Just like beer is destined to have foam, every gold rush has people who have been promoted to the forefront of the times.
In the boom of large models, the vector database is the lucky one.
On the one hand, there is not much breakthrough at the technical level. Vector database is not a particularly new database technology. It has been applied in the field of AI for seven or eight years. In 2015, Google announced the use of RankBrain semantic retrieval to handle search tasks. Compared with N's card, liquid-cooled computing, all-optical network, and upgraded storage, the vector database has no particularly impressive breakthroughs in technology.
On the other hand, the investment boom of vector database is particularly strong. In the first half of the year, it became an outlet for start-up companies, cloud computing vendors, established database companies, and investors to "crowd up and attack". Vector database start-ups such as Pinecone, Chroma, and Weviate have all obtained financing, and some of them have raised hundreds of millions of dollars Dollar. This is still a very impressive achievement under the uncertain investment situation of the global economy.
Unlike GPU cards, which have strong short-term demand and short supply, coupled with the constraints of Moore's Law, even if there is a bubble, it is made of iron. It is also different from "new infrastructure" such as Cunsuan. The strategic value of long-term investment has received unanimous attention from temples and folks.
The vector database, more as an AI basic technology and product, has begun to be known to the public.
With this alone, it is somewhat disturbing to soar in the investment market. In addition, recently, the popularity of large-scale training models has begun to cool down, and the number of visits to ChatGPT has declined. More large-scale models have gone to the fields and mines to "work".
People can't help but wonder how long the investment concept of vector database can fly as the wind of the big model sinks, and whether it will come and go suddenly, leaving companies and investors who have drunk "a mouthful of bubbles" What about messy in the wind?
Let's have a good taste, this glass of beer and foam.
technical beer
Training and using large models is inseparable from a series of AI infrastructure. Therefore, as one of the infrastructure, the vector database does have something. Introducing the vector database, you can drink real "beer".
It is necessary to talk about the technology itself first.
Needless to say, the database is an essential IT infrastructure for storing and querying various data, which can be regarded as the "hard disk" of the data. Then, the vector database is a "hard disk" that is more suitable for AI physique. There are several characteristics to illustrate this point:
1. Necessity.
Vector database, as the name implies, is specially used to store and manage vector data. As a data structure, each vector contains multiple dimensions, and each dimension represents a different feature or attribute, such as the color of an image, the frequency of words in a text, and so on. The AI algorithm needs to learn from massive unstructured data such as images, audio, and text, and extract "features" represented by vectors so that the model can understand and process them. Therefore, vector databases are more suitable for AI applications than traditional relational databases.
2. High efficiency.
Each element has an index for easy access or modification of the value. Based on this, vector databases can quickly find the closest embedding to a given query by grouping and indexing, enabling efficient similarity search while reducing storage and computation costs.
Compared with the traditional stand-alone plug-in database, the retrieval scale of the vector database can be increased by ten times, supporting the peak capacity of millions of queries per second (QPS), and the delay is controlled at the millisecond level.
Imagine, if there is no efficient search technology support, a large language model often has billions or tens of billions of parameters, can only process a limited amount of input data, and cannot search a larger database, then in AIGC, search, and advertising recommendation algorithms The performance of other tasks will be limited.
A public data is that by using the cloud vector database, the per capita listening time of QQ music increased by 3.2%; the per capita effective exposure time of Tencent video increased by 1.74%; the cost of QQ browser decreased by 37.9%. The changes in these data lie in the retrieval efficiency and stable operation Performance, operational efficiency, and recommendation algorithms have been greatly improved.
3. Big demand.
With the acceleration of industrial intelligence, as well as the explosion of large models and other AI applications, the number of AI use cases in various industries continues to increase, resulting in a flood of data and storage and calculation tasks. The length of the embedding vector in the vector database is not limited. Limits, with good scalability, can vary according to AI use cases and models, better handle large-scale data sets.
Moreover, the vector database can expand the time boundary and space boundary of the large model, so that after the training is completed, the large model can also access the latest information of the vector database and learn about recent events.
In general, the vector database is a database that is more suitable for AI physique. It is outstanding in AI tasks and is becoming increasingly popular in the field of machine learning.
So here comes the problem. Some major technology companies that have accumulated in the field of AI for a long time, such as Google, Microsoft, Mate, and BAT, have technology accumulation in vector databases, and they can also export related capabilities and products. In addition, some database startups based on open source technology, such as Pinecone, Weaviate, Odrant, and Chroma, have gained market popularity in recent years.
It can be said that there is no shortage of vector database products and solutions on the market. So in 2023, how does this glass of technical beer bubble up?
bubble on top of the tide
It is not an exaggeration to say that the market status of the vector database is "from 0 to 1".
First, mass-market awareness has only just opened up.
Previously, the vector database was mostly used by AI companies, and it has only become known to the public this year. This cannot be separated from the help of some AI-related companies. At the NVIDIA GTC conference in March this year, Huang Renxun mentioned the vector database for the first time, emphasizing the importance of the vector database for large language models.
Not all enterprises have the ability to self-build the infrastructure required for large models. It is a more flexible choice to train and apply large models through MaaS (Model as a Service), which requires cloud vendors to provide full-stack infrastructure.
Baidu, JD.com, Tencent, Huawei, etc. all mentioned vector databases in their own large-scale complete infrastructure. At present, the MaaS business of cloud vendors has just begun to enter the market, and the industrial implementation of large models cannot be achieved overnight. The acceptance and scale of vector databases are still unknown.
Second, the technology of the vector database has not yet experienced the iteration of "life and death".
Pinecone is the closed source leader. Other competitors are either open source, such as Weviate, or giants, including leading cloud vendors and veteran database vendors such as Oracle and IBM, who have begun to build AI database products and solutions.
Big manufacturers gather together to compete, which means that if there is no major breakthrough in technology, they will fall into high-density homogeneous competition, and quickly enter the red ocean from the blue ocean. And if there is a subversive change in technology, many new entrepreneurs with low technical barriers and low customer awareness will find it difficult to compete with open source ecology or technology giants, and they will easily be washed away by big waves.
Finally, the cost of the vector database has not been reduced to the level of "reproducible on a large scale".
Whether it is a self-built vector database or access through MaaS services, it has not yet reached the level of "paid and available". Generally speaking, enterprises need to vectorize unstructured private data first, generate a matrix of vectors, and then store them in the vector database for large model learning and retrieval. This process involves a lot of engineering, which will consume many developers and time costs of the enterprise.
This requires cloud vendors or database vendors to provide full-link tools to help enterprises complete the entire data vectorization, large-scale model access work, and reduce the difficulty of subsequent operation and maintenance. For example, Pinecone, with its good out-of-the-box product experience, has achieved very large growth, and its B-round valuation reached 750 million US dollars.
Google Cloud, Tencent Cloud, JD Cloud, etc. have also launched a series of external-oriented tools, frameworks, and applications based on years of accumulation of internal applications. But it is only the first step from scratch, and it needs to be "rolled up" to truly mature.
It can be seen that at this stage, vector databases are popular, and there are indeed many practical needs in AIGC, large models, cloud services, etc., but there is still a long distance between "popularization of concepts" and "real usability". The zone in between is where the bubbles grow.
The rivers and lakes are far away, the wind is high and the waves are turbulent. For start-up companies or industry users who have not thought clearly, it is better not to "bring capital into the group" rashly.
Sipping the Brew of the Times
If you are a database manufacturer, or an enterprise that is anxious to deploy large models and AI applications, and hope to drink beer to your mouth as soon as possible, what should you do?
Looking forward to the future, some tracks have a relatively small proportion of bubbles, and the demand is particularly strong.
In terms of market, localized substitution is a good choice.
Against the backdrop of technological competition, coupled with the growing prosperity and technological breakthroughs of my country's database industry, companies in key basic industries such as finance, telecommunications, energy, and transportation have begun to prefer domestic production when selecting databases to ensure data stability and safety.
Foreign manufacturers have explored and accumulated earlier in the vector database, and it will take time for domestic databases to make up for their shortcomings.
At present, strong domestic technology companies such as BATH have accumulated the core independent technology of vector databases, cooperated with them in research and development and customized development, and provided specific optimized vector database products for certain specific scenarios, adding localized alternatives The track is a choice with lower cost, more controllable risks, and clear market demand.
In terms of strategy, don't go alone when you join the cloud ecosystem.
Given that the commercialization prospects of vector databases are still unclear, some insiders said that instead of investing in new vector database projects, it is better to pay attention to which existing databases can be made more powerful by adding vector engines.
Cloud database is one of them. It is the general trend to go to the cloud and use data to empower intelligence. Many government and enterprise customers often choose public cloud or industry cloud to meet their business needs and migrate data to the cloud. The attention and acceptance of cloud database rise.
Large-scale cloud vendors such as Tencent Cloud and Huawei Cloud have high brand recognition and market acceptance, and have cloud-native and AI-native technology stacks and product systems. It is a safer way to mine the vector database together with the cloud ecology.
Like AI and large models, the taste of vector databases cannot be separated from the storage and brewing of time. Whether to disappear like a bubble after the popularity of the big training model drops, or to settle down as craft beer, waiting to become the rigid demand of the next generation of digital infrastructure, to be sipped by industry customers is the choice left to database players and buyers question.
