Shadowing Practice: Master AI in 60 Minutes with Vector Databases! - Learn English Speaking with Video

בחר מצב תרגול

כל מצב פותח חדר תרגול במסך מלא.

712 משפטים

אם המשפטים קצרים או ארוכים מדי, לחץ על ערוך כדי להתאים אותם.

All right, so in this video, we are going to continue talking about vector databases.

So we are now going to start pulling all the things we've learned so far and put it all together.

So the idea is that we're going to take all the knowledge we already have about vector databases, about how to create a vector database,

and then how to put information into the vector database, so vectorized data essentially,

and then being able to search to use the vector database to do similar research

so we can get the information that is closest to the answer

or at least the pieces of data that is closest to the answer that we're looking for.

So we've done that part.

By the way, this is the second part of this vector database series.

So if you want to continue to follow, I would recommend you to watch this video somewhere at the top here,

here or over there, where we got started with vector databases.

So go ahead and watch that and come back and watch this video.

All right.

So that's what we're going to be doing in this video, in this second part of vector databases.

So it's going to be more of a hands -on, but I want also to continue giving you the pieces,

the building blocks for you to really have this bigger picture of vector databases, and how to use them ultimately.

Because remember, ultimately, we are going to be using the vector databases with large language models,

because the idea is to have the vector database to do all the heavy lifting per se,

to be able to query things into our vector database, to get those chunks of information,

because it's real good at doing similarity search and all the great things

that we need to get the semantic meaning of certain documents, certain words, and so forth.

And then the idea is to take that information and then pass

that through a large language model because the large language model knows more.

It knows how to take those pieces of information and sort of summarize everything so we can get the final result.

And that is the main goal of vector databases and the main goal in this video, the second part.

But before we go and we start doing the real world use case coding, showing everything, there is another concept I need to share with you

when it comes to understanding as a building block of vector databases, which is embeddings.

So as you will see, I'll show you in this video, of course, before we start coding,

what embeddings are and what is the importance or the significance of embeddings in this realm of AI -based application vector databases.

That is something to look forward to in this video.

So without any further ado, let's go ahead and get started.

We understand that vector databases are real good at dealing with unstructured data,

which is more than 80 % of overall data floating around in the world.

So let's look at the unstructured data, how that is actually transformed into vectors.

So the idea, of course, is to take this unstructured data and then transform that into the vectors database,

into vectors, essentially what we call ALT's embeddings.

We'll get into that.

So that is the whole idea.

We'll get this unstructured data, transform that into vectors, just the numerical representation with a little bit more to it,

so that then that is what is actually saved into the vector store or vector database.

But let's look essentially what really happens.

So the idea in the grand scheme of things is that when data comes in, let's say we have a huge amount of documents.

It could be different kind of documents types.

It could be videos.

It could be text, articles, audio, and so forth.

Let's just assume this is just documents in general.

The idea is that we need to split that information.

That needs to be split into smaller chunks and then transformed into embeddings and then that embedding, which essentially is a vector,

we'll see the differences in a second here.

This embedding or vector is what is stored in the vector database or vector store.

So essentially, that's what's happening here.

A database, essentially, a database only saves this format.

That's why it's called vector store.

That's the idea.

But now we need to understand what is this embedding.

So embedding vector, so we can call this embedding or embedding vector.

Essentially, it's the same thing, okay?

So what embedding vector is, is that it keeps the content and the meaning of the text.

That's very important.

So it has a little bit more of metadata.

more than just the vector as we know in mathematics.

And the idea is that text with similar content and meaning will have similar vectors.

And that's what we've discussed and we now know that's the whole idea.

Anything that is similar in content and meaning, right?

Similar in content and meaning will have similar vectors.

Makes sense.

So essentially, as we saw before, if we have all of these words, for instance, cat, kitty, flower, eat, run, walk, Europe, and so forth,

cat and kitty have essentially the same similar meaning.

Therefore, their vectors will have similar numbers.

Therefore, their vectors will be also similar.

That's the whole idea.

And so going back to what we've just discussed, that is indeed what is saved in the vector database.

Now let's look at the vector database of vector store.

It's actually the same thing.

Some people use vector store, other people call vector database, just the same thing.

Let's look at an overview of the whole process.

We really need to understand how the vector store internally,

how things work, and how the actual vectors and data

and everything is embedded together so that it is put into a vector store or a vector database.

Same thing.

So again, the idea is that we have documents.

Again, it could be the actual documents or structured data.

It doesn't matter really.

Is that this has to go through a splitting process, generally speaking.

Okay.

And embeds are created.

In this case here, these are pieces of data that are representing these text splits.

And these are essentially what we call the embedding vector.

So now we have, in this case, the embedding vector, which is the numerical presentation of the texts or the documents that were split up,

that were transformed into this different dimensional space.

Okay.

But the good thing here is, in most cases, when we talk about embedding in general, it has vectors, as you can see here,

but also attached to it, we have the original text splits.

Okay?

So, they are attached together.

So, we have the representation of the numerical representation, but also the original text attached.

So, would be how internally, in most cases, the embeddings are organized.

And as I said, if we have a database, that means then something has to happen to that data.

So, The data is just there for the sake of being there, right?

100

So that means we have to query that data.

101

We have to query the vector database, the vector store.

102

So how does that really work?

103

Well, the idea is very simple.

104

You have to remember that this is all about the formats that we have here, the differences in formats.

105

Remember, this unstructured data, it's now being saved as embeddings or vectors.

106

Essentially the same thing, really.

107

And then when we are then trying to query our vector store,

108

obviously the querying, it will come in as a natural language spoken,

109

which needs to be transformed into something which then the vector database will understand to be able to query.

110

What that means is as follows.

111

If you look at this diagram here, I'm going to put all the thing together here.

112

So, the idea is that, remember, we already have this embedding vector here, right?

113

Now, when a query comes, a question, that also needs to be transformed into embed.

114

So, essentially, into a format that the embedding system here will understand, the vector space will understand.

115

Makes sense.

116

So, as I mentioned earlier, we need to create what we call an index.

117

So the index is another data structure that is created by the database, in this case, the vector database.

118

Vector database uses various algorithms to create or to build an index that makes retrieval,

119

ah, important, that makes the retrieval more efficient.

120

And that is what is between the embed, which is part of the query question, has to be transformed into embed so that then,

121

so that this data structure here that is created, essentially an algorithm created by a vector database,

122

is able then to query and find what needs to be found,

123

find the similar or search and compare all of the entries in the vector database.

124

So essentially important to note here, an index is what makes the whole querying the vector database fast

125

and efficient because an index is a data structure that facilitates this search process.

126

So this query or question comes in is transformed into an embedding.

127

And that will be then used to search and compare all the embeddings entries in the Vector Store and the Vector Database.

128

And what will happen internally, it will go ahead and pick the most similar entry,

129

which is then what is going to be sent out to be processed for further processing.

130

So it's going to go ahead and find something that's similar to this embed because it was transformed, and it's all numbers, mathematical numbers.

131

It's going to look what is similar to whatever this embedding that was transformed from the query.

132

And then spits out, per se, picks the most similar entry as such represented here.

133

Does that make sense?

134

And so what happens then, because the whole idea is to use large language models such as GPT and so forth, with that result, the processing,

135

what will happen then, the most similar embeddings that were received from previous process, the most similar entries are here now.

136

These are the ones that are sent to a large language model, take the most similar entry found,

137

and pass it on the large language model along with the question query and so forth.

138

And then, of course, that's when we get the result or answer.

139

So, at the end of the day, that is the whole basis

140

or the whole reason why we have vector databases is to save all that information from raw data,

141

if you will, into embeddings, essentially a vector with a little bit more information, more metadata.

142

And that is what's actually taken

143

and fed into a large language model to with queries

144

or any other questions that need be

145

and the large language model as an algorithm a complex algorithm will know how to take

146

that and spits out the result or answer

147

that we're looking for in the past videos or

148

so i introduced a new word a new concept called embeddings now let's look at the differences between embeddings

149

and vectors the truth of the matter is embeddings and vectors are actually referring to the same thing.

150

However, there are some differences in the definition as well as in roles.

151

When we talk about a vector, as we saw earlier, mathematically speaking, vectors do represent data.

152

It's a mathematical representation of data in an n dimensional space.

153

So, each dimension would be, in this case, a feature of that data as we saw.

154

Now, embeddings are a specific type of vector that are used in most cases in machine learning and artificial intelligence,

155

because they are a little bit loaded, if I may say, than just a simple mathematical vector.

156

Now, the terms embeddings and vectors are often used interchangeably in data science, particularly in machine learning.

157

But they have distinct definition and roles.

158

Let's go ahead and take a look at them.

159

So here are the key differences and relation.

160

So vectors are used in different fields per se.

161

So if you say vector, as we see, we can look at vector mathematically, but also they are used in computations and so forth.

162

So if you were to define vectors,

163

vectors could be considered as generic and are used for a wide range of applications, right?

164

So they are really good at handling numerical computations.

165

Whereas embeddings, when we talk about embeddings,

166

we are talking about a data structure per se that allows this mapping of write data into the vector space, into a vector space.

167

But most importantly, this is very important,

168

is that embeddings also preserve or attach the actual vectors with semantic relationships.

169

Meaning, this is very, very crucial because this is the core of AI as we know it.

170

Artificial intelligence, machine learning, and so forth.

171

Is that we need to have these vectors that also have some sort of metadata attached to it.

172

So, embeddings are that, allows us to have this raw data into vector space, but also we are attaching the semantic relationships meaning to that data.

173

So, the bottom line here is that embeddings are vectors, but not all vectors are embeddings.

174

So, embeddings are vectors that encode semantic similarities between the items they represent,

175

which in this case makes it really great for AI applications.

176

In a nutshell, the term embedding refers to specifically vectors

177

that have been learned in a way that encodes semantic similarities between the items they represent.

178

Think of vectors as vectors, essentially, but loaded with something else.

179

This something else being semantic relationships, meaning that it's encoded to those vectors.

180

So that is the distinction.

181

So whenever I say vectors as it relates to machine learning, AI, vector databases, and embeddings, I'm essentially talking about the same thing.

182

So I may be saying embeddings and vectors is essentially the same thing, but now you know the differences.

183

So now you have pretty much all the tools that we need to move forward and start building our AI -based application,

184

this hands -on that I have for you, which is going to be essentially using the large language model

185

and the data that we have saved we will have saved i should say into our vector database

186

so the idea is that we're going to take certain pieces of documents

187

or large documents or document

188

and then we are going to chunk them up cut them up into smaller pieces

189

and save those into a vector database along with

190

that we're going to be saving of course with embeddings now you should know what an embedding is

191

and so now when we have those embeddings

192

that is what's going to constitute part of this vectorization of our data

193

which is really important as you know

194

because it's easier to do the similarity search in this vectorized data realm per se

195

and now we're going to be using a large language model attach

196

that with our vector database which knows how to go

197

and find things really fast so that we're able to converse with our data, meaning we're able to chat, ask questions pertaining to our data,

198

in this case, our chunks of pieces of our data or our documents and get answers.

199

So now we're ready to go ahead and do that.

200

And let's go ahead and get started.

201

So far, we're able to do the most important thing, which is load the source data.

202

In this case, we have just a few documents or just a few pieces of documents, although they're small, but we could scale it up, right?

203

So what happens now is that we load that into our vector store or vector database.

204

In this case, we're talking about the Chroma DB.

205

And what happens then is we're able to then pass a query in form of text,

206

and that is indeed transformed into an embedding or a vector space,

207

which then is sent out into the vector store or vector database to do the search.

208

So it does the similarity search, and then we're able to see the result.

209

So we get the most similar pieces of that document.

210

So this is wonderful.

211

But what we really want to accomplish is something that is a little bit fuller, because the whole idea behind

212

Vector Databases is that we want to be able to query them with the help, of course, of large language models or models in general.

213

That's the whole idea.

214

So what we want to be able to do is, for instance, we have a bunch of documents.

215

We want to be able to take those documents and split them into smaller chunks.

216

And then once they're split, we are going to transform those into embeds like embeddings.

217

And these embeddings are the ones that are going to be stored in the vector store.

218

And as you know by now, we're able to pass these embeddings, which are vectors that not only have attached them vector,

219

but they also have the original test, the splits that will happen, the splits that correspond to that vector space.

220

So that's what we have so far.

221

But what happens next is as follows.

222

We want to be able, again, you've seen this before, we want to be able to have a question again,

223

which is going to be transformed into an embedding.

224

And then we have the index here.

225

And the index is going to help make this whole process of querying

226

and searching and comparing all the entries in the vector database way faster.

227

Okay.

228

And so what will happen internally then is that we are going to, according to the question or query

229

that was passed in is going to be able to search

230

and compare all the entries in the Factor database and then pick the most similar entries,

231

which is exactly what we've been doing and then showing that.

232

But now we're going to take this to the next step because we don't want just to see, for instance, if I ask,

233

you know, tell me about physics and just give me the chunks in the parts of documents that are related to physics.

234

That is fine.

235

We got that.

236

But we want to go even further to say, tell me about physics.

237

And then what will happen is that we're going to get this comparison that happens in a vector to b.

238

And we're going to pick the most similar entry.

239

This is where we have stopped and get the result so far.

240

We want to take this result, the similarities that were these embeddings that are similar results, right?

241

We're going to pick those.

242

And then we're going to pass them through a large language model.

243

So it's going to take the most similar entry found

244

and pass it onto the large language model with the question and query and everything.

245

So now we have the first question that was asked by the user and went through the whole process.

246

And we're going to take that question and put it all together with the most similar entry.

247

In this case, the chunks of those documents

248

that are most similar related to that question query and pass

249

that through the large language model and then we're going to get the actual answer or the result.

250

So that is what we're going to be seeing how to do in the next video.

251

So now we have this data directory with new articles.

252

So this is just a bunch of articles that are found online.

253

Each one of these articles is inside of its own document, txt document, as you see here.

254

So we are going to use all of these documents.

255

So essentially, we're going to load all these documents into a database or a vector database.

256

But before that, we're going to split them up

257

and then create embeddings as we split them up and then put them in a database,

258

persist that information, and make sure that we're able to do what we did before, which is to query the database,

259

the vector database space, the vector database space, and then be able to extract that information that we received,

260

which is the query as well as the results we got from the query querying our vector base

261

and then push all of that into a large language model.

262

In this case, we're just going to use GPT, use some of their models to actually to be able to ask questions and receive answers,

263

direct answers, not just approximations from our query.

264

So I went ahead and created, of course, a new file here called vector db llm large language model okay

265

so first i'm going to go ahead and do some imports here

266

so most importantly remember i'm getting chroma utils and embedding functions

267

that way we can also pass along not default embeddings

268

we're going to use open ai embeddings just like what we did before

269

i have the chroma db of course the open ai all of

270

that and then let's go ahead and pass in here these variables for our .env file.

271

So load all of those environment variables, which means then I'm going to be able to create our OpenAI API key.

272

Then I'm going to go ahead and get the OpenAI API key.

273

Make sure that you have that under your .env file, just like what we did before.

274

All right.

275

So the next part here, I'm going to create the OpenAI embedding function.

276

So I'll just say OpenAI embedding functions, and I'm using the text embedding three small model there.

277

And of course, I'm passing the OpenAI key.

278

So we're able to actually go and get through the API, get the embedding object functions that we can use here.

279

Okay.

280

So next we're initializing Chroma client with persistence.

281

So this is very important because as you can see here, we've seen this before.

282

We have client, ChromaDB persistent client.

283

We're passing just the path where we want this to be.

284

I'm going to just put actually something like DB so that it goes under this folder here, DB.

285

Okay.

286

We have the previous ones, but that doesn't matter.

287

It's going to go ahead and create a new one.

288

And we have a collection name here.

289

We call document QA collection and I can say that's very good

290

and for then here I'm creating the client so get

291

or create a collection the action

292

so here we're creating the actual collection we pass the collection name

293

which we specify the top

294

and we pass along the embedding function in this case we're going to use the open ai embedding function

295

which we instantiated there so now we're going to put together here all the functions

296

that will allow us to do what we want to do

297

so what we really want do is you want to be able to load all of this data, which is this new articles, into our database.

298

Okay.

299

So to do so, I'm going to get the...

300

I already have the function that does all of that.

301

So I'm going to just go ahead and get it.

302

This is the function, of course, that loads documents from a directory.

303

So we pass the directory path.

304

So we're passing here the directory path, and I have this print that says loading documents from directory.

305

So essentially here we go into this directory,

306

find this directory where we have data and new articles and then we find all of these files that are, they have txt at the end.

307

As you can see if I hover over, these all of them have txt at the end.

308

It's kind of hard to see but there we go.

309

So going through all of that and then create the actual document and then we return the document.

310

So each time we loop through, we are appending the actual document format into ID and...

311

It's going to be the file name and the text is going to be the contents of that document.

312

So we're creating those documents and then we'll return.

313

Now the next part here is going to be to actually split those texts into chunks, right?

314

Because we need to make sure that those are small pieces, those chunks.

315

These are the ones that are going to be transformed into embeddings and then added into the vector database.

316

So we have another function for that, which is this.

317

So splits text, we pass the actual text to be split and the chunk size.

318

For each chunk size, we're going to say a thousand.

319

That's just the number I got.

320

You can make it smaller or larger.

321

And chunk overlap is 20.

322

So this here is to allow for the context to not be lost.

323

Now this is a little bit more advanced, but the idea is that if we have different chunks,

324

larger this overlap, the better the context for each one of these pieces we are going to preserve.

325

Okay.

326

So essentially we're going through everything, the text that we are receiving, and we're just splitting into different chunks, into different chunks and minding,

327

of course, the overlap to be there.

328

And then we'll return all those chunks.

329

Okay.

330

So now let's go ahead and do the actual loading of everything.

331

So let's see here.

332

So I'm going to load the documents from our folder directory here, data, news articles that's all I'm doing here.

333

So I created this function first because it would be easier for us to call them.

334

As you can see here, we're actually calling load documents from directory.

335

Okay, so we have here, first of all, we're specifying the directory path.

336

So we say it's data news articles, which is exactly what we have here.

337

If you change these names, of course, make sure you change this path as well.

338

So we're going to this articles, new articles.

339

In fact, I made a mistake here.

340

let's see, it should be new news article, so I'm going to just say it's new article because I made a mistake here,

341

it should be news, but it says new, that's okay alright, so there we go, we have that set up,

342

and then we are getting all the documents by calling load documents from directory, because remember, this is returning load documents from directory,

343

we pass in the directory path and returns those actual documents, alright, so now we're using those functions we just created here.

344

And now let's go ahead and just for us to see, I'm going to just print and I'm going just to go ahead

345

and say loaded documents and get the length of those documents.

346

This comes in as a list, as you can see here, and I can just get the length.

347

Let's go ahead and run this real quick.

348

Okay, loading 21 documents.

349

Went ahead and created about 21 documents, actually loaded 21 documents, as you see here.

350

So we have probably about 21 documents.

351

Very good.

352

Now I can also print.

353

Just want to see the actual documents like this.

354

We should get a list of documents which we can also see.

355

Okay, you can see now we have this list of all of these documents that we have downloaded.

356

It's a very long list of everything.

357

All right, very good.

358

All right, so we know that the documents are loaded

359

because we can actually see all the documents here when we do print as you see there.

360

Next we need to split the documents into small chunks.

361

So here's what we do.

362

We create a list, create an empty list and we loop through all those documents and we call this split text.

363

So this split text we created, we pass in text, the chunk size, and the chunk overlap.

364

So it will return a list of those documents, the splits.

365

Okay, so that's all we're doing here.

366

I've added some of these prints here so that we can see what's happening as these are called.

367

So now that we have those chunks, we need to transform them into embeddings.

368

So essentially, we're following exactly what we've seen before, what we've talked about.

369

To do so, I'm going to also create a function that does that and I'm going to put that right here.

370

But before we do that, below here, I'm going to say client,

371

openAI, and then I'm going to pass the API key, which I know I have access to that,

372

which is this guy here.

373

So now I can go and create those embeddings.

374

So get OpenAI embeddings.

375

So we pass the text that was going to be transformed into embeddings.

376

So we use here client embeddings .create.

377

Now how do I know how to do this?

378

Well, we'll go back to OpenAI documentation.

379

We can go and let's go to tutorials.

380

So here it will show you exactly.

381

Let me just make sure this is.

382

Let's just go to quick start.

383

I think that's the best place and it tells you exactly how to get started.

384

Install Python.

385

and all that great stuff.

386

And if you scroll all the way down here, we have this chat completion.

387

Now, they have changed a few of their, they've changed some of the way we call things, but it's okay.

388

You can always come here to make sure that you are up to date.

389

So the idea is that we instantiate our client.

390

In this case, I passed along the API key.

391

And then we just say client, chat completion and create and then I'm past the model and then the messages.

392

So we have the role and we have the role for the system and role for user.

393

So this is the prompt for the user, the question that comes in, as you will see in a second.

394

And this is for the system to say, okay, what is it that I want the system to know?

395

That way when we ask questions, this is going to be contextualized.

396

So that's where I got

397

it from this is where you should come to get understand more how to use OpenAI to do things.

398

Now this is actually for the completion create

399

but I can go to embeddings by the way which is the one that I wanted to show you.

400

Let's go to embeddings if you click embeddings you should have the code

401

that says client .embeddings .create and then you have the input and then you have the model that you pass along.

402

All right so that's what we're doing here so I say client embeddings the input the text model.

403

I got the text embedding 3 small because that's all we need.

404

And then we go ahead and from this response here, I get the actual embedding for the text that we passed in,

405

which then is returned here.

406

So it's going to be a list of floats.

407

As you know, all embeddings are vector spaces, really, and so they all come as floats.

408

So now that we have the embedding function,

409

now I need to take each one of those chunks of this document that we created here and create the actual embeddings.

410

All I'm going to do is I'm going to go through this chunked documents, which we have access to here because we got those chunked documents,

411

and we're going to loop through.

412

So as we loop through, we are going to create the actual embeddings.

413

So we're going to generate embeddings for the document chunks.

414

That's all we're doing here.

415

So for the getOpenAI embedding, here we pass the doc, the text field in our doc,

416

as we loop through, and then we pass to the field embedding for that document, for each one of those documents.

417

So now that we have the embeddings, notice the process.

418

We have the documents, we chop it up in small pieces, and now we have to create those embeddings,

419

which are then used to be added into the vector database.

420

You see the flow?

421

Right.

422

And so now we're going to insert into our database because we have those embeddings.

423

So I'm going to copy this and pass in there.

424

So now we can, now you can see here, we have our chunk, chunk documents.

425

So we're going to loop through and start inserting all of that information collections.

426

And so we use collection, appsert, and we pass ID.

427

So this is the same thing we've seen before.

428

We pass the actual documents.

429

And notice here, now what we're doing is we're saying embeddings is doc embedding, because that is what we have been doing here, right?

430

So the document, each document now should have its own embedding as we loop through, because we're calling the getOpenAI embedding,

431

we're passing the text, that chunk of text, and then we create embedding and put that into the document.

432

So the document now has embeddings, that's why we say embeddings.

433

As we loop through, or rather as we loop through each one of those chunked documents,

434

we're going to insert each one and add it in ID, as well as adding the text, but also importantly, appending, in this case,

435

the embeddings along for this vector, this document, as we insert everything.

436

So if you open DB, you can see we have the previous databases here, but now if I were to run this,

437

if all goes well, I'm going to be able to actually insert everything into the document, into the database.

438

So let's clear this and run again.

439

You can see here splitting embeddings and soon you should see something happening here.

440

So it's generating embeddings because it will take a little bit because it's a lot of information.

441

you can see now inserting chunks into db okay let's open this all right

442

so this was created earlier but now you can see

443

if you go to chroma persistence

444

or persist you can see we have our chroma persist database here

445

so what i can do again command p or command shift p

446

and i'm going to say open database and let me go ahead

447

and get the chroma persist because we have a few here so

448

So let's see.

449

I think this is the one.

450

Okay.

451

So I deleted everything.

452

Let's start over.

453

I think that would probably be better.

454

Let's run this again.

455

Okay, so splitting, generating embeddings.

456

That's very good.

457

You can see now database was created.

458

Let's give it a moment.

459

We went ahead and inserted everything into the database, right?

460

Let's see if this is true.

461

So let's go back and say

462

Command -Shift -P.

463

Let's say open database.

464

That's the only one we have right now.

465

Let's open that.

466

And let's go SQLite Explorer just refresh just in case.

467

And you can see if I were to run any of these, you can see that we have all of these.

468

Okay, see all of these are actually the database with all of the text.

469

And you can see this week, apps, Apple, and very good.

470

So all the chunks are actually added to the database.

471

And I can continue to run something else.

472

For instance, search, we can see that we have full text search

473

and each one of these is different chunks with all the overlap and everything.

474

So that is beautiful.

475

Okay, so this actually works.

476

So let's look at collection.

477

That should give us document, QR collection and dimension and everything.

478

Okay, so we're able to get all those chunks converted into embeddings and then take those embeddings and added them into ChromaDB, be the vector database.

479

So we know that we have our database, our vector database, and has all of the embeddings.

480

Everything is good.

481

So at this point here, what I would do actually is I would go ahead

482

and comment all of this out because we don't want to make it so that every time we run this,

483

we ended up doing the same thing.

484

We don't need to do that anymore.

485

So I'm going to go ahead and just comment all of this out.

486

Let's see, all the way, pretty much all the way here.

487

So everything we've written so far, it's okay, just comment it all out.

488

So the next thing we'll do is we are going to go ahead and get the function to query documents.

489

So I'm going to put that function here. So it's very simple.

490

These query documents, once it's called, we pass the question, which is going to be the question, the query, and then the end results.

491

In this case, we just want to at this point, but this is actually not going to be relevant, but I'll just keep it there

492

Because the idea here is that we're going to attach, do this final part, which is we're going to attach all this information,

493

all these embeddings that come there similar to the large language model as well as the question.

494

That's why we're passing this to then make it so

495

that we can ask the large language model to questions about the data that we have.

496

Okay.

497

So we're calling collection query, passing the query text question, which is what we have here.

498

And I'm also passing the end results.

499

Now this is going to be redundant.

500

It would only work or would be needed if we were just to go ahead and do what we did before,

501

which was just to get some of the documents that have indeed the close that gives us the distances between the documents.

502

But we have moved on from that.

503

So this is no longer necessary, but I'm going to leave there anyway.

504

The other thing you can do also, you can actually create the query embedding by calling the getOpenAi embedding

505

and passing the question so we can make it that as embedding.

506

In this case here, I'm just going to go straight and pass for our query here the actual text question.

507

Okay, either way.

508

All right.

509

So then I get relevant chunks here.

510

What I'm doing here after I get the results here, I'm just getting all the documents and then just retrieving the relevant chunks that I need.

511

And I'm printing here, returning relevant chunks so that we have something to look at.

512

And then we'll return the relevant chunks.

513

Now we have the relevant chunks.

514

Let's go ahead and I'm going to just call that so we can see something.

515

So at the bottom here, I'm going to add a question.

516

So this is about, tell me about AI replacing TV writer strikes.

517

So I know this because in the document here that we have,

518

there's something related to writers, right?

519

AI replaced TV writers.

520

So I just ask a question related to that to see if this would actually give us something useful.

521

So I call relevant chunks.

522

I call the query documents, which is what we just created here, and then does everything.

523

It should return relevant chunks for us.

524

So in this case here, these relevant chunks, it's going to be a list of documents.

525

I'm going to go ahead and print them so we can hopefully see.

526

Make sure again that you commented out everything else so that we don't have to go through the whole process again,

527

which takes a little bit of time because everything is in database now.

528

So we can just be.

529

Okay, and just like that, it went ahead and said returning relevant chunks in the must -watch final season of Succession,

530

and it gives all this stuff.

531

So, very good.

532

So, it's giving something related to picture and television producers, refused to engage with the propo, ed, and all the things.

533

So remember, these chunks may look a little bit weird because when we created them, when they were split up,

534

we made sure that we have the overlap to make sure that we don't lose the context.

535

Okay?

536

So there we go.

537

So we went ahead and got the correct information, which is relevant to AI and writing and so forth.

538

Okay?

539

So this is for writer's strike.

540

If I were to ask something like, let's say Databricks.

541

Okay?

542

Databricks acquires AI.

543

I'm going to ask a question about that.

544

Tell me about Databricks acquiring acquisition of AI, something like this.

545

Okay, let's go ahead and save.

546

I'm going to go ahead and clear this and run.

547

And you can see it should go and get Databricks.

548

Voila, today announced that it has acquired Okera.

549

Very good.

550

So this is indeed working, which is exactly what we want.

551

So the next step would be to take these chunks, because remember, these are pieces of data,

552

pieces of information, that then we are going to be able to take this chunk

553

and then pass it through the large language model, all right, so that we can ask actual questions.

554

Instead of just getting these chunks, we want to get the actual answer to say, yes, Databricks did X, Y, and Z, everything direct.

555

So that is the beauty of this whole workflow from having the documents as we saw, chopping them up into smaller pieces

556

and then passing them through the process of creating embeddings and then putting all of that into database.

557

At this point, we're able to get the documents that are related to that question, which is exactly what we've been doing so far.

558

But now we're going to put all of

559

that information with the question through the large language model and then get the direct answer.

560

So we're able to get the relevant chunks, which is a win.

561

Now we're going to get these relevant chunks with the query and pass that through the large language model.

562

So we're going to use, of course, OpenAI API here.

563

If you go back and go to Kickstart, and you can see that we can actually,

564

we have declined already, and we're going to invoke the completion endpoint here.

565

So client, chat, completions at create.

566

So this is what we're going to do.

567

So what we'll do here is we're going to train our system to say, okay, you are a person who does X, Y, and Z.

568

So I want you to take the information I'm going to give you right now

569

and then answer because you are really knowledgeable.

570

And then we pass for the user this is the question, the query, we're going to pass it along here.

571

And then we should have that information, right?

572

The actual result or the actual answer.

573

Okay.

574

So how do we do that?

575

Well, first let's go ahead and get the function to generate a response from OpenAI.

576

Okay.

577

So this is the function to generate a response.

578

So we have the question that needs to be passed along as well as the relevant chunks.

579

So the first thing we need to do here is we're going to add a context.

580

So the context is to say is what we're going to add with our prompt

581

so that the system knows what is the context, what is it that I need to focus on,

582

what kind of skills I have to have to be able to do X, Y, and Z.

583

So essentially that's what we're doing here.

584

So essentially how I'm saying here you're an assistant of question answering tasks.

585

Use the following piece of retrieved context to answer the question.

586

If you don't know the answer, say that you don't know.

587

Use three sentences maximum and keep the answer concise.

588

So this is something that I've wrote for the prompt.

589

You can modify this to whatever else you want.

590

Now here I'm passing the context.

591

So the context I'm passing here as a variable, which is going to be this.

592

So the context is going to be all of the relevant chunks.

593

So it's essentially saying, hey, take all of these relevant chunks that we're going to pass through here

594

and then answer the question that I'm also going to pass through. Because you're knowledgeable.

595

This is your prompt now then we use the response then we use the completions

596

and api here client chat completions

597

that create we need to pass a model in this case just pass gpt3 .5 turbo

598

because it's cheaper and is just really good now there's gpt4

599

and by the end by the time you finish watching

600

this course maybe something way way better you can use gpt4 oh it doesn't matter

601

so this is just an example you we're just going to use gpt3 .5 turbo

602

and then in the messages here this is where we pass the role system

603

so essentially this is exactly what we had here

604

okay so in fact i just got it from here

605

so for the prompt here for the system to say hey system

606

this is your context this is what you're what you're expert in we pass the prompt

607

which is what we added here which which in turn we have the context

608

and the question to be answered by this

609

system by the system

610

and then the user here this is where we pass the actual question the question

611

which is going to pass here very good

612

and all we have to do now because we know

613

that this is the beauty that response requires a question

614

and the relevant chunks to do the magic

615

so we already we already have all

616

that now all we have to do really is just call the generate response

617

so answer to generate response and pass the question

618

which is what we have here and the relevant chunks

619

which is what we have here because we get that from query documents and pass the question.

620

Okay, so now I can just get rid of that and just say print answer and print the answer as such.

621

Look what will happen.

622

So I'm going to, so before we had this, we get all the chunks,

623

that's fine, but now we're going to use the large language model to give us magic.

624

So we're going to object with content that says Databricks acquired Ocara,

625

a data governance platform with a focus on AI.

626

So now I can probe in and get the actual content instead of just getting the object.

627

Okay.

628

So I can say here, answer, get the content.

629

If I save now, if I do the same thing, you can see that now instead of the object, which if you scroll down,

630

the object has all these other metadata too.

631

So, it can be handy.

632

This one should get the same answer mainly, but now we get the actual answer.

633

Look at this.

634

I can even ask something like this.

635

Tell me about the strike in the company or something like that.

636

Because we remember that the writers, there should be a document about about Riders strikes and so forth.

637

So let's run this and see if it's going to work.

638

Okay.

639

Look at this.

640

The Riders Guild of America is striking for the first time since 2007.

641

I can ask other questions.

642

For instance, let's say Fintech's Fintech Space or SpaceX Starship.

643

Something related to that.

644

Tell me about SpaceX.

645

Okay, look at this.

646

Space X Starship is a super heavy launch system, blah, blah, blah.

647

And there you go.

648

So now we can ask questions about all of those articles, all of that information.

649

So I can say, give me a summary, let's say brief overview of the articles, concise.

650

This is pretty fun.

651

okay there we go the articles cover topics such lifts equity

652

uh first quarter results potential upsides amidst its new strategic direction

653

and impact of of downed rounds

654

it also mentions upcoming events like tc city spotlight index venture tech crunch off all of these great great things

655

so now we're able to go through the whole flow of taking documents

656

and splitting them up creating the actual embeddings and then adding those embeddings into the vector store and its vector database, Chroma, that's what we're using,

657

and then be able to, of course, query that, but then get all those results, and then with the question,

658

pass that through the large language model so we get a very concise answer within the context of our documents.

659

And there you have it.

660

We just finished the whole workflow.

661

So we're able to get documents and split them up

662

and then create embeddings and then save those embeddings along with the original text splits into the vector database.

663

So this is wonderful.

664

But then we took things a little bit further.

665

So we took the question, we were able to get the query in question and turn that into embedding.

666

And then we're able then to search and compare all entries in the vector database and then pick the most similar entry.

667

And then the final part, which is the whole idea behind what we've been learning about vector databases and so forth,

668

is that we then pick the most similar entries.

669

Those are picked and they were passed through an LLM, a large language model.

670

In this case, you were just using the OpenAI large language model,

671

model in this case, but you could use any kind of large language model or model out there, okay?

672

And so we put all of these similar entries with the question query, and we passed through the large language model,

673

and we got the actual result.

674

This is the culmination.

675

This is the reason why, main reason why, that we have vector databases, so that they can work with large language models to give users,

676

in this case, create results, direct results and responses.

677

I hope this is making sense and you're seeing the value in learning vector databases.

678

All right, my friends, I hope this was helpful.

679

I hope you learned a lot in this series, starting from the first series,

680

which again, you can go and find somewhere here and you know pretty much all the fundamentals of vector databases.

681

You know how they work, you should know how they work,

682

how to create one and how to now use all this knowledge to actually create something useful.

683

In this case, we're able to use vector databases along with large language models to ask questions about documents,

684

about our own data, in this case, documents.

685

Now, I'm using documents in this case, but could be a database, could be anything else out there,

686

any other unstructured data that we can pass through a large language model, this would also work, right?

687

So I hope this was helpful.

688

I hope you were able to see how vector databases really work and how amazing they are.

689

And I think it's important for you to learn these concepts, even if you are not wanting to be an AI engineer,

690

but understanding these things, understanding these concepts,

691

I think it's going to really put you at a very high level when it comes to understanding how AI works,

692

which is exactly where we're going.

693

AI is not going to go back.

694

It's moving forward.

695

So I think you should really consider continuing learning about vector databases, machine learning, a large language model,

696

understanding this whole AI world.

697

The other thing is this two -part series

698

that I put together for you is actually part of a larger course where I go deeper into vector databases,

699

understanding really how they work and how to use them

700

and leverage them so you can actually start building AI -based applications that are very efficient,

701

that are well put together and scale better, right?

702

So it's part of a larger course, which is available.

703

And if you're interested in learning more about vector databases, then I'll leave a link in the description where you can go and check it out.

704

All right.

705

Thank you so much for your time.

706

And I really appreciate your time.

707

If you're new here, please consider subscribing and liking this video and make sure to share with people, friends, colleagues, grandma, grandpa as well,

708

if they are wanting to learn more about ai large language models

709

and programming in general this is the place to be

710

and let me know what you think of these videos what i'm doing here

711

because your feedback is always always always welcome i always love to hear from you

712

so well i'm here for you so thank you so much again and till next time be well

הורד אפליקציה

כל מה שאתה צריך כדי לדבר בשטף

הגיית AIציון לכל משפט

תרגול IPAשלוט בכל צליל

אוצר מיליםבנה את מאגר המילים שלך

משחק אוצר מיליםלמד תוך כדי משחק

סרוק QR

Download on theApp Store Get it onGoogle Play

FOR YOU

המלץ על סרטון

197 משפטים

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

Daily ConversationUniversity Students

407 משפטים

Why AI Tokens are so Expensive - Computerphile

Daily ConversationUniversity StudentsWorking Adults

30 משפטים

Research Source Mastery: How to Find & Verify Credible Data (2026)

Daily ConversationUniversity Students

125 משפטים

Introduction to statistical prediction models

Daily ConversationUniversity Students

288 משפטים

Art Direction Techniques You Can Use

Daily ConversationUniversity Students

219 משפטים

Vocabulary for Data Engineers - Data Engineering 101

Daily ConversationUniversity Students

64 משפטים

P2 什么是机器学习【2024公认最好的 | 吴恩达机器学习 | 教程 | Machine Learning Specialization(超爽中英!)】

Daily ConversationUniversity Students

74 משפטים

LangChain in 5 Minutes (Explained Clearly)

Daily ConversationUniversity StudentsWorking Adults

אוצר מילים והערות הגייה לשיעור הזה

שיעור הדיבור הזה ברמה C1 מבוסס על הסרטון “Master AI in 60 Minutes with Vector Databases!”. המילים שחוזרות הכי הרבה: vector, embedding, document, database, chunk. בסרטון הזה יש 712 משפטים ו-8815 מילים לתרגול Shadowing. הדיבור נמשך 53:24. הדובר מדבר בקצב טבעי של כ-165 מילים בדקה, קרוב לשיחה יומיומית. 86% מהמילים נמצאות בין 3,000 המילים הנפוצות ביותר באנגלית; כדאי ללמוד את השאר לפני שמתחילים.

אוצר מילים מרכזי בסרטון הזה

15 המילים המתקדמות ביותר בסרטון, עם ההגייה והמשמעות שלהן:

מילה	הגייה	משמעות
vector שם עצם	/ˈvɛktɚ/	וקטור
embed פועל	/ɪmˈbɛd/	לְשַׁבֵּץ
transform פועל	/tɹænsˈfɔɹm/	לשנות
insert פועל	/ɪnˈsɝt/	לְהַכְנִיס
similarity שם עצם	/ˌsɪm.ɪˈlæɹ.ɪ.ti/	דִּמְיוֹן
attach פועל	/əˈtæt͡ʃ/	חיבר
numerical שם תואר	/n(j)uːˈmɛɹ.ɪ.kəl/	מִסְפָּרִי
algorithm שם עצם	/ˈælɡəɹɪðm̩/	אַלגוֹרִיתְם
concise שם תואר	/kənˈsaɪs/	תַּמְצִיתִי
float פועל	/fləʊt/	צָף
mathematical שם תואר	/ˌmæθ(.ə)ˈmæt.ɪ.kəl/	מָתֵמָטִי
folder שם עצם	/ˈfoʊldɚ/	תִּיקִיָּה
spit פועל	/spɪt/	יָרַק
scroll שם עצם	/skɹoʊl/	מגילה, מְגִלָּה
kitty שם עצם	/ˈkɪti/	גּוּרַת חֲתוּלִים, גּוּר חֲתוּלִים

Phrasal verbs שתשמעו

מילה	משמעות
come back פועל	שָׁב

משפטים שכדאי לחזור עליהם

משפטים קצרים ושלמים מהסרטון שאפשר להשתמש בהם בשיחה יומיומית:

Let's look at an overview of the whole process.
Let's go ahead and take a look at them.
We're going to pick those.
We have the previous ones, but that doesn't matter.
That's just the number I got.

דקדוק בסרטון הזה

המבנים שהדובר משתמש בהם הכי הרבה, עם המילים המדויקות מהסרטון:

מבנה	בסרטון
הווה מושלם ממושך (Present perfect continuous) have/has been + -ing — פעולה שהתחילה בעבר ועדיין נמשכת עכשיו	we've been doing · have been doing · we've been learning
הווה מושלם (Present perfect) have/has + past participle — פעולה שקרתה בעבר ועדיין יש לה חשיבות או השפעה עכשיו	we've learned · we've done · we've discussed
סביל (Passive voice) be + past participle — הדגש הוא על מה שקורה, לא על מי שעושה את הפעולה	is actually transformed · is actually saved · be split
משפטי זיקה (Relative clauses) who / which + clause — מידע נוסף על אדם או דבר	databases, which is · vector, which is · embed, which is

הגייה ששווה לשים לב אליה

הדובר משתמש ב-181 קיצורים וצורות מקוצרות, כמו we're, I'm, we've. אמרו אותם בצורה הקצרה, כפי שאתם שומעים אותם.

הצלילים של “th”: algorithm /ˈælɡəɹɪðm̩/, mathematical /ˌmæθ(.ə)ˈmæt.ɪ.kəl/
הצלילים של “sh” ו-“zh”: computation /ˌkɑmpjʊˈteɪʃn̩/, dimension /daɪˈmɛn.ʃən/, dimensional /daɪˈmɛn.ʃə.nəl/, approximation /əˌpɹɑk.sɪˈmeɪ.ʃən/, culmination /ˌkʌl.mɪˈneɪ.ʃən/
מילים ארוכות — שימו לב להטעמה: directory /dɪˈɹɛktəɹi/, similarity /ˌsɪm.ɪˈlæɹ.ɪ.ti/, metadata /ˈmɛt.əˌdeɪ.tə/, numerical /n(j)uːˈmɛɹ.ɪ.kəl/, internally /ɪnˈtɜːnəli/

צלילים שדוברי עברית מתקשים בהם:

/w/ — שפתיים מעוגלות, לא /v/: query /ˈkwɪɹ.i/, workflow /ˈwɝkfloʊ/
/æ/ — פתוח יותר מה-“אֶ” העברית, לא כמו “a” ב-“cup”: transform /tɹænsˈfɔɹm/, semantic /sɪˈmæntɪk/, similarity /ˌsɪm.ɪˈlæɹ.ɪ.ti/, overlap /ˌəʊvəˈlæp/, attach /əˈtæt͡ʃ/
/r/ — הלשון לא נוגעת בשום מקום, לא ה-“ר” הגרונית: query /ˈkwɪɹ.i/, transform /tɹænsˈfɔɹm/, directory /dɪˈɹɛktəɹi/, chroma /ˈkɹoʊmə/, similarity /ˌsɪm.ɪˈlæɹ.ɪ.ti/

איך לתרגל עם הסרטון הזה

הקשיבו לכל הסרטון פעם אחת בלי לדבר, ורשמו את המילים שאתם לא מכירים.
התחילו במהירות 0.75×, תרגלו Shadowing משפט אחר משפט, וחזרו למהירות רגילה כשזה מרגיש קל.
הקליטו את עצמכם והשוו למקור, תוך תשומת לב למילים כמו vector, embed, transform.

מהי טכניקת השאדואינג?

שאדואינג היא טכניקת למידת שפה מבוססת מדע שפותחה במקור להכשרת מתורגמנים מקצועיים ופופולרית על ידי הפוליגלוט ד"ר אלכסנדר ארגואלס. השיטה פשוטה אך עוצמתית: אתה מקשיב לאודיו באנגלית ילידית וחוזר עליו בקול רם מיד — כמו צל העוקב אחרי הדובר עם עיכוב של 1–2 שניות בלבד. בניגוד להאזנה פסיבית או תרגילי דקדוק, שאדואינג מכריח את המוח והשרירים בפה לעבד ולשחזר דפוסי דיבור אמיתיים בו זמנית. מחקרים מראים שזה משפר משמעותית את דיוק ההגייה, האינטונציה, הקצב, הדיבור המחובר, הבנת השמיעה ושטף הדיבור — מה שהופך אותו לאחת השיטות היעילות ביותר להכנה לIELTS Speaking ולתקשורת באנגלית בעולם האמיתי.

טכניקת Shadowing: קרא את המדריך המלא שלב אחר שלב →

Shadowing Practice: Master AI in 60 Minutes with Vector Databases! - Learn English Speaking with Video

בחר מצב תרגול

הורד אפליקציה

How I use AI to learn engineering (anti brain rot studying)

Database Sharding! Designing Data-Intensive Applications chapter 6

The Complete Guide to Hybrid Search in RAG (BM25 + Embeddings + Reranker)

what is a database index?

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

Why AI Tokens are so Expensive - Computerphile

Research Source Mastery: How to Find & Verify Credible Data (2026)

Introduction to statistical prediction models

Art Direction Techniques You Can Use

Vocabulary for Data Engineers - Data Engineering 101

P2 什么是机器学习【2024公认最好的 | 吴恩达机器学习 | 教程 | Machine Learning Specialization(超爽中英!)】

LangChain in 5 Minutes (Explained Clearly)

אוצר מילים והערות הגייה לשיעור הזה

אוצר מילים מרכזי בסרטון הזה

Phrasal verbs שתשמעו

משפטים שכדאי לחזור עליהם

דקדוק בסרטון הזה

הגייה ששווה לשים לב אליה

איך לתרגל עם הסרטון הזה

מהי טכניקת השאדואינג?