As large model technology develops – is it possible that RAG technology will become obsolete?
RAG technology: Evolution and future in the era of large models
Core content:
1. RAG technology will not be replaced by large models, but will continue to iterate and upgrade
2. RAG technology's unique advantages in cost, real-time and domain adaptation
3. Future development of RAG technology: becoming the core infrastructure of intelligent systems
“ RAG technology will not be eliminated in the future, but will continue to upgrade with the iteration of technology. ”
There are many different opinions about RAG technology. Some people believe that RAG technology is a phased solution for large models and will eventually be eliminated; but some people believe that RAG technology cannot be replaced and its role is still irreplaceable; and there are also views that RAG technology will be upgraded with the iteration of large model technology, and more will form coordinated development.
From the perspective of enterprises, due to the uncertainty of the future of RAG technology, some people still have various concerns about RAG.
So, let’s discuss today about the future development of RAG.
What does the future of RAG look like?
Regarding the question of whether RAG technology will be eliminated, the author believes that it will not be eliminated; but it will definitely continue to iterate and evolve with the development of large model technology.
Let’s take a closer look at why RAG technology will not be eliminated.
First of all, the reason why RAG technology is being eliminated is simply that the capabilities of large models are becoming stronger and stronger, the parameters are becoming larger and larger, and the generalization ability is becoming stronger and stronger, so external enhancement tools such as RAG will no longer be needed in the future.
But let's think about a problem. Although the capabilities of the big model will become more and more powerful, its capabilities are always limited. It is impossible for the big model to truly become an omnipotent "god". At least in the foreseeable future, this situation is almost impossible to achieve.
Secondly, the current model mainly uses the pre-training method, although there are also feedback learning and other methods; but the large model has not really learned to learn independently and is still unable to process real-time data, which is where RAG comes into play.
Furthermore, the first principles advocated by Marx are all based on physical laws. As the capabilities of large models increase, their volume and storage space will also continue to increase. However, we know that physical storage has its limits, and it is impossible to store everything.
Furthermore, even if we do not consider the issue of physical storage, the large model relies entirely on its own capabilities to store all data; but what about the cost?
For many companies, the cost of training and fine-tuning a model is unacceptable; even if it is acceptable, if there is a cheaper and more convenient RAG technology, why not choose RAG instead of training or fine-tuning?
The cost of training and fine-tuning models includes not only financial costs, but also technical costs, time, and iteration costs.
For example, training or fine-tuning a model in the financial field can only be applied in the financial field; but what if there is a need to deal with problems in the legal field?
At this time, a new legal model needs to be retrained. Although fine-tuning and training can make the model perform better in specific fields, its shortcomings cannot be ignored.
However, there is no such concern when using RAG technology. You only need to train a model for an NLP task; then, when processing financial issues, load data from the financial field; and when processing legal issues, load data from the legal field.
Therefore, RAG has advantages that model training does not have:
Low cost
Good real-time performance
Flexible field adaptation
Moreover, judging from the current development of RAG technology, RAG technology has also undergone several versions of iterations; from the initial pure text retrieval RAG technology to the current similarity retrieval, graph structure storage, cognitive enhancement and distributed RAG, etc.
Furthermore, RAG technology can be combined with technologies from other fields, such as intelligent agents, to realize the autonomous retrieval function of RAG.
RAG will not disappear, but will evolve into the core infrastructure of intelligent systems. Its value lies not in replacing large models, but in building a sustainable and evolving cognitive system. Enterprises that reject RAG will lose key competitiveness in the deep waters of AI implementation.
