What Happened
The company has combined the capabilities of two of its products: the Hologres analytical database and the Model Studio AI model platform. As a result, data developers can now call large language models (LLMs) directly from SQL – without Python, separate infrastructure, or the need to understand how the models themselves are structured.
Simply put: you write an SQL query, call a model within it, pass it text or an image, and receive a response as data that you can then work with.
Who Needs This and Why
SQL is a language that virtually everyone who works with data knows. It has been around for decades, is easy to understand, predictable, and integrated into familiar workflows. A vast number of analytics systems and pipelines are built on it.
When AI capabilities become accessible through this same language, there's no need to change the tech stack, learn new skills, or wait for help from another team. A data engineer can immediately incorporate a language model into an existing data processing workflow, just like any other function in a query.
This is especially relevant for teams that don't have dedicated ML specialists or GPU infrastructure. The entire computational load remains on the Model Studio side – the user doesn't have to worry about it.
Practical Applications
This integration enables several use cases that previously required separate tools and expertise.
PDF document analysis. You can pass the content of a document to a language model in an SQL query and ask it to extract the necessary information: summarize it, find specific data, or classify it. This is useful, for example, when working with contracts, reports, or technical specifications stored in a database.
Image understanding. Models capable of processing not only text but also images are also accessible via SQL. This means you can, for example, automatically describe images in a table or extract structured data from them – all within a single query.
RAG – Searching with proprietary data. One of the most popular approaches in enterprise AI is so-called RAG (retrieval-augmented generation), where the model answers questions by drawing not only on its training data but also on the company's specific documents or knowledge bases. Such a scenario can now be implemented using SQL, without building a separate architecture.
Why It's Not as Simple as It Looks
At first glance, this might seem like mere «syntactic sugar»: a model call wrapped in a familiar query. But a more significant integration challenge lies behind it.
Hologres must be able to correctly pass data to Model Studio, receive responses, handle errors, and integrate all of this into the database's transactional logic. In essence, the boundary between the analytics system and the AI platform becomes transparent, which requires tight integration at the architectural level, not just a simple «wrapper» over an API.
For the end user, this is all invisible. But it is precisely this invisibility that creates value: the developer doesn't need to think about what's happening «under the hood».
Limitations to Keep in Mind
The solution is tailored to the Alibaba Cloud ecosystem. If a team is already using Hologres and Model Studio, the integration feels seamless. If not, making the switch would require a deliberate decision to change or expand their infrastructure.
Furthermore, the quality of the results still depends on how well the prompts to the model are formulated. SQL simplifies access, but it doesn't eliminate the need to understand exactly what you are asking the model and in what format you expect the answer. Simply put, «asking in SQL» is not the same as «asking correctly».
What This Says About the Broader Trend
This move is part of a broader trend: AI capabilities are being increasingly integrated into the tools people already use for their work, rather than existing separately and requiring specialized knowledge to access.
Databases, code editors, spreadsheets, and task management systems are all gradually becoming entry points for AI. This isn't because it looks «nicer» in a presentation, but because that's where people actually spend their work time.
The ability to call a language model from SQL is a small but telling example of how the barrier to entry is being lowered. Not for everyone at once, and not for all tasks. But for a specific audience – data engineers and analysts – it represents a tangible shift in what is accessible without additional effort.
