How chunking and embeddings work

When a document is attached to a Knowledge Base, the platform runs an indexing pipeline:

Source document → Parse → Chunk → Embed → Searchable index

• Parse extracts text from the source document.
• Chunk splits the text into passages sized for retrieval.
• Embed turns each chunk into a vector using the KB's embedding model.
• Index writes the vectors so they can be matched against queries.

What matters for you

• Indexing is asynchronous. After you attach a document, it is not searchable immediately. Search visibility is eventually consistent.
• The embedding model at query time must match the embedding model used for indexing. Mixed dimensionality is not supported.
• Chunking configuration is per-KB. If retrieval quality is off, the chunk boundaries are often where you'd tune first.

Chunking parameters, embedding model choice, and re-indexing endpoints are exposed in the interactive OpenAPI.