Original video in Chinese.
Key Takeaway
- The topic feature in Meita AI Search is its core update, offering knowledge base functionality with support for multi-user collaboration and API calls.
- Creating and sharing knowledge bases accelerates the flow of knowledge and makes “RAG as Service” possible.
- For AI to become useful, domain knowledge and domain experience must be added, and Meita addresses this through topics and workflows.
- The topic feature strengthens AI’s input capabilities by integrating search results, documents, URLs, and articles.
- Meita strengthens AI’s output through features like concept maps and glossaries, improving the usability of knowledge bases.
- The future form of a question-answering engine will be more refined knowledge flow and deep collaboration between users and AI.
I added an AI customer service bot to my blog. Now you can ask it any AI-related question. For example, what is overfitting? How does a large language model predict the next token? Or can a large language model think like a human?
Behind this AI customer service bot is a huge knowledge base, with nearly 2,000 professional papers and articles. And behind the knowledge base is Meita’s RAG technology.
A few days ago, Meita AI Search launched its topic feature, which is really the knowledge base feature that everyone needs.
Users can save documents, web pages, articles, and even Meita’s search results into a knowledge base. Meita will clean and parse this material, and then the large language model answers user questions based on it.
For me personally, the two features I like most in this update are sharing and API.
In the topic sharing settings, if you turn on the “editable” permission, the people you share with can also upload materials into the topic’s knowledge base. That means multiple people can operate the same topic together.
For example, in an office setting, everyone on the team can update the topic knowledge base in a timely manner, so the AI keeps up with your business progress. In an academic setting, if a group is doing research together, whoever finds a valuable paper can upload it. It’s especially convenient and useful.
As for the API feature, it allows the large amount of valuable material accumulated inside a topic to be applied in more places. For example, the topic I mentioned at the beginning, containing 2,000 professional papers, was created by Meita itself. After I got the API information, with Cursor’s help, I didn’t write a single line of code—just pure conversation—and in less than half an hour I had built the AI customer service bot.
This is basically the wildly popular RAG as Service overseas. If your needs are only lightweight, then there’s really no need to deploy those RAG systems locally, which consume a huge amount of resources and are extremely complicated to debug. The cost-performance ratio is way too low. Really, after messing around with it myself, I found that using an existing solution is pretty good.
Looking back at the topic feature: creating a knowledge base is bringing knowledge in—Input; sharing and API are letting knowledge flow out—Output. Taken together, Meita is really doing one thing: accelerating the flow of knowledge.
Hello everyone, welcome to my channel. Modestly speaking, I’m one of the few creators in China who can explain the Why and How of AI clearly. What I offer is more valuable than tutorials. Remember to hit follow. If you want to connect with me, come to the newtype community. There are already 600 friends who have paid to join!
Back to today’s topic: the topic feature in Meita AI Search.
Meita is a question-answering engine product. But I think it probably won’t stop at being a question-answering engine, or rather, a question-answering engine should not be its final form.
I said in an exclusive video for the community before that if AI is going to become useful, it must make up for two things: domain knowledge and domain experience.
A lot of knowledge is non-public, circulating only within specific fields, and AI can’t get access to it. Or, even if it is public and online, you still simply can’t find it on the internet. If AI cannot access this domain knowledge, then it cannot stay synchronized and aligned with the user at the level of knowledge and information, and collaboration becomes impossible.
So AI needs RAG and knowledge bases to make up for domain knowledge.
Now look at domain experience. Experience is even more non-standardized and private than knowledge. For example, if you run a foot massage shop or a beauty salon, from the greeting when the customer first walks in to guiding them toward a package, there can be dozens of experience-based steps in total, and all of these live in the boss’s head as the best way to handle things. If AI can’t access domain experience, then it can only act like a rookie and force its way through.
So AI needs Prompt and Agent Workflow to connect with SOPs and make up for domain experience.
If we look at question-answering engines from the perspective of knowledge and experience, we find that on the output side, question-answering engines already do very well—they give you an answer, not a web page. But on the input side, information on the internet only supplements and updates the general knowledge they already have. Domain knowledge and domain experience are still lacking.
So Meita has done two things:
Beyond the scope of search, Meita adds workflows. For example, marketing analysis, stock analysis, and so on. This is an elementary way of teaching domain experience to AI. There is still a lot of room for iteration afterward.
The other is the recently launched topic feature. It gives AI an external knowledge base. This knowledge base contains four types of content.
First, search results. You can save them directly into a topic by clicking the plus button to the right of the question. In this way, many search results we find useful can be categorized and stored, entering the knowledge base in article form. Meita does this very well.
Second, documents. For example, PDFs and Word files. My most common approach is to skim through a paper first to get a rough idea, otherwise I don’t even know what I should ask or talk about. Then I hand it over to AI for further discussion.
Third, URLs. For example, some companies’ official websites contain a lot of press releases and product information. In that case, you can let AI crawl the content inside as a reference. This counts as a strongly directed instruction—telling AI not to search around aimlessly, but to look at this website.
Fourth, articles. By creating a new article, we can paste text directly into it. That saves us the trouble of creating a document first and then uploading it.
Through this external knowledge base containing these four types of content, Meita supplements AI with domain knowledge it originally lacked, or information from private domains, as well as public-domain information that needs special attention from AI.
All of the above belongs to strengthening input. And in terms of strengthening output, Meita solves a headache-inducing problem:
As the information in the knowledge base grows larger and time goes on, we may start to feel a bit lost, unsure what exactly is in it or what we should ask. At that point, the concept map, glossary, learning suggestions, content summary, and suggested questions become very important.
As long as your knowledge base has more than 5 items, Meita will automatically generate these for you.
It will read all the materials, establish links between the authors and the corresponding topics, and present them in the form of a concept map. This is especially useful for things like papers.
If you have a strong learning need, then the glossary, learning suggestions, and content summary will definitely help you. The larger the knowledge base, the more obvious the effect.
If the original combination of search plus AI is the 1.0 form of a question-answering engine, then the topic feature, after strengthening input and output, is the 2.0 form of a question-answering engine. It has advanced from a larger, broader funnel to a smaller, more precise one.
So what would the 3.0 form of a question-answering engine look like? I suggest Meita consider adding Perplexity’s Page feature into topics.
Because from 1.0 to 2.0, the information inside undergoes active refinement by the user and becomes closer to production needs. Users will be more motivated to hold multiple rounds of conversation and think within a topic. At this point, a tool like Page is especially suitable, because it materializes and productizes the process we’re used to: searching and constructing ideas at the same time.
With the help of a Page-like feature, users and AI work together and refine the information, or knowledge, once more, forming a new, next-level funnel. This is what I think the 3.0 form of a question-answering engine should look like, and the logic behind it.
From 1.0 to 2.0 to 3.0, I have a feeling: the flow of knowledge is like a stream of water, splitting from the main trunk into countless branches, and eventually converging back into the main trunk to complete a cycle. During this flow, the value of products like Meita lies in channeling and accelerating it.
OK, that’s it for this episode. After watching, remember to go try the product. See you next time!