โมดูลสร้างเวกเตอร์และการสร้างเวกเตอร์ฝังอัตโนมัติ

Vector search needs every object to have a vector. You can compute embeddings yourself, or let Weaviate do it via a vectorizer module that embeds objects automatically on import.

A vectorizer is a pluggable module (e.g. text2vec-openai, text2vec-cohere, text2vec-huggingface) that Weaviate calls to convert text into vectors. You choose it when defining a class.

The vectorizer is set in the class schema.

With a vectorizer set, you import plain objects and Weaviate embeds them for you. No need to call an embedding API yourself.

• Send object properties
• Module generates the vector
• Object stored with its vector

Not every property should influence the vector. You can skip properties (like IDs or timestamps) so only meaningful text contributes to the embedding.

By default Weaviate can prepend the property name to its value before embedding. Disabling 'vectorizePropertyName' keeps the vector focused on content rather than field labels.

You can also set vectorizer to 'none' and supply precomputed vectors at import. Useful when you already run a custom embedding pipeline or need a model Weaviate does not host.

Each vectorizer accepts options in moduleConfig, such as the model name or dimensions. Set them at the class level to control how embeddings are produced.

The same vectorizer embeds your query so it lives in the same space as stored objects. This is why mixing models or changing the vectorizer after import breaks search consistency.

When to use each:

• Auto-embedding — simplest, less code, Weaviate manages calls
• Manual vectors — full control, custom models, batch pre-processing

Pick a vectorizer per class, skip irrelevant properties, configure the model, and let Weaviate auto-embed on import. Or set 'none' to bring your own vectors. Keep the vectorizer consistent between import and query.

You learned that Weaviate vectorizer modules auto-embed objects on import. Configure the module per class, skip non-meaningful properties, set the model in moduleConfig, or use 'none' to bring your own vectors. Keep the vectorizer consistent between import and query for valid search.