Shadowing Practice: The Complete Guide to Hybrid Search in RAG (BM25 + Embeddings + Reranker) - Learn English Speaking with Video

Elige un modo de práctica

Cada modo abre una sala de práctica a pantalla completa.

847 oraciones

Si las oraciones son demasiado cortas o largas, haz clic en Editar para ajustarlas.

In this tutorial, we're going to build a hybrid retrieval system from scratch that combines BM25, dense embeddings, reciprocal rank fusion,

and a re -ranker to then bring it all together.

This is really the production stack that works and that wins in 2026 when you want to build serious production -ready systems.

By now, almost everyone kind of like knows how to build rack systems

and they've dabbled with it and know what effect their databases, but almost no one understands how to build them from scratch

and optimize them and evaluate them for your own data set

which is exactly what we're going to do in this video

and now a quick word of caution before we dive into things

because this tutorial is not going to be like most other RAC tutorials

that you'll find online

and on YouTube in this one we go really deep

this is really for engineers developers who already have a basic understanding of what RAC is

and need to use this technology really professionally within their jobs

or the clients you're working for and you want to build a system that you truly own and understand.

All right, so let's kick things off.

And the first thing that we're going to look into is the actual data set that we are going to use.

Because for any retrieval task, the data is the most important part.

And that's really where we should start.

So we're going to use this data set, the financial QA data set, which is also part of the BEAR benchmarks.

So this is a benchmark for information retrieval.

So let's quickly cover that first, because this is very important to understand because whenever you start with a new data set,

you ideally want to have some kind of like ground truth to decide what makes a good retrieval system

because otherwise we cannot optimize, right?

So what's nice about the Bayer data sets is that they created these pairs of queries, relationships, and corpuses.

So as we go through the code base and go through the examples, this will become more and more clear.

But for now, you should understand that the data set that we are going to use

consists of a whole bunch of financial questions and statements that are linked together.

So in the corpus, there are answers.

And these are similar to your own documents that you would, for example, put in your retrieval system, whether they are company documents,

your sales playbook, customer support questions, pretty much anything that you want to put in your corpus.

That's the starting point.

Now, then also we have questions that can be answered using documents within that corpus.

So let's actually check what this looks like in the code base so we can actually get a feel for this.

So I am currently in the hybrid retrieval folder.

Then under data, there is a download financeqa .py file.

So this is a one -time setup that you can run to download the corpus.

So you'll need to set up your project.

So I have a PyProjectAtomal over here.

I recommend to run UfiSync in order to get this all up and running.

We are using the datasets library here from HuggingFace.

And because of that, we can simply run this and it will load all of these datasets.

So we got the corpus, the queries, and the relationships.

Remember how I was talking about this over here?

So this is a one -time run that you can do.

I already have them downloaded over here.

So for me, it will probably skip things.

So just run that and then it will store that data into parquet files.

So this is a very efficient way to store this data on disk that we can then also later load in.

So that's the data set.

But now let's actually get to the exploration phase.

So I'm going to the one explore data file.

And what I'm going to do is I'm going to use the Jupyter Interactive session here on the side.

So we can walk through it step by step.

If you don't know how to do this, I'll link another video below to explain exactly how I set up my cursor to work like this,

similar for Visual Studio Code.

So we're starting over here with our data folder that we just downloaded that data to, so here you can see the full path to it.

Now we're going to load the parquet files that we just downloaded to get a feel for what this actually looks like.

So if I now look at the corpus, you can see this is currently a pandas data frame.

So if you've never worked with pandas, then this is pretty much an Excel spreadsheet, but then within the Python programming language.

So you can see the table over here.

This is also why I like using the Jupyter Interactive Session because it renders these tables really nicely.

So you can see we have a whole bunch of rows, a whole bunch of documents, and there are three columns, each containing an ID.

Title is empty in this case.

This is a default for the bare datasets, but this particular finance QA doesn't have titles.

So we just have the text.

Now let's look at the first element in here to get a feel for what the corpus looks like.

What data are we actually working with?

So I'm using some pandas notation over here where I'm going to say, look, I'm going to take the text column.

So that would be corpus and then the text, which you can see over here.

And then I'm using the ilog0 notation over here to say, let's now just look at the first element in there.

So let's see, let me actually, if I clear this for you and run this one more time.

So this is the first document that shows up in the corpus.

And on its own, it doesn't really say much to us right now.

So let's actually continue with the queries.

So now I'm going to load the other parquet file, which is the queries.

So let's run that.

And you can see that now this also results in another data frame.

Exactly the same structure.

We got an ID, title is empty, and now we got some text in here.

So let's again look at the first query that is in there.

So if I print this over here, you can see right now, instead of a document with information in it, we have a question.

So what is considered a business expense on a business trip?

And you can also kind of like iterate through this to see, okay, what is actually in here?

Starting a new business online.

Let's pick some random ones.

New business owner.

How do Texas work for the business versus individual?

So you can see these are the actual queries that show up.

Now, then the third data set that we have in here is the QRLs, which is the relationships.

Now, this is where it gets interesting because here we create a mapping between a query ID and a corpus ID.

So if I just look at the first record over here, you can simply say that we have query ID 8.

So question 8 within that query list relates to or has a relationship to the document with the following ID.

And that is indicated by the score is 1.

So what you'll find is that in this QRLs, all of the scores are one, the ones with a zero are not in there.

So these are all of our relationships.

Now, you need to spend some time here to really understand this.

And that's why I also created this image for you to dive a little bit deeper into that.

Because everything that we're going to do later down the line comes back to documents,

queries, and the relationships between them so that when we run the retrieval system, we can see when we ask this question,

100

does our retrieval system, does our hybrid system come forward with the right documents that have a relationship with the query?

101

So make sure you really understand this

102

and then also understand how you can map this back to the code base

103

because this is really the mental model that you need in order to go forward.

104

So then step number two of this first file is we're going to make a simple filter on all of the queries.

105

Because if we look at currently all of the queries that we have,

106

you can see this is a big list over here where we have a lot of queries.

107

But not for every query.

108

We have a document that's associated to it within this data set that we are using.

109

So this is a simple filter that we do where we take the QRLs.

110

So let's see.

111

We take the IDs over here.

112

So these are all the questions that we have and we take the set,

113

which is Python's way to deduplicate a list or a series in this case.

114

So here we have the deduplicated set of that.

115

And then we just do a filter on the test queries.

116

And again, this is a little bit tricky to understand if you've never worked with Pandas, but pretty much what we do,

117

look, if I run this, we now have a filtered list.

118

So you can see that we went from how many did we have?

119

We had 1 ,700 to start with, and now we have 648 left.

120

So that's just a filtering mechanism that we do over here.

121

And then to show you one more thing, if we inspect one query and it's relevant documents, here's how you can see how everything comes together.

122

So let's say we have the following query.

123

So now we have the first test query over here.

124

And if I print that here, you can see what that looks like.

125

So this is where should I park my rainy day slash emergency funds.

126

So this is the first query

127

that we now have in the data set where we also have documents backing it up with the relationship.

128

What we then do is again, pandas notation.

129

We go into the QRLs data frame and we perform a search to check for the query.

130

And this is similar to doing a lookup or a SQL where statement.

131

And now you can see that for the query, where should I park my rainy day emergency funds,

132

which has this ID, you can see that if we apply a filter, we get the document IDs.

133

So bringing this all together, we can now see

134

that this particular query with this ID links to these documents meaning this question can be answered through these documents

135

which is exactly what was shown in the image over here okay

136

so now that you understand the data set

137

and also know how to work with this locally we're going to dive into the actual retrieval pipeline

138

and we're going to work through this step by step

139

so first we'll get into bm25 then the embeddings then rff re -ranking

140

and then bringing everything together so that's going to be the flow of the video building each and every individual component first,

141

and then bringing them together later in the code base to form one cohesive pipeline.

142

So we're going to start with BM25, which is a keyword -based way for information retrieval.

143

And with all of the methods that we're going to cover, I'm going to walk through the code base.

144

I'll explain as much as I can during the video itself

145

but also what I want you to know is that there is a docs folder.

146

And for each and every one of the techniques,

147

there is a additional markdown document with all of the formulas and the math behind it as well, if you want to go deeper,

148

as well as the original papers where all of these techniques came from.

149

But I cannot go into all of the details, otherwise this would literally become like a five to 10 hour lecture.

150

But I'm just going to give you what you need in order to follow along.

151

And then again, if you want to go deeper, the information is there for you.

152

So BM25, also known as sparse retrieval, more on that in a bit, will do keyword overlaps.

153

And then there is a weighting mechanism

154

that is applied to it to actually perform the math to see what we need to retrieve.

155

This is really good for catching exact terms, identifiers, rare words, but it will miss paraphrases.

156

So this is a rather naive way of searching and retrieving information, but in some scenarios can work highly effective.

157

So let's dive into this first and see what it looks like.

158

So if I come back to my code base and then jump into bm25 .py,

159

we have the setup over here and we are going to use the following library.

160

Let me actually open that up for you.

161

So this is a Python BM25 implementation.

162

So this is what we will be using.

163

This is also a dependency in the project.

164

So pyproject .toml, you can see BM25s.

165

So now you know where this import is coming from.

166

And the nice thing about this implementation is that we don't have to worry about the mathematics.

167

We can just rely on what was already built and use this

168

so we are going to start with loading our data directory again

169

so this is the same one we were using

170

and now we're also going to create an index directory

171

so let me show you over here

172

there is an indices folder over here where we are going to store the information

173

so we are going to use that later

174

so for now it's only the paths that are in here so now we're going to simply read our corpus

175

again similar to what we already did right

176

so remember this is the data frame

177

and now we're going to convert it to a different data format to make it work with the bm25s

178

so we're going to create lists out of it

179

so what you can see right now is we'll use pandas to list

180

so we can get the doc ids

181

which is just a list of all the ids

182

and then we also get the text and we do the same

183

so now we've converted so you can see this is just a huge list

184

so now we convert simply from a data frame to lists.

185

And we put all of the documents in there because we want to create, keep the corpus the original size, right?

186

So remember how we made a filter on the questions.

187

The corpus should be really big because otherwise we kind of like cheat on the retrieval task, right?

188

Because if we only leave in the actual relevant documents, we just reduce the size because the more data there is,

189

the more errors a retriever can potentially make to retrieve nonsense.

190

Okay, now we're going to tokenize the entire corpus.

191

And this now becomes pretty easy because we can use the BM25S library for this.

192

So simply what we do is we call the .tokenize method.

193

We put the list of documents in there

194

and we can also filter for English stop words so

195

because it's a keyword based mechanism and it will create calculations

196

and a score really on how frequent the words are occurring in certain documents etc removing stop words can help

197

because these are generally noise so if we now convert

198

that to tokens you can see that process takes two seconds

199

or so and now

200

if i look in here you can see we get the following tokenized data over here

201

but let's actually figure out okay what is this i see some ids i see a vocabulary

202

but this is all very messy to me still right now

203

so let's actually have a look in that

204

so what essentially token tokenized does it's it creates a tuple of token ids

205

and vocabulary items and

206

if we look at the first element in the tokens

207

you can see this is now you it creates one inner list per document

208

So this is the first document in our corpus, but now with simply these IDs.

209

And these IDs, they reference to the vocabulary.

210

So if I now look into the vocabulary, which is a dictionary with a string and then an integer,

211

you can see there is a lookup over here right now.

212

And how you can read this is as following.

213

It's simply a mapping.

214

So because this is the actual first document in our corpus, you you can see that it starts very sequentially, right?

215

So we have zero, one, two, three, four, five.

216

And so you might ask yourself, okay, why is this all the same?

217

But then here you can see, for example, that we have five again.

218

So five is over here.

219

So we can see what is five?

220

Five is training.

221

So you can see that this is now just a representation

222

of how to structure a document using IDs and a lookup vocabulary table.

223

That's how this is set up.

224

So again, you might need to play with this a little bit to figure things out.

225

If we now take the second element over here, you can also see that right now the ordering is going to be completely different, right?

226

And you can always link it back to the original item in the corpus.

227

So for example, if we are over here, remember, right?

228

So let's say we have the corpus and we take the text out of that.

229

So this is just the same exercise.

230

And we take the iLog1 you can see

231

that over here this is the first document

232

so it removes the stop board

233

so here you can see for example you see training

234

and then here is training one more time so

235

that is how this is the five over here

236

and the five over here reference training and we can see

237

that in the document over here so you can see this is something

238

that you just have to go through you have to kind

239

of like play with this a little bit for it to start makes sense

240

and that's why in this tutorial I want to just spend some time to just go deep into it

241

because especially nowadays with ai it's

242

so easy to just kind of like one -shot these systems

243

and oh yeah just use bm25 do a hybrid system

244

and then just like let it run without actually knowing what's going on behind the scenes

245

and there's just no way to understand that without tinkering with this

246

and actually just looking at those ideas and then making those matches

247

so we created the tokens

248

and now the next step in here is to create the retriever

249

so now we can again just use the bm25 as library very simple we have a retriever

250

and then we say retriever .index and we plug in the tokens

251

so this is now how we create the retriever which is ready

252

and loaded locked

253

and loaded with our own data now it's also one one thing to worth mentioning here is

254

that bm25 also has different implementations different methods so you can see

255

that the default method is the lucene in here

256

which is also indicated now again

257

if you want to go deeper into this here is how the calculations behind

258

this work as well as additional further readings and resources

259

that you can dive into now the next thing

260

that we're going to do is we're going to simply persist this index to the disk

261

so we're going to create a directory

262

and then we call can call the retriever save

263

and then we just simply put a path in there of where we want to store our index

264

and then we write that so if we run this

265

and i come over here i already had it

266

but you should now see

267

that all of the data all of the information is stored in here

268

so one of the interesting things about bm25 also is

269

that you don't really need a often need a database for this this can just live locally on on disk

270

and if you for example look at right now

271

so let me actually show you the size of this you can see

272

that it's 33 megabytes on disk

273

and this is using a corpus of 57 000 documents although short documents the interesting thing is

274

that for most use cases out in the wild

275

when people are building these rack systems these information retrieval systems your corpus is almost never

276

that big unless you're working on enterprise applications, multi -tenant solutions.

277

If you are automating within your own company, doing this as a freelancer or an agency to help local businesses,

278

your corpus is almost never this big.

279

So those files can easily just live on a file.

280

You can even put them into version control.

281

You could also see the indexing was really quick, right?

282

It took like one or two seconds to do this.

283

So no database really needed in order to store this.

284

You just load it in memory and use it.

285

Okay, and then go to step number four, actually running a query.

286

So, okay, we created up until this point, it was all prep work, right?

287

But this is where we're actually going to do the first step of information retrieval,

288

where we take a query and we search it against the entire corpus and then retrieve the document.

289

So here is a implementation of how we do that.

290

And again, because we are using the BM25S library, it's very simple.

291

We just call the, we use the retriever and we call .retrieve.

292

So let's walk through this together to see what this looks like.

293

And I'm going to actually take the following query over here.

294

So where should I park my rainy day funds?

295

And then let me also actually put k is 10 in memory over here.

296

So we're going to retrieve the top k is 10 documents

297

what we then do is we follow the same tokenization process

298

on the query as what we did with the corpus

299

so this is simply just turning any query into query tokens

300

so now you can see our question of where should i park my rainy day funds is tokenized as follows

301

and this removes the stop and then just gives a token to every word in here.

302

And then here you can see how that is built up because the query starts with where and then should.

303

And then you can see I is removed from it.

304

And then park is number three, which is index two.

305

You get the idea.

306

This is how you walk through that tokenization process.

307

And now one thing that's also worth to mention here is

308

that the tokenize is a deterministic mechanism over here that will always yield the same output.

309

So this is not using our customized or trained data at this point because we don't need it.

310

We simply need to tokenize it.

311

Where it does get custom is on the retriever step, remember, right?

312

Because we created the index on that.

313

So that's where our own custom vocabulary and index comes into play.

314

So we take our query tokens that we have

315

and then we run it against the retriever with k is 10 in this case

316

and then what you'll find is that we get the indices

317

and we get the scores and now what we do is we simply bring

318

that together in a list we loop over that

319

and then let me actually show you what that looks like

320

if i run this all together

321

so we can get a feel for what this looks like let's see over here let me actually clean

322

this up and run this one more time so it's the top now you can see that using our BM25 index,

323

if we tokenize where should I park my rainy day funds, these are the actual queries that we retrieve from the corpus.

324

So you'll find the scoring over here, then the index of the document itself and the actual text, five documents retrieved.

325

And then real quick, the fact that you're still watching this video signals to me that you're pretty serious about AI engineering.

326

But one of the big problems that we have right now is

327

that no one is really trained to be an AI engineer to work with LLMs because everything is so new, right?

328

And there are actually very little courses and resources online that really teach proper AI engineering end -to -end.

329

And I even ran into this issue in my own company

330

when I tried to hire AI engineers and figured out that actually nobody's properly trained for this.

331

So that's why I created the Gen AI Accelerator program.

332

This is a six -week track.

333

We've had engineers from booking

334

Intel, Electronic Arts, even from Meta going through this, and it will teach you in six weeks everything

335

that you really need to know to build production ready ai

336

applications all the way from foundation system design principles architecture rack of course llm monitoring

337

and evaluations and also very important deploying your ai applications to the cloud

338

and to actual production environments

339

and now we also have a 30 -day money back guarantee for this program

340

because we've been optimizing this for over a year already

341

and literally everybody loves it so

342

if you don't like it you can try it out literally just shoot me a message

343

and i will happily refund the entire amount for you.

344

And here's what makes this even better.

345

We have officially partnered with Credly to turn this into an actual certification program.

346

So because of this, you'll also find that a lot of students go through this program via their company.

347

So you can also reach out to your manager and see if this fits within your learning budget.

348

All right, so if that sounds interesting, make sure to check out the first link in the description.

349

And now let's get back to the tutorial.

350

So that is the first step of our hybrid retrieval system, BM25.

351

So sparse retrieval, keyword overlap with TF -IDF style waiting.

352

And it's excellent to catch exact terms, identifiers, and rare words, but it will miss paraphrases.

353

Now, let me scroll a little bit down over here to

354

pretty much explain why each piece earns its place in this retrieval pipeline and why they work so well together.

355

So for terms, like you can see over here, like the exact specific terms, BM25 is super good.

356

But it's often, it runs into problems where you have paraphrasing,

357

users asking questions in other terms than how they show up in the documents.

358

And that's where we can get into embeddings using dense embeddings.

359

And this is really why it's

360

so nice is dense embeddings really win exactly where BM25 loses

361

and loses exactly where bm25 wins they're kind of like two counterparts to it that work really well together.

362

So let's now look at how we're going to set up this.

363

And if you followed any previous

364

Rack tutorial, then you are already familiar with these concepts and it will probably be quite easy to understand.

365

So we're going to walk through the embeddings file over here, the three embeddings .py and we're going to use text embedding 3 small for this.

366

Now this requires an OpenAI API key.

367

You can use open source local text embedding models, but for a actual proper retrieval system, I recommend whenever possible,

368

just use one of the better models out there because it will simply improve the results of your system overall.

369

So we're going to follow a quite similar setup.

370

And we'll identify some directories in here.

371

And we'll create some simple functions that we can use to create the embeddings.

372

So you can see we're going to use the OpenAI client over here.

373

And we can use client .embeddings .create.

374

We'll plug in the model, which is the text embedding 3 small model, which will output an array of 1536 dimensions.

375

So that will be the length of the vector embedding in there.

376

And this is a simple batch function that we can use to loop over this.

377

Now, what we are going to do is we're going to read the corpus again.

378

So same structure, right?

379

We did this also with BM25, but now we're going to take a different approach.

380

So we take the IDs, we take that list

381

and now what I want you to be aware of is that

382

when you're going to run this since we're going to do it over the entire corpus

383

this is going to incur some costs

384

so it will be a couple of cents to kind of

385

like run through this we only have to do this once

386

and then it will be cached because we store it to disk

387

but I just want you to be aware of

388

that again for the sake of like following

389

and demonstration you could also swap this with an open source embedding model how to do that

390

Like you can literally simply open this in cursor cloud code and say, look, we're using OpenAI.

391

Can you implement an open source embedding model for this?

392

And it will set you up.

393

The process will be the same, but the output that you will, and the results that you will get in the end will be different.

394

So let's go step -by -step over here

395

and then zoom out a little bit to understand what's going on behind the scenes.

396

So we're again going to take all of the text in here.

397

We're simply going to split that up.

398

And this is where the magic happens.

399

So there's an if statement in here that says, look, if we already have the embeddings,

400

which we're going to store to dense embeddings .npy over here, if that file is already in place,

401

which you can see is the case on my system because I already run it, we will simply load it into memory by using np .load.

402

And if it's not, we'll go over it and we'll build the index, which you can see is the function we specified over here.

403

And this will loop over everything.

404

So it will take batches of size 256 and it will loop over the embedding model over here to create those vectors.

405

Now, one step to kind of like zoom out.

406

And just for your reference, again, if you've never worked with vector embeddings before,

407

you probably want to spend some time to kind of like go over this

408

and understand how factor embeddings work and how we are essentially capturing the semantic meaning in a numerical representation.

409

This is now again the point of this tutorial where I

410

cannot really go deeper into all of this to explain exactly how this works.

411

That's why I added the document over here.

412

What I want to show you right now is show you how this comes together and how we can use these embeddings.

413

Because if I now simply run everything over here and then clear all of this out,

414

you can see that we now have the embeddings over here.

415

So now we have the factors.

416

And we can even say, just to get a feel for what this is, we can take the first one.

417

So this is the factorized version of the first document in our corpus

418

and you can see the exact shape over here, right?

419

So this is how long the array is.

420

And this is because we're using the text embedding three small model.

421

Now, one little kind of like tip

422

that I want to give you here is that if you've worked with Rack in the past,

423

there is almost a 100 % certainty that you used a vector database to store your vectors, right?

424

So this, I'm assuming that for most people following this tutorial, you are familiar with this process of taking the corpus

425

and then turning them into embeddings and then storing them into a vector database.

426

What we are doing over here, what you can see over here, is instead of using a vector database,

427

we're simply using a NumPy array.

428

NumPy is a Python library that's, here you can see it's used for computations.

429

It's essentially the library that all of like AI

430

and deep learning is pretty much built around

431

because this is what actually stores the matrices and does the matrix multiplications that all like neural networks,

432

machine learning models, any types of models really are based on.

433

And we can leverage that really without needing a vector database.

434

We can simply just store these vectors to disk.

435

And that is what you saw over here.

436

So we have an embeddings .npy file over here.

437

And where we do that is the following.

438

So you can see over here, we simply do, after we create all of the embeddings.

439

So here you have all of the embeddings over here.

440

In all other RAG tutorials, you'll find that, okay, now we have the embeddings.

441

Let's put them into any type of vector database.

442

whether that's Quadrant, LensDB, VV8, BGFactor, let's put it into a vector database.

443

But for any corpus really under, I don't know, a million of chunks really in there, you don't really need it.

444

Similar to what I explained with the BM25, right?

445

We had like the 30 MB file.

446

Similar with the embeddings, we can also just run this in memory

447

and then use that without the extra complexity so that's the np .save

448

that we run over here we get the npy embeddings file

449

and let me show you just for the sake of clarity show you how big

450

that file is so you can see this is bigger than what we had with the bm25

451

so this is 350 megabytes on disk but again nothing too crazy too large that any deployment,

452

any compute, any VPS cannot handle, right?

453

And again, this is a pretty substantial corpus.

454

We got 56 ,000 documents in there.

455

So just a quick little tip to show you that you often don't really need a vector database.

456

So now let's continue to creating the search again, right?

457

So how do we use this?

458

Now that we have these embeddings, How do we use it?

459

So now here we have a function.

460

Let me clean this up for you.

461

A function to perform the search.

462

And what we're going to do in this case, we're again going to take the same query.

463

So we have the query, where should I park my rainy day funds?

464

We'll put that k is 10 into memory as well.

465

So we're going to retrieve the top 10 documents and then later show five over here.

466

And now the next step is to create that embedding.

467

So if I run this file over here, you can see that we now turn that where should I park my rainy day fund?

468

We create the embedding similar to what we did with BM25.

469

Remember, we did the tokenization.

470

So we did the prep work.

471

So it is then in a format that we can search against our entire corpus.

472

So we're doing that in the same way over here.

473

But remember here we're making an API call to open AI.

474

So we need

475

that model in order to get this computation with BM25 we could all do it locally because of the BM25S library.

476

Now here there is one step in there

477

which is for this particular embedding model that we're using it's not needed

478

but I wanted to show this in here anyway and

479

that is we perform a normalization and normalization means

480

that we rescale the factor so that the Euclidean distance becomes one in total.

481

Now here is that explained again in more detail

482

but why we include this in here is even though open ai

483

if we use the embedding models they are the vectors are already normalized

484

so we don't have to do it

485

but you can see we pre -normalize it over here during the embedding process on the corpus

486

and we also do it during the search on the actual query this is just a safety precaution

487

so that if you use any other embedding model

488

that doesn't automatically get you normalized factors we perform that step

489

and this is super important because

490

when we're working with normalized factors we can make the retrieval part easier

491

because again if you've worked with factors with rag before then you've probably heard about cosine similarity

492

which is how we actually calculate the similarity the semantic similarity really between two uh two factors within the space

493

that we created

494

so here's a 2d representation over this although in the in the real world

495

it is much more complex as you can see over here

496

but if we use normalized factors the calculation becomes simpler

497

because we can essentially just take the dot product it's explained over here why that works again

498

This is the mathematics behind it.

499

But if we have normalized vectors, we can simply take now our doc embeddings normalized, which is our entire corpus.

500

Remember, we can take our query vector, which we just created, and we can simply use this at notation over here,

501

which is NumPy's way of calculating the dot product.

502

Now, what we're essentially doing over here is we're performing a matrix multiplication.

503

So this is the entire matrix over here.

504

And we pretty much take every array in here and we multiply it with the query fact.

505

And then we get our scores.

506

So now we get all of the scores in here, which we can then sort to the top K.

507

So now you can see the following array over here.

508

And in the function, we will return those particular IDs.

509

So again, this is quite deep to kind of like get into, especially if you want to understand the math behind it.

510

But this is what's going on behind the scenes in most factor databases

511

when you offload your data to that and call the kind of like retrieve or get similarity function.

512

We have now implemented that in almost pure

513

Python really from scratch.

514

Literally just the only kind of like thing that we're leveraging over here is the NumPy library.

515

So if I now store this into memory and then just run the entire thing, you can see that now,

516

let me scroll up, we got back to the same as what we had with BM25, right?

517

We have a query.

518

We can search that against our corpus.

519

And now we, in this case, you can see K is five.

520

So it prints the five most relevant documents, but now through a different mechanism, through a dense search.

521

So you can see the scores, the IDs, any documents associated with it.

522

So that is step number two of our retrieval pipeline, right?

523

So remember to zoom out a little bit, really kind of like summarize

524

and get a clear understanding of where we are now at we have two different ways of creating these indexes

525

and then two different retrieval mechanisms so that we can get five documents

526

or ten documents over here

527

and over here using these exact methods over here where we now also know the strengths

528

and the weaknesses of each one exact keyword matches and the exact inverse, namely more paraphrased and actual meaning behind it.

529

So how do we now bring that together?

530

That is step number three.

531

And for that, we're going to use a technique called reciprocal rank fusion.

532

This is another algorithm.

533

And this is, I would say, the simplest out of everything that I shared.

534

If you really want to kind of like understand some of the math and the calculations behind it, this one is quite easy to follow.

535

Pretty much what we do is you can see over here where, again, what I just explained, we have two ways of getting rankings.

536

Five documents via this, five documents via that.

537

What the RRF essentially now does is it takes, it has this formula over here.

538

K equals 60 tends to be the kind of like norm that was also introduced.

539

It's the smoothing factor, which is almost also introduced in the original paper, which kind of like still works and everyone defaults to.

540

So you can now see if for this particular document,

541

so the green one, BM25 ranks it at number one and dense ranks it at number two.

542

You can see the formula.

543

We do one, we divide it by 60, which is the smoothing factor K, plus and then the rank.

544

So this is for BM25, it's ranked number one.

545

And here for dense, it's number two.

546

So this now becomes the calculation and this becomes the score.

547

You see, it's a little bit small, but that is how it's calculated.

548

Now, if you look at the purple dog over here, you can see rank number four, rank number one, rank number four,

549

rank number one, do the calculation.

550

And then you pretty much sum everything.

551

So this is a way to simply bring together different retrieval algorithms where the original scores cannot be directly compared

552

because it wouldn't make sense to compare the score, the similarity score, like the cosine similarity of dense ranking,

553

and then rank it against BM25, because they're on different scales that would be comparing apples to oranges.

554

So this is a way to say, look, we just care about the rank.

555

And that's how we bring them together.

556

So that's a short little introduction.

557

Again, more information, original papers can be traced back over here.

558

So let's see what this looks like in the code.

559

If we go to RFF, file number four, you can see that same smoothing factor over here, which we start with.

560

And here is the pure

561

Python implementation of what we do.

562

So we pretty much loop over all of everything in our list over here.

563

And then this is where we do the calculation.

564

So here you can see that exact kind of like one divided by K plus rank, right? similar to

565

what we did over here.

566

That is the pure Python implementation.

567

So let's store that in memory.

568

And now we can actually move quite quickly over here because right now you know how the BM25 works,

569

how Dense works, and we have the corpus.

570

So now it's all about the ranker.

571

And what I did in this step, so you can see in the folder, you can see a utils in here,

572

and you can see that we are performing some imports.

573

So from utils retrievers, you can see that I created a let's see the BM25 retriever

574

and the dense retriever so it's a simple I created two simple classes

575

that we can import

576

so we don't have to kind of like redo the code in all of the files

577

so this is also like if you want to take this into your own project

578

import them from the utils simply take these

579

and put them in in whatever you want to use

580

so okay coming back to where were we the rff over here yeah okay

581

so now we just simply load it we have everything

582

and then we define a new function over here to perform the hybrid search

583

because this is not we're now bringing them together right

584

so you can see we perform a bm25 .search

585

and a dense .search same query same candidates that we're retrieving

586

and now what we want to do is

587

because we're bringing things together

588

we want to bump up k a little bit to say look

589

you can just like just get a lot of documents

590

you can take top 50 like just show me what you got

591

and then we bring together and then we can take a subset from there

592

and simply rank them using our reciprocal rank fusion function remember that we create it over here to perform the calculation.

593

So let's see what that looks like.

594

So if I now take my query, where should I park my rainy day funds?

595

And I'm actually going to walk you through it over here.

596

Let me see.

597

Let's do it like this.

598

I should do that one by one.

599

So let's first do, we get K, K is 10.

600

This is really why I love the interactive window, because this is ultimately the best way to learn,

601

because I can run this all at once and immediately show you the results.

602

But if I run this over here, you can see, let's first focus on the BM25 search.

603

So here's the BM25 search and here are the IDs.

604

Remember, K is 50, so we get 50 documents back.

605

Then let's do the dense search.

606

So here we get the dense IDs.

607

Also, 50 documents that we get back.

608

Now, here we're going to perform the rank fusion by putting the two ID lists in here.

609

So you can see we put that into an actual list.

610

So we just stack them together.

611

So you can see here's the first 50 and then here starts the second list.

612

So that is the input.

613

And then if you go over here, you find that this takes a list with a list of strings,

614

which is exactly the format that we have.

615

And that's how it performs the calculations so

616

that is like walking through it step by step if we then run this and show everything

617

now we can see everything coming together

618

so let me show you what this looks like

619

so here we can see that for our original query that we've been using over and over again

620

here's what bm25 will give you back as top five here is what dense will do

621

and then here is the hybrid with RFF.

622

So here we can already start to see these differences, right?

623

And just so you know, right now we have no way really yet of telling or seeing what is correct and what is incorrect.

624

We can only do that once we get to the evaluation

625

and we map the original query back to the original documents, right?

626

I think we checked this in the beginning, but I don't remember which IDs actually belong to this query.

627

But you can see, for example, that this one really stands out over here you have

628

that duff deer over here which is not in bm25

629

and then you can see with rff this is moved a little bit to the top

630

and you can see uh starting with in addition this one this is now actually ranked number one

631

and i could see why because you could see it is number two for bm25

632

and also number two for um for dense

633

so both have a very high rank in uh in both results

634

and that's why it's kind of like pushed up

635

and then the bonds you can see it ranks over here bm25

636

but it's not in the dense so that's why that moved down

637

so this is how this is now all coming together

638

and we get a fused ranking of the two all right so

639

that was step number three of how we can bring the results together

640

now there's going to be one more final step and

641

that is we're going to add the re -ranker

642

so what we're going to do over here is

643

if you look at the process of like the dense embeddings that is what you would call a by encoder.

644

So both the query and the corpus, they go through the vectorization process in separate paths.

645

When using a cross encoder, we can do that at the same time.

646

So the query can go through the encoder together with all of the documents.

647

And let me actually show you over here again what that looks like without going too deep into everything.

648

What a re -ranker essentially does is it looks at all of the results

649

that are retrieved and now also already combined using RFF and then it's going to reorder them based on actually looking at,

650

okay, but now wait, what is the question and what are the actual documents that I have?

651

And going through this.

652

And there are various models that we can use for this.

653

You can use open source models,

654

but I'm going to use one of the best re -ranked models out there and that is a model from Cohere.

655

You can set up an account on Cohere

656

and you can use the free API key because they have a free plan available to follow along.

657

This is what I would recommend, but you could also swap it out for an open source one.

658

It is explained over here what that would look like.

659

So we're going to use Rerank V4 fast.

660

So let's show what that looks like.

661

And this is actually now also a very simple setup because we've done most of the work already.

662

So let's go into the Rerank file.

663

You can see that right now we set the model and we use the cohere client, we get all of our retrievers.

664

And then the only thing that we do is beyond the hybrid candidates.

665

So this line over here is exactly what we did with the previous step, right?

666

To rank them using RFF.

667

Now, all we do is we take that same list

668

and we add a re -ranker on top of it and then we showcase everything.

669

So let's walk through this step -by -step again.

670

So you can get a feel for what this looks like.

671

We'll load everything.

672

Now, let's see.

673

Let's go into the search here and then the show again.

674

Let's take our beautiful query again.

675

Okay, so let's do the same trick, right?

676

We say case 10, candidates is 50.

677

So now, if we do this, these are our 50 candidates.

678

And remember, here RFF is already applied.

679

So what we did, coming back, we should come back to So we did BM25 search, then search, then reciprocal rank fusion,

680

which is RRF, by the way.

681

I just noticed that for the past couple of examples, I said RFF, I think.

682

Sorry about that.

683

So now with this one little line over here, we perform all of these.

684

And now we add the re -ranking on top of that.

685

So here is how that works.

686

You can now see all of the candidate IDs that we have.

687

So we have all of the IDs and then we have the texts.

688

So we get back to, okay, what are all the actual documents that were retrieved?

689

And now this is where the magic happens.

690

We do a co -air, we do the re -rank, we put in the model and now we put in a query.

691

Query, same, Canada text over here.

692

Top N, we set that to 10 in this case.

693

So let's see what we get out of that.

694

So this is another API call that we make

695

and now you can see we get a re -ranked version of it 10 results

696

and also some scores in here and we let

697

that return to we let the function return all of that

698

so now that we went through step by step let me run everything

699

so we can get again to let's see to a top five

700

so now you can see

701

if we do the rrf only the hybrid these are the documents

702

that get retrieved and now if we do hybrid plus the re -ranking you can see

703

that it's different so this is now on number one you can see

704

that duff there again which through rff was on number two

705

but now it's on number five over here

706

so you can see all of these differences are very subtle

707

and we are constantly looking at the top five

708

but what you should always keep in mind is if you scale this to many questions many documents in production environments, also with, I don't know,

709

top 10, top 20, this is actually going to make a difference

710

because the difference between a rack system like kind of like not working

711

or sometimes working is all about which documents show up when you ask which question.

712

And these subtle nuances, they make all the difference and that's how you're going to optimize your system.

713

So that's, it may kind of like feel very repetitive and it's like, yeah, it all kind of like got,

714

get we all kind of like get the same results right but we don't and that's why now we can actually

715

go to the last file evaluate so

716

if you made it up until this point great job we're almost there

717

because now we're going to perform the evaluation to see

718

if what we're doing actually makes sense

719

so coming back to let's see our documents over here let me come to the ndcg

720

so normalized discounted cumulative gain

721

so this is in information retrieval one of the golden standards to measure retrieval quality again

722

there is some math behind it to better understand this

723

but we again look at the rank of the retrieved documents

724

and then whether they're relevant or not and then we compute a score between zero and one.

725

And this is how we can compare different iterations.

726

So what you should keep in mind there is if we run this, higher means better score.

727

We change something, we try a different embedding model, we try a different chunking strategy, we try a different re -ranker, we try a different step,

728

higher score is better, lower score, worse.

729

That's how we're going to optimize this so how i have implemented

730

that for you again you can see the implementation in pure python over here

731

i'm not going to going to go through everything

732

but this is how we can calculate it and

733

if i now scroll down and pretty much well let's go let's go for this step by step

734

because we have to function and then we get we load all of the queries first

735

so let's get our entire corpus ready.

736

So now we have everything ready.

737

Now we have these, the retrievers.

738

So here's where we're going to wire up the retrieval mechanisms, right?

739

And let's see, we have some other helper functions in here.

740

And this is where, let's see, we're actually going to run everything.

741

And I'm simply just going to do this.

742

And then we're going to reverse engineer, because as you can see what we're now going to do is for each query

743

that we have within our data set

744

and by the way we limit it to let's see we limit it to 50 in in this case

745

so I'm going to take a random sample of 50 queries

746

and then we're going to run into that

747

so I'm not going to do the full full corpus size

748

so you can play with this and pretty much

749

if you look at where we're setting

750

that up you can see over here we're taking a random sample

751

so here this is this is 50

752

and we set a random seed of 42 meaning every time

753

you run this the exact slice of the data is exactly the same okay

754

so now this is completed

755

and why i do 50 in this case just to kind of like speed things up

756

and this will also put you within the uh within the free plan of of cohere

757

that you can still use it I believe without running into API errors

758

so okay but now we have some numbers over here

759

so let's look at them and check what's actually the differences

760

because here you can really see the impact of what we're building

761

so you can see that for our score

762

and remember a score between zero

763

and one is what the ngcg at 10 will give you

764

but we multiplied it by 10 or sorry by 100 to kind of like get the percentage in there.

765

This is also usually how it's reported within all of the research papers.

766

So if you see, if you look at, if we only do BM25, we get a score of 28, which is not that great.

767

And that is because this particular data set

768

that we're using is known to not be very effective with direct keyword -based matching.

769

Then if you look at dense, you can see that clearly performs better.

770

So the vector embeddings, that is already a lot better at retrieving the right documents.

771

Now, you can see if we then combine that using RRF, you can see that we are somewhat in the middle.

772

So between these two.

773

But now here is where it gets interesting.

774

If we now take the hybrid and we rank it, all of a sudden we jump up to 47.

775

That is a major increase over dense alone really while just literally adding one simple re -rank step.

776

Because remember also from your code base over here, this is literally all that we add to it.

777

That is what we add to it.

778

And because of that, we jump to this particular score.

779

So that is how you can evaluate your methods and your approaches.

780

And now one, the most important thing to keep in mind is do you always need this entire stack?

781

And the answer is no. So you should figure out for your corpus what the best combination of this is.

782

And I've now shown you an experimental setup that you can use to do this.

783

Because in this case, it's actually interesting.

784

It could be that maybe if we take BM25 out of it

785

and we just do dense and we maybe increase the retrieval or the top K, we increase that in here and then apply the ranker.

786

We may even get higher.

787

We may even get to 50.

788

So this is now the.. experimental part that you need to get into.

789

And that's, I think, one of the main benefits of like going through this entire video

790

and really trying to understand it rather than just like copying samples

791

that you find online or letting cloud code one shot it, is understanding the mechanisms behind it.

792

And what's, let's see, what I want to show right now is finally, now that we're coming to the end of this project,

793

is how you can apply this to your own project.

794

And the setup is like from a code perspective, it's almost exactly the same.

795

So you decide what kind of documents do you want to use, right?

796

Are they Notion, Zendex, whatever, take your data.

797

Check if you need to chunk.

798

So you like, you ideally want to chunk.

799

First of all, you have a hard limit of what you can actually send into the embedding model.

800

But even like in most like best practices, you want to everything over like 1000 tokens.

801

see if you can already split it up, for example.

802

So this is a step we did not have to do in this project

803

because the corpus was already like there was small isolated documents, right?

804

Okay, then you want to pick an embedding model.

805

Then you're going to build the indexes, wire everything up, and then you're going to build the evaluation data set.

806

And this is the most important gap that you'll probably encounter within your own project.

807

And that's why I created an entire guide for this as well because we use the bear style evaluation set, remember, right?

808

And that's why we have this, Curries, Relationships, and Corpus.

809

And that's why we can run this.

810

That's why we can say anything meaningful about the adjustments that we make.

811

In the real world, you don't have this.

812

So then the question is, okay, but how do I improve my system beyond just kind of like vibes

813

and just like running

814

it? how pretty much like 90 probably even more people like working with ai are doing

815

it just chuck everything into the vector database

816

and like just just let it run

817

and then just like check off of vibes this is how you get like scientific about this

818

so how do you create

819

that well let's see coming back to the guides

820

this is the one i created the guide over here on how to do

821

that but the thing is it's it's it's pretty simple you already have the documents

822

that should be your starting point

823

and then you're going to use an llm with a prompt

824

like this you are a user with a question the following

825

document contains the answer contains the answer generate a single realistic question

826

that a user would ask in their own word

827

so rules question must be answered by the document

828

and you can loop over this

829

so you can even use like a small cheap model for this

830

and And now you can exactly structure your own data set with document IDs

831

and questions and relationships in the format that we used in this tutorial.

832

And in that way, you can start to improve it over time, add more to it.

833

And this is just super simple nowadays with LLMs.

834

So that's how you build your own eFiles really, and how you really put this into practice beyond just the tutorial.

835

All right, and with that, we've come to the end of this video.

836

Now, as always, if you find it valuable, please leave a like and also consider subscribing.

837

And if you want to go really deep on AI engineering and you love tutorials and lectures like this,

838

actually covering what works in production and not just like most of the AI crap that's out there online,

839

then make sure to check out the Gen AI Accelerator.

840

It's a six -week track where I pretty much transition

841

you from being a general developer into an actual AI engineer by showing you exactly how we build

842

and deliver all of the client projects for DataLumina, our AI development company, where we've been working on AI projects for close to a decade already.

843

So if that sounds interesting, make sure to check it out.

844

And then if you want to continue learning here on YouTube, I'm going to link this video up here, which is another approach to information retrieval,

845

but then using an agentic approach.

846

So how to go about a Gentec rack is what I explained in this video.

847

So I also highly recommend to check out that.

Descargar app

Todo lo que necesitas para hablar con fluidez

Pronunciación con IAEvalúa cada oración

Práctica de IPADomina cada sonido

VocabularioConstruye tu banco de palabras

Juego de vocabularioAprende mientras juegas

Escanea el QR

Download on theApp Store Get it onGoogle Play

FOR YOU

Recomendar Video

712 oraciones

Redis Crash Course - the What, Why and How to use Redis as your primary database

Daily ConversationUniversity StudentsWorking Adults

111 oraciones

Executing Monitoring, Evaluation and Learning (MEAL) | Complete Practical Guide

Daily ConversationUniversity StudentsWorking Adults

90 oraciones

Did Google just dethrone OpenAI and Anthropic?

Daily ConversationUniversity StudentsWorking Adults

30 oraciones

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

Daily ConversationUniversity Students

125 oraciones

Introduction to statistical prediction models

Daily ConversationUniversity Students

412 oraciones

1. Roadmap for backend from first principles

Daily ConversationUniversity Students

68 oraciones

what is a database index?

Daily ConversationUniversity StudentsWorking Adults

31 oraciones

Usability Heuristic 2: Match Between the System and the Real World

Daily ConversationUniversity StudentsWorking Adults

137 oraciones

JUST $1 For Billions of Tokens?! (DeepSeek v4 Pro)

Daily ConversationUniversity StudentsWorking Adults

Vocabulario y notas de pronunciación para esta lección

Esta lección de conversación se basa en el vídeo “The Complete Guide to Hybrid Search in RAG”. Este vídeo tiene 847 frases y 10557 palabras para practicar shadowing. La parte hablada dura 59:17. El hablante habla a un ritmo natural de unas 178 palabras por minuto, cercano a una conversación cotidiana. El 87 % de las palabras está entre las 3.000 más comunes del inglés, por lo que el vocabulario es fácil de seguir.

Vocabulario clave de este vídeo

15 palabras del vídeo que vale la pena aprender, con su pronunciación y significado:

Palabra	Pronunciación	Significado
query sustantivo	/ˈkwɪɹ.i/	consulta, pregunta
corpus sustantivo	/ˈkɔɹpəs/	corpus
rank sustantivo	/ˈɹæŋk/	fila
dense adjetivo	/ˈdɛns/	denso
vector sustantivo	/ˈvɛktɚ/	vector
retrieve verbo	/ɹɪˈtɹiːv/	recobrar
database sustantivo	/ˈdeɪtəˌbeɪs/	base de datos, banco de datos
tutorial sustantivo	/ˌtjuːˈtɔːɹɪəl/	tutoría, tutorial
hybrid sustantivo	/ˈhaɪ.bɹɪd/	híbrido
token sustantivo	/ˈtoʊkən/	seña
calculation sustantivo	/ˌkælkjuˈleɪʃn̩/	cálculo, calculación
mechanism sustantivo	/ˈmɛk.ə.nɪ.zəm/	mecanismo
rainy adjetivo	/ˈɹeɪni/	lluvioso, pluvioso
essentially adverbio	/ɪˈsɛnʃəli/	esencialmente, en esencia
mathematics sustantivo	/mæθ(.ə)ˈmæt.ɪks/	matemáticas, matemática

Phrasal verbs que vas a escuchar

Palabra	Pronunciación	Significado
go through verbo		atravesar
walk through verbo		explicar paso a paso, enseñar
come together verbo	/ˈkʌm tʊˈɡɛðɚ/	juntarse, concitarse
check out verbo		investigar, chequear
get back to verbo		responder / contestar (en un rato), volver / tornar a contactar
go over verbo		sentar, pasar
zoom out verbo	/ˌzuːm ˈaʊt/	alejar el zoom, quitar zoom (el)
bring together verbo		concitar

Frases que vale la pena repetir

Frases cortas y completas del vídeo que puedes reutilizar en la conversación diaria:

All right, so let's kick things off.
Remember how I was talking about this over here?
What data are we actually working with?
Let's pick some random ones.
Can you implement an open source embedding model for this?

Gramática en este vídeo

Las estructuras que más usa el hablante, con las palabras exactas del vídeo:

Estructura	En el vídeo
Present perfect continuous have/has been + -ing — una acción que empezó antes y continúa	we've been optimizing · we've been using · we've been working
Oraciones condicionales if + oración, will/would + verbo — una condición y su resultado	if it's not, we'll go
Present perfect have/has + participio pasado — una acción pasada que sigue importando ahora	they've dabbled · you've never worked · we've converted
Voz pasiva be + participio pasado — importa lo que ocurre, no quién lo hace	are linked · can be answered · is considered

Pronunciación a tener en cuenta

El hablante usa 151 contracciones y formas reducidas, como we're, I'm, don't. Dilas en su forma corta, tal como las oyes.

Los sonidos de “th”: mathematics /mæθ(.ə)ˈmæt.ɪks/, smooth /smuːð/
Los sonidos de “sh” y “zh”: calculation /ˌkælkjuˈleɪʃn̩/, essentially /ɪˈsɛnʃəli/, implementation /ˌɪmplɪmənˈteɪʃən/, fusion /ˈfjuː.ʒən/, evaluation /ɪˌvæljuˈeɪʃn̩/
Palabras largas — cuida el acento: mechanism /ˈmɛk.ə.nɪ.zəm/, essentially /ɪˈsɛnʃəli/, mathematics /mæθ(.ə)ˈmæt.ɪks/, implementation /ˌɪmplɪmənˈteɪʃən/, vocabulary /vəˈkæb.jə.lə.ɹi/

Sonidos difíciles para hispanohablantes:

s + consonante al inicio — sin añadir una “e” delante: smooth /smuːð/, scratch /skɹæt͡ʃ/, stack /stæk/, scroll /skɹoʊl/, swap /ˈswɑp/
/v/ — no es /b/: los dientes tocan el labio inferior: vector /ˈvɛktɚ/, retrieve /ɹɪˈtɹiːv/, vocabulary /vəˈkæb.jə.lə.ɹi/, dive /daɪv/, evaluation /ɪˌvæljuˈeɪʃn̩/
/z/ — sonora, no /s/: mechanism /ˈmɛk.ə.nɪ.zəm/, representation /ˌɹɛp.ɹə.zɛnˈteɪ.ʃən/, zoom /zuːm/, nowadays /ˈnaʊ.ə.deɪz/, transition /tɹænˈzɪʃ.ən/

Cómo practicar con este vídeo

Escucha el vídeo entero una vez sin hablar y anota las palabras que no conoces.
Empieza a velocidad 0,75×, haz shadowing frase por frase y vuelve a la velocidad normal cuando te resulte fácil.
Grábate y compara con el original, prestando atención a palabras como query, corpus, rank.

¿Qué es la Técnica de Shadowing?

Shadowing es una técnica de aprendizaje de idiomas respaldada por la ciencia, desarrollada originalmente para la formación de intérpretes profesionales y popularizada por el políglota Dr. Alexander Arguelles. El método es simple pero poderoso: escuchas audio en inglés nativo y lo repites en voz alta de inmediato, como una sombra que sigue al hablante con solo 1-2 segundos de retraso. A diferencia de la escucha pasiva o los ejercicios de gramática, el shadowing obliga a tu cerebro y músculos de la boca a procesar y reproducir simultáneamente patrones de habla reales. Las investigaciones muestran que mejora significativamente la precisión de la pronunciación, la entonación, el ritmo, el habla conectada, la comprensión auditiva y la fluidez al hablar, convirtiéndola en una de las metodologías más efectivas para la preparación del IELTS Speaking y la comunicación en inglés en el mundo real.

Técnica de shadowing: lee la guía completa paso a paso →

Shadowing Practice: The Complete Guide to Hybrid Search in RAG (BM25 + Embeddings + Reranker) - Learn English Speaking with Video

Elige un modo de práctica

Descargar app

Master AI in 60 Minutes with Vector Databases!

Analyze SEC filings with Transformers for fun and profit

LTEC 782: Design-Based Research in Education (Session 01)

Redis Crash Course - the What, Why and How to use Redis as your primary database

Executing Monitoring, Evaluation and Learning (MEAL) | Complete Practical Guide

Did Google just dethrone OpenAI and Anthropic?

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

Introduction to statistical prediction models

1. Roadmap for backend from first principles

what is a database index?

Usability Heuristic 2: Match Between the System and the Real World

JUST $1 For Billions of Tokens?! (DeepSeek v4 Pro)

Vocabulario y notas de pronunciación para esta lección

Vocabulario clave de este vídeo

Phrasal verbs que vas a escuchar

Frases que vale la pena repetir

Gramática en este vídeo

Pronunciación a tener en cuenta

Cómo practicar con este vídeo

¿Qué es la Técnica de Shadowing?