Original video in Chinese.
Key Takeaway
- QAnything is a knowledge base product suited to ordinary users. It supports creating multiple knowledge bases and can handle documents and web content.
- QAnything’s robot feature can publish knowledge bases as links for team collaboration or AI customer service.
- QAnything has some innovations in RAG technology, using Rerank technology (two-stage retrieval) to improve retrieval accuracy.
- The article emphasizes the advantages domestic vendors have in AI applications, as well as the importance of knowledge bases as users’ data assets.
- Future directions for knowledge bases include text splitting based on semantics and support for multimodal content.
Full Content
Today I want to introduce a knowledge base that ordinary users can get started with right away.
I have a feeling: domestic vendors are about to start competing intensely in knowledge base products. Right now, there are roughly two camps stirring into action.
One is the model vendor camp, like Moonshot AI and Minimax. While developing large language models, they will definitely build consumer-facing products around knowledge bases. Let me use an analogy so you can understand:
If AI is [water], then the chatbots everyone has today are [bottled water]. These [bottled water] products are already being sold everywhere, so their value will definitely keep dropping. Even a leading product like ChatGPT will face pressure from user churn.
So, developing new product categories around this [water] called AI is something every model vendor must do. And knowledge bases are already a recognized necessity: there is demand on the consumer side, and there is also a market on the business side. Plus, adding to a chatbot makes logical sense, so everyone will definitely move in this direction.
The other camp is traditional internet companies. The reason is simple too.
What goes into a knowledge base? User data assets. And these are the data assets users care about most. Wherever those data assets land, that is where users will stay or migrate to. So whoever can make good use of large language model technology and first build the best-performing, easiest-to-use knowledge base product will be able to defend their turf in this round of AI competition, and even poach users from others.
Among the traditional internet company camp, the one I see moving relatively fast is NetEase. This company has always had a strong product gene. The product I want to recommend in this issue is [QAnything], which I recommended in Knowledge Planet a couple of days ago.
I have introduced many knowledge base projects before. To be honest, they all require a certain amount of hands-on ability to get running, and they are actually not very suitable for ordinary users.
I think for everyone, in this AI era, getting started and actually using it matters more than anything else.
QAnything is a product that is especially suitable for ordinary users. The product is very intuitive, and it is even better than many foreign products.
Take knowledge base creation and selection, for example.
Many similar products either only have one big knowledge base, or although they allow multiple knowledge bases to be created, you can only select one knowledge base and can only chat with documents within that one knowledge base.
QAnything supports creating multiple knowledge bases. So you can manage materials just like you would with folders. For example, I created three knowledge bases:
- One for papers related to large language models, all PDF documents;
- One for articles from my newtype public account, which are actually the scripts for my videos;
- One for all kinds of articles I come across in daily life and want to save.
If you want to choose different knowledge bases, it is very simple. Just click a few times and you immediately understand what it means.
When it comes to building applications, you can always trust domestic vendors.
I especially like QAnything’s slogan: everything can be asked. That is the technology trend.
At present, the objects you can ask questions about are documents and web pages. Once multimodal large language models become faster and cheaper later on, video will definitely be supported too.
I won’t say much about the document upload feature. Everyone can try the [Add URL] feature more. I uploaded copies of good public account articles I came across into it. Because I found that I often can’t remember which article I saw a certain viewpoint in. Now that I have a knowledge base, I can just ask the AI directly. It’s basically fuzzy search, and it’s pretty practical.
On top of the knowledge base, the Youdao team also added a robot feature. You can set some prompts for the robot, connect it to a knowledge base, and then publish it as a link.
In my view, the robot feature serves two purposes.
First, share the link with coworkers. For example, you can arrange for an intern to regularly upload team documents into the knowledge base, and then publish it internally in the form of a robot. That will definitely help the team.
Second, share the link with customers. For example, you can put the link in the public account menu bar and use it as AI customer service.
The reason I had this idea is that I noticed that in a knowledge base, besides uploading document sets, you can also upload question-and-answer sets, which is the QA everyone is most familiar with. For example, company introductions, product introductions, and so on. Every company definitely already has this information ready, and once you upload it, you can use it directly. A simple AI customer service setup is done.
After using it these past few days, I found that QAnything’s accuracy is pretty good. The Youdao team is paying attention to RAG technology, and they use Rerank technology, which the official description calls [two-stage retrieval].
Rerank is not some especially profound technology. About half a year ago, I saw experts on YouTube introducing it and sharing code. Its principle is very simple:
Based on the user’s question, we first filter out 50 relevant text chunks from the vector database. But of course, you can’t feed all 50 into the large language model. On the one hand, there are limits on context length; on the other hand, some of those 50 text chunks are definitely only somewhat relevant. At this point, you enter the Rerank stage and rank those 50 text chunks by relevance. For example, we set it up so that the 3 or 5 most relevant ones are sent to the large language model.
After this whole process, because a Rerank step has been added, retrieval accuracy will definitely improve. But there is also a cost: speed decreases.
There are many nuances in RAG technology. What I just talked about was Rerank in the retrieval stage. There is also a lot of room for improvement in the earlier text-splitting stage.
The traditional approach, no matter how you set the chunk size, is never the most suitable. The ideal approach is to split based on semantics, so that the contextual meaning is not forcibly cut off. So who makes this judgment? Of course, the large language model does.
These kinds of new discoveries and new technologies have been emerging overseas all along. I hope domestic vendors can also keep a close eye on them. I’ve found that our understanding of technology domestically lags far behind. This kind of information gap is even bigger than the technology gap.
OK, that’s it for this issue. Next, I’ll introduce more products with lower barriers to entry so that more people can get started quickly. If you have any questions, come find me on Knowledge Planet. See you in the next issue!