liveBook · Manning

Multivalued vector field

As explained in AI-Powered Search

A multivalued vector field is a search-index field that stores multiple vector values for one document.

A multivalued vector field stores multiple vector values for one document, allowing search systems to represent many document chunks without splitting them into separate documents. section 15.2.1

The field can contain an arbitrary number of chunks, including overlapping chunks. A sentence or paragraph may therefore appear in multiple vectors while remaining attached to its original document, preserving context across retrieval. section 15.2.1

Overlapping document chunks

Instead of indexing every chunk as a separate document, store several overlapping chunk embeddings in one document field. Retrieval can then compare many matches while retaining the document’s broader context. section 15.2.1

How AI-Powered Search covers it

How deep it goes
Defines storage and chunking behavior, but stops short of implementation details or ranking algorithms.
Examples
Uses arbitrary and overlapping chunks, including repeated sentences or paragraphs, as the concrete data model.
What you can do after
Evaluate chunking trade-offs and understand how one document can retain context across multiple vector matches.
How it connects
Connects RAG with embeddings, vector databases, approximate nearest-neighbor search, and contextualized late interaction.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI-Powered Search. The section links point to the parts of the book it draws on.