double check that the documents state is correctly being set in the web search tool node. You can also try to adjust the system prompt of the generation LLM to guide it in the right direction.

[@MichaelHubbard]

Thanks you!

[@leonvanzyl]

Thank you very much for the super. Your support help a lot 🙏

[@micbab-vg2mu]

Great - I will try it:)

[@leonvanzyl]

Let me know how it went 💪

[@krishnankuppaswamy7553]

Neat and super useful. You have a very good knowledge on this subject and following your contents closely. Unable to import the flow on the Flowise Cloud version 2.1.4. I am on trial period and trial started just a while ago. Status: 500 Error: exportImportService.importAll - Cannot read properties of undefined (reading 'length') Any hint to make corrections to the json file would be useful.

[@leonvanzyl]

Thank you! It seems a lot of people are having issues with importing my flow. Investigating

[@krishnankuppaswamy7553]

@leonvanzyl Thanks 🙏

[@AIbel3arabi]

can you start doing VS Vector Shift tutorials?

[@tecnopadre]

Hey Leon, thnks for sharing. When the last LLM that generates the last answer isn't this the way a second LLM generates the answer to the same question for second time? Why wouldn't you take the previous answer if you already validate it with the condition Agent?

[@leonvanzyl]

There's no way to do a "pass through" of the generated answer. I guess you could change the last system prompt to "Repeat the following words" followed by the generated answer. I personally just like to summarise the answer. It won't add anything new,. Good question though. Either approach would be fine.

[@tecnopadre]

@@leonvanzyl I'm just trying to save inbound and outbound fees. 😅

[@kelekcibo9172]

HI Leon, we have many Tickets in our Jira Project. I wanted to use your template for having a "JiraAI", who answers questions to tickets, etc. I am using self-hosted qdrant. I have an agent , but it always says stuff like, I dont know where to find it, etc. I am inserting the tickets via n8n. With Flowise, I only want an agent-/chatflow who answers questions in regard to tickets. Which video do you suggest me to watch to implement my scenario? Is it possible to use qdrant in the document store?

[@maniecronje]

Fantasties!

[@sharankumar31]

Very nice

[@leonvanzyl]

Thanks

[@choijungho1]

Very surprising and admirable! I loaded the json file to test it locally. "Failed to import: Error: exportImportService.importAll - Cannot read properties of undefined (reading 'length')" An error occurs. Is there a solution?

[@changtimwu]

Hi! I got a similar error. My flowise version is flowise@2.1.3

[@changtimwu]

The error occurs when importing the JSON file under Chatflow instead of using the "Load agent" option in Agentflows. BTW, Flowise should display a more user-friendly error message for this scenario.

[@szood]

can you teach how to set up flowise with a database please?

[@youtubeccia9276]

excellent stuff

[@leonvanzyl]

Thanks 🙏

[@Chasingaxl]

Is this Miro that you using for the diagramming?

[@leonvanzyl]

It's called Eraser.io

Thanks

[@leonvanzyl]

Thank you very much for the support 🙏

[@revotech5980]

can flowise be used for production ready products? thankyou

[@leonvanzyl]

Absolutely! I've got client apps running on Flowise in prod for over a year now. It's super stable

[@joseignaciodussaillant7031]

I keep getting "Error buildAgentGraph - Cannot read properties of undefined (reading 'name')" when trying to test the flow. Any tips? havent been able to find the solution anywhere

[@TeamsWorkAI]

Great !

[@leonvanzyl]

🙏

[@wortellemon2943]

after save the project, when i click chat button its automaticly give me a blank white screen

[@Aaron7k]

Thankss plz do in n8n

[@EinfachJens-16]

great video. workflow file is not working on my maschine

[@leonvanzyl]

What seems to be the issue?

[@EinfachJens-16]

@@leonvanzyl my fault. load the workflow wrong

[@pushingpandas6479]

How install Lama3.2 on a digital ocean server?

[@leonvanzyl]

In that case I'd actually suggest you look at using Groq. It's an awesome platform for using Llama 3.2. I have a video on using Groq and Llama on my channel actually.

[@김정률-i9n]

Can I request a course on vector database metadata processing?

[@leonvanzyl]

Great idea

[@k1r0vsiii]

you re a flowise gandalf leon

[@leonvanzyl]

🤣 Thanks

[@tonyseno]

Like and subscribe !

[@MariushAI]

Is it possible to create a similar workflow in n8n?

[@leonvanzyl]

Well, the principles that I teach in this video is not specific to Flowise. It should be easy to implement in n8n. Maybe I should create an n8n tutorial on this?

[@MariushAI]

​@@leonvanzyl definitely yes

[@danzreview]

GG

[@김정률-i9n]

How to use flowise nowhere else in the world!!!!! COOL!!!

[@BirdManPhil]

Leon, forgive me if im wrong, but are you not for hire?

[@leonvanzyl]

Send me an email and I'll see if I have capacity 😁

[@orangtimur6812]

Why my flowise ui different 😢

[@leonvanzyl]

You need to update your Flowise instance maybe 🤔

[@univer6979]

If my app needs to work offline without internet connection, in that case can we further improve the response by asking it to search the vector again? I had tried creating a chatbot as per your earlier video kzbin.info/www/bejne/onuyi5yIpL-BgNk but i could still see some hallucinations and non complete responses even though the answer was there in the vector store. Any thoughts on how to improve it?

[@plannedillusion]

🤣 I'm actually building a similar system.

[@leonvanzyl]

Great minds think alike?? 🤣

[@DresElMagnifico]

Thanks

[@leonvanzyl]

Thank you very much for the support 🙏

[@harrykekgmail]

Thanks

[@leonvanzyl]

You're welcome