The Opportunity for Products Like ima

Originally published in Chinese on February 25, 2025. ima is Tencent's AI knowledge-base app. This piece answers three questions about the category: where its core bottleneck lies, whether a knowledge base survives once models get strong enough, and whether the category belongs to big companies or small ones. What is the core bottleneck for knowledge-base products now: model capability, hallucination, or the underlying RAG and OCR? Current models handle retrieval and automatic organization of content in a knowledge base without any trouble. The core needs of a knowledge base are mainly collecting, organizing, and using content. Once a user puts all their knowledge content in one knowledge base, the model only needs to categorize and organize it. The core is understanding the content and producing output where the user needs it. If the product can go further and refine and upgrade what is in the user's library, helping the user move from getting information to using it, that is the next step up. By the state of the technology, extracting, categorizing, and organizing text is hardly difficult now and is a fairly mature stack. The core is what users need from the knowledge base. Model capability already meets the requirements, and to some extent the capability may not even be needed for collecting content. As things stand, hallucination can be largely removed after several rounds of reasoning, since reasoning corrects a request along several dimensions. For the knowledge-base product line, the underlying technology has basically no bar left, and what remains is mostly innovation in how the product is applied.