May 2 '26, Saturday
May 3 '26, Sunday
May 4 '26, Monday
May 5 '26, Tuesday
19:08·2 SPEAKERS·2065 WORDS
ASo I can give you access to my code if you want to look. But I don't think it will help that much
Bbecause I'm also constrained by now what Dell will give to me and what I'm allowed to use.
AAn LLM can. Tool call. Tool calling. Tool calling is basically. You might just expose those two things.
BHow does a fuzzy search work?
AThat has a different question that has nothing to do with LLMs. Right. How does fuzzy search work?
BI don't know what tool calling means, so I thought that's what it is.
ANo. First figure out the deterministic parts of your project. Things that don't require AI.
BOkay.
AYou should be able to build an execute SQL function that is not. Has nothing to do with AI.
BRight, I already have that. I have an execute SQL function. I have a fuzzy search function. Well, formatted database. No, I don't have a fuzzy search function.
AOkay, so that needs a lot of work on its own.
BFuzzy search is not the part that AI does.
ANo.
BHow is fuzzy search done?
AFuzzy search is done by multiple ways. Like there are so many different ways of doing. It depends on your data shape. I can give you an example with my thing where I take every interaction that I create on my website and I embed it into a embedding space. And that embedding space is a vector space. And then I tell the LLM, you can query this embedding space with whatever string you want. So that string is also embedded, is embedded. You find the nearest neighbors and then you return all those closest neighbors.
BSo this converting of data. Oh, LLM. Do you have an embedding model?
AYeah, you have to have an embedding model.
BOkay.
AFor semantic search. Yeah.
BSo then what is the LLM doing? Converting the results of semantic search into natural language. Is that what the LLM does?
AAnd like being able to find things on its own, like as a human, you might what to search for. That's what the LLM is deciding. It might do four parallel searches. It might run some deterministic code and do some semantic search and try to merge those results. That intelligence part is what the LLM is doing.
BOkay.
AUnderstanding the results.
BSo the fuzzy search is also based on an embedding model that is able to take any data piece, including the question and convert it into a point in n dimensional vector space and then find nearest neighbor.
AYeah, that's basically.
BOr make an answer based on nearest neighbour.
AMy fuzzy search does a little more. But this is not amazing because embedding is not a great. Like it's not Lossless, right?
BCorrect. That's what I was reading.
AThen I've also set up full text search. SQL offers full text search on text columns.
BOkay.
AWhich is a deterministic text data.
BSo this is not relevant.
AThen what are you going to embed if you don't have text data?
BI mostly have like numerical data that I want to be able to ask questions like what is common between failures across a mayor in the last quarter.
AThen you don't need semantic source at all, right? You're not looking for data semantically, right?
BI am looking.
AYou are not looking for data semantically, right? This. How would you do this? You would write some SQL code, run it, see the result, and then maybe write some Python file to analyze it on top of the SQL, whatever data was.
BI guess what I want is for the LLM to be able to run a variety of SQL queries that explore various avenues of what possible problems could be.
ASo you don't need a fuzzy search.
BOkay.
AYou know, like if you don't have text data.
BFuzzy search is specifically for text data. Then I don't need.
AYeah, you don't need it. Like fuzzy search is for like text data only, right? You can't fuzzy search over.
BYeah, and I don't have. I don't have any like search survey answers or like.
ASo all you need is an execute SQL tool and a tool which tells the LLM. I mean, again, this is a separate optimization. Don't have to worry about it. Forget about that for now. Tool call. Okay, now we come to the tool call part you have. So exactly the parts that you will do with your intelligence is what the LLM can also do. But just expose the to SQL Execute SQL tool called to the LLM.
BHow do you expose.
AOkay, to expose and tool call. How does an LLM API work? You just send.
BI don't know.
AThere's a server. It's very simple. There's a server. You send a request complete, give your answer. It will just give that text back, correct. Okay. And then when you put one more text, it takes all three of those texts, puts it back, pushes back.
BCorrect.
AApart from sending text requests, you can also specify context, other things. Not, not that is the context. What we're sending is the context.
BNo, but like, you can attach files or like, you can attach.
AYou can attach files, you can specify output JSON formats. Now, okay, where like, don't output a text text message, only output like a JSON message in this format and you can attach tool calls.
BOkay.
AWhich means the LLM can choose to either respond to you in normal language or it can choose to tell you to call one of those tools and send the response back.
BOkay. So it can format the API like call with the necessary parameters and tell you run this and then.
AYeah. In the tool call, you give it a function, tell what the function does, give it all the parameters that the function takes and tell what each of those parameters mean. It will understand all of that and then. And based on your question, it will decide what tools should I call. It can't call the tools from the server. It just tells you you run these tools because you have access to them. But whatever output you get, give it back to me and I will understand and answer and continue. Okay, so that's all you need. That's what tool calling is.
BOkay. Now how do we make this automated? Like how? Like I can't manually, like put my Execute SQL function every time and ask it a question and then it gives me a query which I run and then I feedback, right?
AYeah.
BSo how do I make a pipeline of this?
AThis pipeline? Banana B. Just use any SDK. Use a Versal AI SDK.
BBecause I don't also I feel like I'm fully lost with this.
AYou understood the framework that ABHI built? No abhi.
BYeah, but now how do I look at. Look at what these tools that are available and know which is correct for me. Windsurf I know is my GUI.
AWe are looking for an LLM, a library that calls LLMs. We are not looking.
BWhat am I looking for? Please tell me, what
Aare we looking for?
BFor? I'm looking for Dell Server LLM Model Hosted or Uska Muje Programmatic Calling API.
AThat's true. That's one thing you need and I
Bdon't know how to look for this, but am I correct in what I'm looking for?
AYou're looking for the API? Yes. Hopefully you're also looking for not just the endpoint, but some kind of tooling around the endpoint so you don't have to do REST calls. Well, the REST API calls directly some Python bindings on top of the. That's an SDK. A Python SDK. I'm sure Delga that's what you're looking for. But exactly
Bsay, because I know other teams have started doing stuff like that.
AIs there no LLM search for your entire.
BYou want to see it. They've made it so that it's absolutely useless.
AWhat about like just. Python version? Use Python version. Use Karakto. Roger. What about PIP installing stuff? Can you PIP install Anything or what?
BYeah, I think I can PIP install anything. Python versioning is controlled by company Portal, same as UI install Python from company portal. But like let's say I'm here, okay@askdel.del.com and I can pick Opus 4.7 and then I wanted to search for search all of confluence.
ASo to explain what I do, maybe it gives more context. I use the Vercel AI SDK which is just a SDK on top of different APIs. Anthropic's API, OpenAI's API, Gemini's API and tries to condense them into the same interface so I can Talk to these LLMs in a unified way. And there all I do is I specify tools tools parameter equal to the list of tools that I want the LLM to be able to be like to be called and then parameter the max turn count or something.
BSo let's say if today you're using with your with the Vercel SDK and your code today you're using Sonet 4.7 and tomorrow you feel like it's a one line change because the SDK does the job of unifying like for.
AFor you don't huh. That. That is true. But even when like each of these model providers also have their own SDK like OpenAI has an OpenAI SDK Claude has a. But this SDK kind of unifies all those SDKs as well. But yeah, more specifically.
BIs this something like where I might find something that I need?
AThis looks everywhere.
BThere's like I think this agentic platforms. I don't know what that is either. On that platform as a service.
AI don't know. Why don't you just copy this whole page tell Claude what you're looking for and it'll give you the best candidates for it. Okay, you're just looking for a Vercel AI SDK that connects to Dell APIs. Dell APIs tools you will specify.
BBut the Dell AI tools endpoint. Okay. The Del AI endpoint.
AControl a. Control c. Yeah. Snipping to control c. Ah bro, smart. Tell it what you're looking to do
Bfor I have this conversation is already that.
ABut have you like it does not know what we just discussed? No.
BOkay, so I start a new chat then I tell it I am trying to build a system for me to be able to ask questions in natural language from a database that contains categorical slash numerical data.
AOkay. This is very high level. Just tell it I have the. I'm just looking for a Vercel AI SDK alternative that I can use internally within my company, Dell. So I, I want to be. I want to write some Python code that calls Dell's internals internal LLMs. That's what you want. And then now tell it. Are any of these relevant or should I look somewhere else?
BI don't like having conversations that are not code related on cloud code with.
AThis guy. Can't do anything.
BBut when I want to chat, I would rather do it here. And then when I want to implement, I would rather do it there. I chat here. This is my chat.
AGood point.
BOkay,
ADo like a. Okay, It. You don't need rag. Just to clarify. Rag is dead anyway.
BRag is dead.
AMalav. You don't need rag. RAG is essentially the other thing which is fuzzy search. And the LLM works. RAG is essentially exposing fuzzy search as a tool call. That's it.
BThat's what RAG is exposing fuzzy search as a tool call. Okay, understand.
ABut that's it. Yeah, that's it.
BSo I don't need that for this data.
AFor this data, you don't need it. But also RAG is just that. You don't need a special system to implement rag. You just need a fuzzy index over your data. That's the hard part.
BOkay,
AYeah. These two things can run in a loop. Basically. I would also just have this chat in cloud code and right there and then it can tell you register, give me these keys. I will construct the python code for you and do everything for you. Yeah. Gateway interception in. Or restart. Anyway, you understood? No, what you're looking for, I guess.
6 episodesMay 6 '26, Wednesday
May 7 '26, Thursday
May 8 '26, Friday
Open this page on desktop.