Make AutoGen Consistent: CONTROL your LLM agents for ACCURATE Postgres AI Data Analytics

[@TheBlackClockOfTime]

This is low-key the most important channel on KZbin.

[@J3R3MI6]

Agreed 👍🏽

[@ryzikx]

no, that goes to david shapiro

[@effectentertainment7882]

Ok I subbed

[@effectentertainment7882]

​@@ryzikxlove David not only his ai channel either. That man is amazing.

[@TargetedTV2025]

Subd

[@j.hanleysmith8333]

This is the series that we were all hoping for. The other KZbinrs just go over the toy examples that the repost themselves advertised. I love your focus on building powerful, valuable pieces of software

[@Jump-2-the-moon]

The toy examples always frustrate me. Like, how the hell is creating a Snake game or a task manager going to give me any benefit? I want real world scenarios that take a lot more customization and ability.

[@codescholar7345]

This is the series I'm looking for! Thanks!!

[@agenticmark]

When your prompts get into superprompt territory, put your prompts in text files, one prompt per file, them just load the string into your source code. It makes the code a lot cleaner and makes it easier to edit your prompts. This is actually pretty unique for autogen info on youtube. Most people just recap the examples already given from the AG team. Your orchestrator reminds me of round-robin routing. Love the style of this video. Good editing, good pacing, good setup.

[@JustMaier]

Agreed. It’s rare to run into real examples. They’ve gotta exist, but I’m guessing most of them are on GitHub, not KZbin.

[@mickelodiansurname9578]

With the addition of fine tuning from OpenAI then keeping them in a text file is even more useful... cos they are fine tuning gold when they work. In fact it might be good to dump them all into json or create embeddings and put them in a store.

[@agenticmark]

absolutey@@mickelodiansurname9578

[@mareklewandowski7784]

Passion emanates from your videos. I just joined the ride. I feel like we'll have hundreds of thousands of views here really really soon .

[@J3R3MI6]

More AutoGen 💎🙏🏽

[@koliom]

The most detailed and deep dive explanation of Autogen capabilities I found so far! Well done! Thank you!

[@HostileRespite]

Data fidelity is critical. Absolutely right.

[@everythingiscoolio]

Everybody needs to be working on this. We have a chance here to massively change the world.

[@ventureinozaustralia7619]

Not a coder (I'm a low code guy) so I'm following from a distance but this video is brilliantly inspiring, so well articulated, especially the summation at the end.

[@jtbrower]

Your introduction “raw engineering” is exactly what I have been looking for as a Software Engineer myself. Thank you for that intro, you have my follow now.

[@MichaelWoodrum]

Any chance on releasing your repo for analysis?

[@0201alex]

That intro hooked me to comment. Love the way you set viewer expectations

[@stevehoff]

Nice presentation, great detail and a concern for your consumer's time. Well done.

[@DOSputin]

Mad respect. Thanks for spreading the good word, and for sharing your adventure.

[@toruhiyo]

Love stepping off the hype train into some real engineering wisdom. Your content is stunning! Awesome work! 👍🔧

[@CrypticConsole]

I like the tone set by the intro ngl

[@SkyJensen]

Thankful to come into this space and find immediate applicable functionality

[@chrismachabee3128]

Hi, fantastic video. Day 1 here, first timer. I have a joke. I have to work backwards to catch up.Really too new, but I am digging it.

[@saziedhassan3976]

This is just splendid! Thank you so much for taking the time to make these videos. If I could only explain in words how helpful these are. REQUEST HERE: Would love to see an example of agents that work together to retrieve information from both documents (using vector indexing for instance) as well as postgres db. Most business intelligence reside on both and bringing them together could be really powerful for driving key insights. Thanks again!

[@jcd528]

Keep doing it... I will keep viewing... Thanks

[@yiouyou]

May I ask how to get the code in the video? Thanks.

[@nhapham9502]

Please share the source code.

[@n0rrland]

Fantastic work and really exciting to see where this new "team building" is gonna take us! Thanks for the content, looking forward to the next session! 😊

[@thomasgormanable]

Great video, I'm a very novice hobbiest programmer but the orchestrator reminds me of what ChatDev has implemented. It has a process of meetings where up to 2 agents of certain roles are in each meeting and they have a conversation until a condition is satisfied. Chatdevs limitation is that only 2 agents can be a part of a conversation, Autogen solved that with the chat manager but lacks the who should be a part of which conversation. What you have implementated is kind of a merger of the two. Again great video

[@0201alex]

Been using ChatDev for this reason but now the game changed thanks to this video I’m about to make my own agents!

[@ironknight132]

Excellent video sir! Thank you!

[@pioggiadifuoco7522]

well done mate! keep going!!!!

[@wowzande]

Hey bro, Thanks for the video. Cheering you on to build on the momentum..keep going, dont stop. From the heat and the fury. Make sure you dont forget to rest.

[@Jandodev]

The tooling for this is everything!

[@psykepro]

Dude this session is golden! Thank you very much!

[@apereiracv]

I love this and I love you lol. Great content.

[@Bobaleaux]

It would be phenom, it you PR this to the AutoGen repo as a new class of conversation. However, if not, any other options?

[@spputty2127]

I like where you're going with this Dan, I also see great possibilities multi agents. I haven't coded for 30+ years but I'm refreshing/updating my memory and it's like riding a bike really. I want/hope to incorporate low level local agents to speed conversations and save tokens.

[@belatorok5630]

Excellent, LLM is like an Einstein, multi agent is a group of Einsteins but if there are no good orchestrator they are only play their fun games. :) The problem with these LLMs is that they operate too freely non-deterministically, so they need to be kept in check with a well-proven plain traditional algorithm. So they can make really wonderful things.

[@MidSeasonGroup]

Hi Dan, thanks for putting this out. I was having the same issues and I definitely love the route and pivot you took here. I’m curious if you’ve seriously considered a full custom solution yet.

[@tubedewy11]

Hi any chance you are going to share code. Fine if not.

[@cjbobby]

at 23:33 you said that you recommend something for logging. i couldnt really catch it, but i suppose it is some sort of library for logging? could you let me know pls?

[@ernesttan8090]

amazing! how do you figure out all that without any documentation?

[@moisesbessalle]

Amazing! Thank you for this!

[@devhau5]

Love this video! Thank you

[@fierybore1223]

Great work man thank you for these resources!

[@tomski2671]

Thank You for the video, unfortunately I don't have the abilities or time to get into at this low a level. However have you considered creating a GPT agent using your ideas to generate code. That might have its own challenges and require maintenance, but might be worth a try.

[@MrAloha]

I keep trying not to learn Python, and it looks like I'm gonna just ask the LLM to teach me to code like I'm a 5th grader. 😁 Last night, I created a conda env and installed Autogen Studio. Today I'll work on building my team.😳 Thanks for the info!🙏

[@Silkari]

Looks like the type of thing I want to explore, especially the Broadcast Conversation you demonstrated, however, can't follow the video or learn about how you achieved it. Can you share code?

[@ba70816]

I had to edit my original comment! This is exactly what I am doing at this minute! Or at least I was trying to figure how to do it! Nice! Are you on the AutoGen Discord?

[@vecyang]

May have more videos on real world use case on autogen, for exaple, marketing research with back-and-forth auto critic.

[@ph8787]

Hey, do you have to run all the flow to debug? If not how are you debugging?

[@w_chadly]

I'm going to be that guy. digging the glowing keywords. share the theme pls 😁

[@李应健]

nice work！

[@VaibhavRokde-n7r]

provide code link as well.... That help us

[@dunguyen7395]

thank you for your video, it really helpful.

[@faisalIqbal_AI]

Thank for sharing knowlede

[@NenJiDaPassiv]

So cool. I've been thinking along these lines also. I suspect that that this sort of strategy could open up opensource llms to operate with greater efficiency and accuracy. I would love it if you could show us an attempt, success or fail either way, that employs a small llm locally to accomplish some simple tasks. Of course we would like have gpt4 at the helm. Haha ive been chewing through tokens lately and really hope I can find a cheaper solution.

[@Artholos]

This stuff is great! I’m getting great things out of my autogen programs when using GPT 3.5 and 4… but dang is that expensive! I’ve been trying to use local LLMs but very poor results. Would you perhaps be able to make a tutorial on how to get good stuff from local LLM agents? Love you work 🎉 can’t wait for the next one!

[@agenticmark]

Yeah, I run Vicuna locally for autogen, and compared to chatgpt, its just not as good. It will be a matter of time before these open source models catch up. Also the prompting techniques change depending on the model you are using. tricks that work for one, will not get good results on another.

[@thomasgormanable]

I believe there is rumours that OpenAI is reducing it's API cost by a factor of 20 times. There is a developer conference in the next month or so with that being an expected output.

[@Luffy-ss8og]

Great job man. Really amazing work. Can you please share a link to the code? That would be really helpful :D

[@daddydiy9008]

I made something similar utilizing autogen but ran into the issue of having datasets that were too large to be summarized by a single agent. Ultimately this meant that I could never get a unified report. I’ve since figured out a way around it, but I’m curious how you’re doing it. Could you elaborate please?

[@bonadio60]

Excellent content, I am playing with autogen too and I feel the same as to need of more control. For me the function_call should be executed by the agent it self not to return it to another agent to execute it.

[@faridullahkhan1]

thank you, love the autogen videos you creating. 1 ask: is is possible to have a bit more complex interaction between agents, using interactions between groups? multiple groups with agents in each group. thank you

[@leosoulas5897]

is there a github somewhere so we can all contribute to this codebase?

[@eliasgvinp2141]

Have you considered using langchain instead of autogen? Langchain is much better suited for building predictable pipelines, processing data, imho

[@kraketito8999]

Can you code with us, so we can explore it further Thanks for the video

[@prasenjitgiri919]

can you have an orchestrator that has more than one tool, for eg. a chat, a sql agent and also a recommendation predictive algorithm - how does one orchestrate all of these? And in serverless, this is what im finding challenging

[@ColinTimmins]

Hot damn, great video. Love hearing your thought process. Just found your channel, subbed for more! =]

[@李应健]

Can I have the source code to learn

[@mickelodiansurname9578]

@IndyDevDan Your point about the overwhelming number of github projects and their actual utility after they are installed is well taken... like you I spent ages installing and testing and dumping most of them... my best guess is Aider, Autogen and CodeInterpreter are the ones's I'll focus on for my projects and forget the rest... And to be honest had you not started this channel with a pragmatic approach you did, then likely I'd have done so myself... cos there is no much utter shit out there that its unbelievable. I will disagree with you that your views and subscribers are not a focus... they should be Dan... cos this is the utility driven channel I've been waiting for man...

[@0201alex]

Aider is the future

[@vecyang]

Could we share the source code by text? Really insightful structure!

[@rastinder]

how to config a agent that can execute code?

[@navneetchoudhary808]

Can anybody tell me how to use the output from the autogen as everything is going into the terminal. How to reference those output variables?

[@shaftymaze]

Thank you

[@picklenickil]

How do you integrate into so many single thread app framework.. streamlit.

[@aaronward9140]

use chainlit instead

[@theDrewDag]

No matter how precisely I follow this tutorial, the function is never run. I keep on getting "***** Response from calling function "run_sql" ***** Error: Function run_sql not found." This did not happen before writing the orchestrator. I am 99.9% sure code is correct. Can you upload the github of this project?

[@JoseMiguel_____]

same issue :(, did you resolve it?

[@jpit5992]

I also have the same issue - any solutions? PROMPT v1: display all the users who have a hotmail email address TABLE_DEFINITIONS: CREATE TABLE USERS_table ( ID integer PRIMARY KEY AUTOINCREMENT, Created text NOT NULL, Updated text NOT NULL, Authed text NOT NULL, Plan text NOT NULL, Name text NOT NULL, Email text UNIQUE NOT NULL ) CREATE TABLE sqlite_sequence(name,seq) CREATE TABLE JOBS_Table ( ID integer PRIMARY KEY AUTOINCREMENT, Created text NOT NULL, Updated text NOT NULL, parentUserID integer NOT NULL, status text NOT NULL, totalDurations integer NOT NULL, FOREIGN KEY (parentUserID) REFERENCES USERS_table (ID) ) PROMPT v2: display all the users who have a hotmail email address Use the TABLE_DEFINITIONS to satisfy the database query SQLite_TABLE_DEFINITIONS_CAP_REF CREATE TABLE USERS_table ( ID integer PRIMARY KEY AUTOINCREMENT, Created text NOT NULL, Updated text NOT NULL, Authed text NOT NULL, Plan text NOT NULL, Name text NOT NULL, Email text UNIQUE NOT NULL ) CREATE TABLE sqlite_sequence(name,seq) CREATE TABLE JOBS_Table ( ID integer PRIMARY KEY AUTOINCREMENT, Created text NOT NULL, Updated text NOT NULL, parentUserID integer NOT NULL, status text NOT NULL, totalDurations integer NOT NULL, FOREIGN KEY (parentUserID) REFERENCES USERS_table (ID) ) -------------- SQLitedata analysis multi-agents ::: Data Engineering team. Orchestrator Starting------------- --------- Running iteration 0 with (agent_a:Admin, agent_b: Data_Engineer)------------ basic chat(): Admin -> Data_Engineer Admin (to Data_Engineer): display all the users who have a hotmail email address Use the TABLE_DEFINITIONS to satisfy the database query SQLite_TABLE_DEFINITIONS_CAP_REF CREATE TABLE USERS_table ( ID integer PRIMARY KEY AUTOINCREMENT, Created text NOT NULL, Updated text NOT NULL, Authed text NOT NULL, Plan text NOT NULL, Name text NOT NULL, Email text UNIQUE NOT NULL ) CREATE TABLE sqlite_sequence(name,seq) CREATE TABLE JOBS_Table ( ID integer PRIMARY KEY AUTOINCREMENT, Created text NOT NULL, Updated text NOT NULL, parentUserID integer NOT NULL, status text NOT NULL, totalDurations integer NOT NULL, FOREIGN KEY (parentUserID) REFERENCES USERS_table (ID) ) -------------------------------------------------------------------------------- Basic chat() : reply with { "role": "assistant", "content": null, "function_call": { "name": "run_sql", "arguments": "{ \"sql\": \"SELECT * FROM USERS_table WHERE Email LIKE '[e-mail]'\" }" } } function_call() Admin -> Data_Engineer basic chat(): Admin -> Admin Admin (to Admin): ***** Suggested function Call: run_sql ***** Arguments: { "sql": "SELECT * FROM USERS_table WHERE Email LIKE '[e-mail]'" } ******************************************** -------------------------------------------------------------------------------- Basic chat() : reply with {'name': 'run_sql', 'role': 'function', 'content': 'Error: Function run_sql not found.'} basic chat(): Admin -> Data_Engineer Admin (to Data_Engineer): ***** Response from calling function "run_sql" ***** Error: Function run_sql not found. **************************************************** -------------------------------------------------------------------------------- Basic chat() : reply with Here is the SQL query to fetch all users who have a hotmail email address: ```sql SELECT * FROM USERS_table WHERE Email LIKE '[e-mail]' ``` --------- Running iteration 1 with (agent_a:Data_Engineer, agent_b: Sr_Data_Analyst)------------ basic chat(): Data_Engineer -> Sr_Data_Analyst Data_Engineer (to Sr_Data_Analyst): Here is the SQL query to fetch all users who have a hotmail email address: ```sql SELECT * FROM USERS_table WHERE Email LIKE '[e-mail]' ``` -------------------------------------------------------------------------------- Basic chat() : reply with { "role": "assistant", "content": null, "function_call": { "name": "run_sql", "arguments": "{ \"sql\": \"SELECT * FROM USERS_table WHERE Email LIKE '[e-mail]'\" }" } } function_call() Data_Engineer -> Sr_Data_Analyst basic chat(): Data_Engineer -> Data_Engineer Data_Engineer (to Data_Engineer): ***** Suggested function Call: run_sql ***** Arguments: { "sql": "SELECT * FROM USERS_table WHERE Email LIKE '[e-mail]'" } ******************************************** -------------------------------------------------------------------------------- Basic chat() : reply with {'name': 'run_sql', 'role': 'function', 'content': 'Error: Function run_sql not found.'} basic chat(): Data_Engineer -> Sr_Data_Analyst Data_Engineer (to Sr_Data_Analyst): ***** Response from calling function "run_sql" ***** Error: Function run_sql not found. **************************************************** -------------------------------------------------------------------------------- Basic chat() : reply with I'm sorry, but I'm unable to run the SQL query at the moment due to a technical issue. However, your SQL query seems correct. It will fetch all users from the USERS_table whose email ends with '@hotmail.com'. Please ensure that the table name and column names are correct and exist in your database. --------- Running iteration 2 with (agent_a:Sr_Data_Analyst, agent_b: Product_Manager)------------ basic chat(): Sr_Data_Analyst -> Product_Manager Sr_Data_Analyst (to Product_Manager): I'm sorry, but I'm unable to run the SQL query at the moment due to a technical issue. However, your SQL query seems correct. It will fetch all users from the USERS_table whose email ends with '@hotmail.com'. Please ensure that the table name and column names are correct and exist in your database. -------------------------------------------------------------------------------- Basic chat() : reply with APPROVED -----------Orchestrator complete----------- Ochestrator was successfull

[@theDrewDag]

@@JoseMiguel_____No man.

[@jpit5992]

Its driving me crazy@@theDrewDag

[@JoseMiguel_____]

I just found that you need to add function_map=function_map to your engineer assistant too@@theDrewDag (Im not sure why)

[@HistorIAsImposibles776AC]

If I had a company I would hire anyone here 😅

[@hidroman1993]

Would be great if you showed how much it cost

[@Jonathan-ih9sm]

The channel exploded because a lotta people made surface level videos about autogen and your video really stood apart.

[@zyxwvutsrqponmlkh]

Get on with it.

[@paullopez_ai]

I like to content, I would watch more videos if you reduce all the hand motions. It makes viewers dizzy and it’s distracting. Just some input.

[@n0rrland]

I kind of like the hands movement, makes you feel more present :D

[@MegaGutemusik]

N1

[@milleniumdawn]

You not sharing the code, is doing a lot less sharing that you think you do.

[@alfonsopayra]

what the heck is going on with those moving hands???

[@alfonsopayra]

MY GOD STOP MOVING YOUR HANDS MAN

[@WiseWeeabo]

stop waving your hands around

[@contractorwolf]

bring it @IndyDevDan, there is too much hype. I came to build

[@Seaileanu]

The "jazz hands" bro...