The evolution and promise of RAG architecture with Tengyu Ma from Voyage AI

0:07TENGU MA, Welcome to KnowPriors. with just a little bit of an overview of like your research agenda to date, because I think uniquely it covers a broad range of fields within and around deep learning from like theory to RL to embeddings and optimizers. So can you talk a little bit about sort of how you pick the directions you have? Yeah. So I think most of the papers I wrote have some theoretical thinking in it. I guess maybe that's the commonality. And besides that, I think I worked on quite a few topics, as you mentioned, ranging from the theoretical understanding, mathematical proofs of deep learning systems, all the way to practical large language models, reinforcement learning, deep reinforcement learning.

1:11And these days, recently, I think what we are working on more centralized to efficiency of training the large language models and improving the reasoning tasks for large language models. So my vision is that in the future, the efficiency is very important because we are running out of data and compute. So we have to either use the data much better and use the compute much better. And also reasoning tasks seems to be a pretty important direction. And also in some sense, kind of like a risky direction in the sense that we don't know exactly how fast we can solve those challenging reasoning questions yet.

1:53Can you mention a few of the key papers or work that you or students in your lab have done, just so our listeners can look them up? In the very early days, I think I worked on some of this matrix completion, optimization for matrix completion. That's like 10 years ago. And then I move on to embedding models, like sentence embeddings, vector embeddings. One of the papers we wrote is a very actually simple paper where we average the word embeddings to get sentence embeddings. And then we did some of these transformations using PCA to make the performance much better. That was even before Transformer came out.

2:32And then I think I move on to Transformers, large language models and contrastive learning, which is the new way of training the embedding models. especially the direction started with some of the papers on using contrastive learning for images. And we work on improving those and understanding why contrastive learning can work. And recently, we work on optimizers for large language models. For example, one of the paper we wrote last year was Sophia, where we found that we have a new trend optimizer, which can improve the training efficiency by 2x for pre-training. This is great. Atom is very old at this point.

3:12Yeah, he's 10 years old now. I think that's the interesting part about it. So optimizers, you know, I think people have tried in the last 10 years for so many times. There were so many papers published which has improvements over Atom in various cases. But so far, Atom is still the default algorithms for training large English models. And that's why we thought that it's the time to really, we spent a lot of time on this. Like I think I started probably around 2018, 2019. And I asked a few students to work on this. And finally we had one paper out after a few years, after a few failed projects and failed ideas.

3:54And recently I think one of the Facebook friends actually used this in their large scale multimodal training. And they found that on that scale, I don't know exactly how many parameters there are, but I assume it's kind of more than 100 billion parameters. They found that on that scale, there is a 1.6x improvement in the efficiency of the training. So that's like$10 million versus$16 million. That's super exciting. Yeah, I think, you know, Sophia has an opportunity to be really, really impactful. You started a company last year taking leave from Stanford. given your work has been theoretical, but with practical applications, what drove you to do that?

4:41I think I came to Stanford partly because there's a very strong industry connection here at Stanford compared to some of the other universities. And also probably entrepreneurship is just part of my career plan anyways. And in terms of the timing, I felt that this is the right timing in the sense that the technologies are more and more mature so that it seems that the commercialization is the right timing right now. So, for example, I think one story I have is that, you know, I look up some of my slide stack for my lectures at Stanford CS29 seven years ago when I started to teach at Stanford. At that point, machine learning, we have a lecture with Chris Ray and the machine learning on applied machine learning.

5:32So how do you apply machine learning industry? And there are seven steps there. So the first step is you define your problem. The second step is you collect your data and you choose the loss function, you train it and you iterate so and so forth. So it's pretty complicated at that point. And now the foundation model arrives to power and in a new foundation model era, the only thing you have to do is that you have to, you know, someone will train a foundation model for you and then you turn a prompt and you add a retrieve augment generation on top of it. And that's pretty much, that's it. So applying machine learning AI to an industry environment is much, much easier than seven years ago.

6:15And that's why I felt that this is probably the right time to commercialize many of the technologies because the technologies are more mature. Yeah, this is actually, I mean, a core premise even for the investing fund that I started a conviction in that, But, you know, somebody is doing the bulk of the work for you in a more general way. And so the application of AI in industry is just much, much cheaper, right? Because you only do, you know, the last few steps or a different set, but last few steps in essence. So maybe you can talk about like the, you know, just given wide range of research, the problem you focus on with Voyage that you saw with customers.

6:51Yeah. So with Voyage, I think we are mostly building these two components. re-rank and embeddings for improving the quality of the retrieval or the search system. So the reason why we focus on this is because we talked to so many customers and we found that right now for implementing Rack, the bottleneck seems to be that, you know, it's not very hard to implement it, right? You can just connect the components and have your Rack system ready very quickly. But the bottleneck seems to be the quality of the response. And the quality of the response is heavily affected or is kind of almost almost bottlenecked by the quality of the retrieval part.

7:31If the large language model see very relevant documents, then they can synthesize very good answers. Even like LAMA 70b can do that very well. Can you just give like a sort of general intuition for like what a RAG system is and some of the applications of it? Yeah. So I guess just a little bit of background. So a retrieval augmentation generation, the idea is that there's retrieval step, there's a generation step. So the main point here is that if you just use a large language model as a black box, you know, as is, then the large language model wouldn't know anything about the prepared information inside the company.

8:04And it doesn't know enough context about the use cases. And the retrieval generation stack is about you first retrieve some knowledge from, for example, inside a company, and then use this knowledge and give the knowledge to the Latin English model so that the Latin English model can generate or synthesize a good answer without any hallucination. This has been found to be very useful to reduce the hallucination rate. And so there's two steps. The first step is to retrieve some relevant information given the query and then this relevant information are given to the Latin English model. The retrieval step is important because once the Latin English model sees the relevant information, it can reduce the hallucination rate dramatically because it used the relevant information as an anchor to refine the answers in some sense.

8:55And what we are doing here is that we want to improve the quality of the retrieval or the relevancy or accuracy of the retrieved documents and information. And the way to do this is that there are two steps. The first step is that you vectorize all of your documents or all of your knowledge base. So you turn the documents to vectors, you turn the videos into vectors. You turn your code into vectors. Code into vectors, everything into vectors. And so the vectors are the representations of each piece of the knowledge or documents and all the other indices. So and then you put these vectors into a vector database and then you search the right of information using the vectors as indices.

9:36Where are you seeing RAG applications today? Like what are what are customers building or, you know, what are the most common systems? Yeah, so we have a lot of users and they are all over the places. I think, you know, we have even a customer who is a chemistry company who is building this RAC system to understand their chemistry documents or products in descriptions. And I think just it's almost everywhere, like finance, legal, code retrieval, code generation, so on and so forth. I think it can be applied to almost any cases. And also for, even for individual users where you have a lot of like individual personal information and you want to have a RAC system on a phone so that you can access your past information in a much more easy way.

10:28And you want to retrieve, for example, we all have seen that, you know, when you search your documents on your laptop, it's actually pretty hard. You have to use the exact file name. It will be much easier if this search can be semantic based. Mm-hmm. RAG is a relatively new architecture. I think your average enterprise technology leader had not heard the term before the last year or so, and it became popularized in a researcher for the last few years. But there is already a debate, I think, you know, in terms of opinions from people at different large labs and in academia about whether or not you need a RAG architecture to work on proprietary data.

11:12And just to sort of describe some of the alternative views, I think there's kind of two alternative points of view given. One is a sort of agent chaining architecture where you are inputting your data and knowledge, you know, chemistry, code, law, finance, whatever documents into a series of LLMs that just operate with instruction on it, for example, to summarize or categorize it. Or you simply feed everything into LLMs with infinite context or actively managed context versus explicitly vectorizing anything. And so I would love to get your reaction to, you know, that as an alternative to RAG. Actually, there was also a debate last year about RAG versus fine tuning.

12:01And I think that debate was kind of like getting to a consensus now. It sounds like RAG is much easier than fine tuning. And fine tuning in many cases doesn't work because you need a lot of data to see the results. And there are still hallucinations even after fine tuning. And now, as you said, the debate becomes right versus agent changing or long context. So maybe let's talk about long context first. So I think there are probably two answers to this from different angles, because long context right now is not practical yet. So we have to kind of anticipate what long context transformer can do and then do the debate at a future time in some sense, or anticipate the debate at a future time.

12:49In the near term, I think the long context transformer, where you just put in all the proprietary data, 1 billion tokens into the context of the transformer, will be very, very expensive. If you use the price right now, it's going to be just impossible to do it. It's probably like 5, 10 magnitudes of difference, depending on how many documents you have in the context. Of course, you can bring the cost down by, for example, one approach is you cache the activations of all of the internal operations of the documents you put in the context. So that will bring the cost down by a lot. But I think still, if you do the calculation, theoretically, it's still much more expensive than RAC.

13:33So I think that's the more practical answer. So in terms of cost, it's going to be much more expensive than REC because you have to save all of these activations or intermediate computations in the GPU memory, most likely, or maybe in CPU memory of all the 1 billion tokens contacts. You know, you may argue that, okay, over time, everything will become cheaper and cheaper, but RAG will be cheaper as well, right? Because many of the technologies under RAG are neural network based and the GPUs will become cheaper, the neural networks will all become smaller. So my prediction is that RAG will be much cheaper than long contacts going forward.

14:16And another way to think about this is that maybe just from the first principle, right? So my analogy of log context is that, so in some sense, the context is the short-term memory in some sense, right? And the reg is more like long-term memory in some sense. So the question is, for example, when you answer any question, why you have to go through the entire library every time, like put all of the entire library in your short-term memory for answering a single question, right? It sounds like the right approach should be that for every single question, you retrieve some subset of the information and use those to answer the question.

14:53That seems to be the most efficient way to do that. It should be some kind of hierarchies in some sense in terms of how we solve the problem so that we can get the best efficiency. Even when we do the computer architecture, like the hardware stuff, right? So you have a different level of caching, right? So you have disk, you have CPU caching and so forth. So in that sense, I feel like the more hierarchical two-level kind of like system like RAC is more cost efficient. Yeah. I mean, the analogy certainly makes sense. I think there is another thread of discussion of like, what does long-term memory for LLMs look like where, you know, it is something managed by the LLM itself.

15:34But I do not think that is a well-answered question. And like RAG may just be a part of that answer. So the embedding model that we rank are in some sense, the large language model that are managing the long-term memory. Of course, there might be variants and other ways to manage the long-term memory, but I think it will be somewhat similar. It's going to be like more, you know, the technology always evolves, right, gradually, right? So maybe two years later, Voyage or maybe other companies will have a new version of the long-term memory, which is based on, you know, embedding models, but, you know, kind of like extending the embedding model in some way.

16:10That's entirely possible. Yeah, I do think it's useful to sort of contextualize for people who are not working with sort of data sources for LLMs at scale every day, like what sort of token limitations are. Right. You know, we go from a few thousand tokens to something like Gemini 1.5 Pro, a code context window of a million tokens. Right. And so if that's short of that in word count, that's maybe five books or like 25, 30 ,000 lines of code and obviously like limited amount of video and audio. And so I think the ability to make reasoning decisions on more than that amount of data is obviously going to be needed.

16:56And so the questions to me are really like, you know, does efficiency matter both from a cost perspective and a speed, like a latency perspective, right? And how much can you push the context window? And like, you know, does hallucination management matter? And so I think there are lots of arguments for like RAG being very persistent here. Yeah, yeah, exactly. Okay, and just to add a little bit on that. So 1 million tokens, 5 books, right? So by many companies has 100 million tokens. That's 100X difference, right? So 100X, you know, for cost is a big difference. That could be just, you know, 100K dollars versus like$10 million.

17:34$10 million is unacceptable, but 100K sounds okay. Yeah, I think that's probably what's going to happen. At least for many of the companies, right? So right now, if they have 100 million tokens, I don't think they can use long contact transformers at all. because it's way too expensive. Right. And the simplest thing for me is actually for a system to look at the entire code base or some representation of the entire code base versus the portion of it that could fit into context today. Yeah. What about the other piece, like the idea of agent chaining and using LLMs to manage the data in that form?

18:09So agent chaining, this is a growing area and many people are doing research on it. I think it's a little bit less well-defined in some sense. The first level bit I would say is that I think it's kind of orthogonal to embedding models and re-rankers to some degree because even when you have agent chaining, you still probably use embedding models as part of the chain. You probably do iterative retrieval as part of the chain and of course you use large language models as part of the chain as well. In some sense it's an orthogonal direction. So I probably would rephrase the agent training as more like an iterative, multi-steps, retrieval, augmented, large language model, augmented system.

18:53So, and some part of this retrieval probably is done by a large language model. Sometimes part of the system is done by a small large language model. And some part of the system is done by embedding model, so on and so forth. So in that sense, I feel like it's somewhat kind of orthogonal. Yeah. And I feel like some of of the motivation for agent chaining to begin with is the same efficiency motivation as RAG. Yeah, exactly. Right, but if you use a very, very large language model to manage the system, the knowledge system, I think you again lose the efficiency, right? So it has to be a somewhat kind of like smaller model to manage the knowledge.

19:33And then at that point, embedding model might be the right thing to do in that agent chaining framework. Maybe another angle to look at this is that whether we should do iterative retrieval versus just retrieve at once. I think iterative retrieval is definitely useful, especially because now there are still a lot of headroom in the embedding model's performance. So that's why sometimes you have to retrieve multiple times because the models are not clever enough. However, in the long run, my suspicion is that iterative retrieval will be useful, but it will be a bit less useful if the embedding models becomes more and more clever.

20:14So once the embedding models are more clever, then maybe one wrong or two wrongs is going to be enough. Mm-hmm. If we go ahead and just assume that RAG is at least a dominant architecture for enterprise use cases where you care about proprietary data that is large with reliability, how do you go about improving like a RAG system, right? You can improve the LLM itself, but what are the other components that you guys are working on or what are the maybe challenges from the user, the builder's perspective to improve retrieval quality? Yeah, so I guess there are a few ways, right? One way is that you improve the prompting of the large English models.

20:50So, for example, you could tell the large English models to abstain if there's no relevant information in the retrieved documents. But because the large language model is so good these days, I think you don't need a lot of prompting anymore. It just responds to the instructions so well. And then the next thing is to improve the retrieval part, which is the bottleneck in my opinion, because most of our users found out that if they improve the retrieval quality directly, that affects the response quality. And improving the retrieval part, I think there are two ways. One way is you improve the embedding model.

21:25One way is that you improve some of the other things on top of that. For example, how you trunk the data, whether you do iterative retrieval, whether you put in some of the meta information in the data, so on and so forth. So basically, I would say there are two ways of improving. One way is to improve the neural networks, either the embedding models or the revampers, or you improve the ways to use neural networks with software engineering, right? Better trunking iterations or other kind of like heuristics or kind of like tricks on top of that. So what we are specialized in is that we want to improve the new networks because that requires a lot of heavy lifting.

22:02That's a very data-driven approach. We train our new networks on trillions of tokens at least, and we fine-tune them for special use cases. And this is something that probably a company should do instead of the end users should optimize themselves. And my long-term vision here is that some of the software engineering layers on top of the networks will be less and less needed when the networks are more and more clever. So for example, right now we already see that trunkings becomes less needed because the context window becomes longer and longer and the long context embedding model, no, relatively long context embedding model.

22:41Long context here means like 10K, for example, maybe 16K so that you can put a 50 pages PDF into it. Because this long context embedding model becomes much better, there's less of a need to trunk the documents into pieces of like five, 12 tokens. And I think this will happen, you know, in other dimensions as well, right? So maybe in the future, you don't have to turn your images into description of images and then give it to the text embedding model. That's what people are doing right now. Like everything is turned into text and then use a text embedding model. But when the embedding models are more clever and multi-model, then you don't have to do that anymore.

23:20Can you talk a little bit about just like the intuition for how fine-tuning or domain-specific embeddings improves performance? Yeah, fine-tuning and domain-specific embedding models are what we are very good at at Voyage. So just to have some context here. So what we do is that we start with a general purpose-based embedding model, which is also what we trained from scratch. and from there we first fine-tune or continue pre-tune, whatever you call it, on some domain-specific data. So for example, we fine-tune on two trillions of code snippets, tokens, and then we get a code embedding model and we do the fine-tuning on one trillion legal tokens and that's how we got the legal embedding model.

24:08And these domain-specific embedding models, I didn't use any of the proprietary data so that everyone can use them. but they really excel in one particular domain and the performance in other domains are not changed much. And the reason why we do this is because the number of parameters in the embedding model is limited. So because you only have like, you have a latency budget, something like maybe one second, sometimes like 200 milliseconds. You know, some people even want 50 milliseconds. And then basically it's impossible to use more than 10 billion parameters for embedding models. and we have limit parameters, any customization is very important because customization means that you use the limit number of parameters on the right tasks, the right domain, so that you excel in that domain.

24:55There's no way that you can use these 10 billion parameters to excel in everything. So that's why you have to specialize in one domain. And we have seen like 5 to 20 % of improvements by this domain-specific fine-tuning, depending on the particular domains. For code, we have seen 15 to 20 % of improvement, partly because we have a lot of data there. And the headroom there is also bigger because code retrieval requires a lot of deep understanding of the algorithmic part of the code. And for legal domain, the baseline is a little better, so the headroom is slightly smaller. So that's why we see 5 to 15 % improvement depending on the data sets.

25:35For some of the very complex legal data sets, we have seen bigger improvements. Just to make sure that our listeners can picture exactly like where the latency cost is coming from here in a search system like your data, you know, has been vectorized by an embeddings model. But then every query also needs to be translated into an embedding and then compared to the embeddings of your knowledge in order to feed that that LM for the generation that you want. Right. And so there's a sort of there's inference time latency here as well. I just think that's not obvious if somebody hasn't built a RAC system.

26:11Yeah, exactly. Exactly. So basically, at the inference time, you have to first turn the query into vectors and then do the search with vector database. And actually, related to this, the dimension of the vectors you produce also affects the latency for the vector-based search. If the dimension of the embedding is only 100, then it's going to be much, much faster than when the dimension of the embeddings is 1 ,000. So, and actually this is something we are very good at as well. So we produce embeddings that is like a 3x, 4x smaller dimension than some of the competitors. Yep. That makes, I mean, intuitively, you are creating embeddings models that use a limited number of parameters and dimensions, just given the sort of latency budget that any application has to create the best possible representation of proprietary data or domain-specific data.

27:08Yeah, exactly. And going back to the domain specificity and fine-tuning, so the second level of customization is that we can customize to a particular company, right? So we fine-tune on the proprietary data of a particular company and we can see 10 % to 20 % improvement on top of the domain specific fine-tuning as well. So, of course, there's a total budget in terms of how much additive improvements you have there, right? So if you start with like 50 % accuracy, then you only have 50 % headroom. But if you start with 90%, you only have 10 % headroom. So the improvement, absolute improvement varies a little bit across the domains.

27:48Maybe just advice to people who are building RAG systems. At what point do they begin to invest in some of these retrieval components? Yeah, I think they can do it even from day one as long as they have a prototype available. So basically my default suggestion for our users is that when they have the RAC, first of all, of course, you want to connect the components and at least see some response. And then probably do some kind of like basic profiling in terms of the latency and the quality. So you can check the retrieval quality, meaning that how often you retrieve relevant documents. There are some default ways to evaluate the retrieval quality.