How a 3‑Person Team Got 12k Users Without Marketing Using OSS Vector Bucket for a Low‑Cost AI Platform

Since 2025, the rapid spread of large‑model capabilities has enabled many young creators to move from content consumption to building their own virtual worlds. Matrees exemplifies this trend: a three‑person startup that, without any paid promotion, attracted more than 12,000 Gen‑Z users who have collectively authored over 45 million words of world‑building content.

Initially, Matrees relied on a self‑hosted open‑source vector database to support Retrieval‑Augmented Generation (RAG). User‑created world‑building elements—characters, items, events, races, geography—were vectorized and stored so that, when a user converses with the AI, the system can retrieve relevant context before generating a response.

As the user base grew beyond 7,000 distinct worlds, the hidden operational costs of the self‑managed solution escalated. Scaling storage and memory became a major burden, and the open‑source vector store required dedicated compute resources, regular version upgrades, backups, monitoring, and fault‑recovery—all of which consumed a disproportionate amount of the tiny team’s engineering effort.

Two months before writing this article, the team discovered Alibaba Cloud’s newly released OSS Vector Bucket. Built on top of Alibaba Cloud Object Storage Service (OSS), the service natively integrates vector storage, indexing, and retrieval. Unlike traditional stand‑alone vector databases, Vector Bucket adopts a “storage‑compute separation” architecture: vector and file data share the same storage layer, while indexing and retrieval are performed asynchronously via lightweight APIs.

The most compelling advantages for Matrees were full management and high cost‑effectiveness. No instance deployment, no pre‑purchased compute, and no need to handle cluster scaling or nightly alerts. Billing is based on actual usage—separate charges for storage, retrieval, and writes, with a free monthly write quota. For a seed‑stage startup, this “pay‑as‑you‑go” model reduced infrastructure expenses by roughly 90 % compared with maintaining a self‑hosted vector database.

Vector Bucket also solved Matrees’s most painful data‑synchronization issue. Previously, structured world‑building data lived in MySQL, while vectorized representations were kept in a separate vector store, requiring continuous sync and consistency work. As part of the Alibaba Cloud ecosystem, Vector Bucket integrates smoothly with MySQL and other Alibaba services, shortening the data‑sync path and making it more controllable.

The end‑to‑end workflow now looks like this: user‑created elements are stored in MySQL; when a user opens an “inspiration dialogue,” the system triggers embedding of those elements via an embedding model, writes the resulting high‑dimensional vectors to the corresponding Vector Index together with scalar metadata (type, relationships, etc.). During a dialogue, the user’s query is vectorized, a semantic search is performed on the world’s Index, the top‑K results are filtered by metadata, and the selected elements are injected into the large language model to generate a response that respects the world’s logic.

Vector Bucket’s storage‑compute separation enables asynchronous index construction without dedicated memory servers; a single bucket can host up to 100 index tables, satisfying multi‑world isolation needs; its model‑agnostic design allows seamless swapping of embedding models; and its serverless, on‑demand pricing cuts the total cost by about 95 % relative to a self‑built solution.

By encapsulating the complexity of vector storage behind a cloud service, Matrees’s developers can focus on the core creative experience, helping young creators turn imagined worlds into reality. As the platform’s slogan says, “From a seed to a world.”

References:

Matrees platform: https://www.matrees.cn/

Alibaba Cloud OSS Vector Bucket: https://www.aliyun.com/product/oss/features/vector-bucket