Does Weaviate Work in China? PIPL Cross-Border, Data Residency & Self-Hosting

For a product that runs Weaviate for users in mainland China, the first question is usually whether its endpoint can be reached. But reachability is not where the China decision is made. Weaviate is a vector database — its own repository calls it an “open-source vector database that stores both objects and vectors” — which makes it a data-at-rest store: you load it with the content you want to search, the data objects (text, attributes, metadata), and it keeps those objects alongside the dense-vector embeddings computed from them, for semantic and hybrid search and retrieval-augmented generation. For a China-facing product that content routinely describes people, so it is personal information. By default Weaviate runs on Weaviate Cloud (WCD), hosted on AWS in the United States and Europe with no mainland-China region; the core database is also BSD-3-Clause open-source and fully self-hostable. That puts three prongs on the table: a cross-border transfer and residency problem, the fact that embeddings do not anonymize personal data, and the in-country self-host lever.

Weaviate in China at a glance

What you actually store — indexed content and its embeddings

Weaviate is not a page to render; it is a store you fill with your own content. You create collections and load data objects — documents, support tickets, product records, knowledge-base articles, user profiles — each with properties (the raw text and attributes) that Weaviate keeps. For each object Weaviate also stores one or more dense-vector embeddings, numeric representations produced by an embedding model from that object’s content, which is what powers semantic and hybrid search and retrieval-augmented generation. Crucially, Weaviate keeps both: its own repository describes it as an “open-source vector database that stores both objects and vectors.” So the store holds two representations of your users’ information — the source objects and the vectors derived from them. When the objects describe people in mainland China — names, messages, behavior, support history — every one of them is personal information, and the embeddings are personal information too (more on that below). This is a persistent data-at-rest store, not a transient call.

It’s a residency and cross-border-transfer problem — and embeddings don’t anonymize it — under PIPL

Two things bite at once, and most teams only see one. First, the cross-border transfer. The moment you index a mainland-China user’s object into a Weaviate Cloud cluster hosted in the US or EU, you upload their personal information offshore. Under the Personal Information Protection Law that is a cross-border transfer of personal information (数据出境), and the duty is on you, the handler — not on Weaviate. Articles 38–40 require notice, a separate consent distinct from any general terms, and one transfer mechanism: a CAC security assessment, the CAC standard contract, or certification. Above the regulatory thresholds, or where the data qualifies as “important data,” the data-export security assessment (数据出境安全评估) may apply before anything leaves. A critical information infrastructure operator or high-volume handler also owes an in-country storage duty under the Cybersecurity Law’s Article 39 (formerly Article 37 — the data-localization article was renumbered by the 2025 amendment, in force since January 1, 2026, with its substance unchanged) that an offshore index cannot meet; the query traffic — search strings and the records they return — adds its own log of personal information.

Second — and this is the prong most teams miss — embeddings do not anonymize anything. A vector embedding is computed from the source text or image; it is not a random token and it is not anonymous. Published research shows embeddings are subject to inversion (reconstructing approximate source content from the vector) and membership inference (testing whether a given record was in the store), so a collection of embeddings of your users’ data is itself a collection of their personal information, carrying the same residency and cross-border duties as the raw text. “We only send vectors, not raw data” is false comfort — and in Weaviate the point is moot anyway, because it stores the source objects alongside the vectors. None of this turns on how fast the endpoint answers; it turns on whether your China users’ content had a lawful basis to leave the country.

Reaching the endpoint isn’t the question — keeping the index in-country is

So “can we reach Weaviate from China?” is the wrong test. The productive question is how to keep your China users’ objects and embeddings on a lawful, in-country footing — and Weaviate gives you an unusually strong lever, because the core database is genuinely open-source. Weaviate’s own install documentation states it is “available as a hosted service, Weaviate Cloud (WCD), or as a self managed instance,” and that “Self-managed instances use the same Weaviate Database as WCD.” The core database is BSD-3-Clause licensed and runs under Docker or Kubernetes — “Kubernetes is ideal for scalable, production deployments” — so the identical engine can run on mainland infrastructure, keeping the index and the embeddings in-country rather than offshore. Be precise about the license, though: the Weaviate Database core is BSD-3-Clause, while some enterprise add-ons in the repository’s wl/ folder carry a separate Weaviate License — check what you actually deploy.

The lawful path is to keep the data-at-rest store in-country — self-host the open-source engine on mainland infrastructure where you have the operations capability, or use a managed deployment that runs in-country through a licensed local operator — while minimizing and pseudonymizing what you vectorize (index the least identifying content, drop raw identifiers you don’t need) and obtaining the consent the transfer rules require. What this is not is a tunnel that ships the same objects and embeddings offshore anyway; that keeps the endpoint responsive and leaves every PIPL duty exactly where it was. This is a risk map, not a verdict: whether a transfer mechanism, in-country storage, or a reworked consent flow applies to your deployment depends on what you index, how much, and about whom — settle the specifics with counsel before you build.

The lawful path — map, localize, deliver

There is a lawful way to run a vector database for a China-facing product, and it has a shape. First, map: our China team inventories what you load into Weaviate — which objects and properties, what personal information they carry, whether embeddings of personal data are stored (they are), where the cluster runs (a WCD region or self-hosted), the query-log exposure, and on what consent basis China-user data is indexed. We frame the technical picture; you settle the legal conclusions with counsel.

Then localize: keep China-user objects and embeddings in-country. Because the Weaviate Database core is open-source, that can mean self-hosting the engine inside China where you have the capability to run and maintain it, or using a managed deployment that runs in-country through a licensed local operator — in either case minimizing and pseudonymizing what you vectorize and governing any residual transfer. Localizing means keeping the data-at-rest store on an in-country path — never a tunnel that ships the data offshore anyway. Where a license bears on a domestic managed option, that sits with the licensed operator and your counsel; 21YunBox is advisory there.

Then deliver: the China-facing app that queries the index is itself a public-facing service in the mainland, so it carries an ICP filing (备案) duty and needs compliant, in-country delivery — the 21YunBox Optimizer, in front of the stack you already run, with no rebuild and no re-platform. The result is a search-and-retrieval setup that runs legally and compliantly for your users in China. 21YunBox never uses or suggests circumvention of any kind. We are a compliant overlay and partner — not a competitor to Weaviate.

Get a compliance assessment →

Related reading:

• Cross-border data transfers under PIPL
• China’s Cybersecurity Law (data localization, Article 39)
• China’s data-export security assessment
• How to get an ICP filing for China

RELATED TECHNOLOGIES

Elasticsearch

Elasticsearch

OpenSearch

OpenSearch

Qdrant

Qdrant

Typesense

Typesense