# What is Indexing?

Also called Search Index, Index Building.

Indexing is the process of organizing content into a structure that makes search fast, whether an inverted index mapping terms to documents or an approximate nearest neighbor index over embeddings. Without an index, every query would have to scan the entire collection, so the index is what makes retrieval practical at scale.

A lexical inverted index records, for each term, the documents containing it along with statistics used for scoring. A vector index instead organizes points in embedding space so neighbors can be found without comparing against everything. Both trade build time and storage for query speed, and both must be updated as content changes.

Vector index families have distinct characteristics. Graph based structures give strong recall at low latency but consume more memory. Cluster based structures scan only a few partitions per query and are cheaper to store. Quantization compresses vectors and reduces memory sharply at some cost to precision. Parameters control where each one lands on that curve.

Freshness and rebuilds are operational realities. Some index types accept incremental inserts cheaply while others degrade as data is added and need periodic rebuilding, and deletions are frequently implemented as tombstones that keep occupying space until compaction. Changing the embedding model requires embedding the corpus again and rebuilding the whole index.

Index design also carries the filtering burden. Because most queries are scoped by tenant, permission, or date, the index must apply those constraints without either scanning everything or returning too few results. Systems handle this with pre filtering, post filtering, or partitioned indexes per scope, each with different cost characteristics.

## Key points

- Structures content so queries avoid scanning everything.
- Inverted and vector indexes solve different problems.
- Parameters trade recall against latency and memory.
- Deletes often leave tombstones until compaction.
- Changing embedding models forces a full rebuild.

## In practice

A collection of two million passages is indexed with a graph based vector structure. A query examines a few hundred candidate nodes rather than two million vectors, returning the top matches in about twenty milliseconds. When a tenant filter is applied, the search restricts traversal to that tenant's partition so results stay both fast and correctly scoped.

## Related terms

- [Vector Database](/en/glossary/vector-database)
- [Embedding](/en/glossary/embedding)
- [Document Ingestion](/en/glossary/document-ingestion)
- [Hybrid Search](/en/glossary/hybrid-search)
- [Recall and Precision](/en/glossary/recall-and-precision)

[Back to the AI Glossary](/en/glossary)