loading
loading
The prebuilt, searchable data structure holding your embeddings (or keywords) so retrieval is fast instead of scanning everything.
You don't compare a query against millions of vectors one by one — you build an index once that organizes them for sub-second lookup. For vectors that's usually an ANN index like HNSW or IVF, which trades a tiny bit of accuracy for huge speed; for keywords it's an inverted index mapping each term to the docs containing it. Building and updating the index is the ingestion half of RAG, querying it is the retrieval half. Example: re-indexing nightly when new docs land so the retriever sees fresh content.
Plainly
Think of Index as roads and power for the app city. The prebuilt, searchable data structure holding your embeddings (or keywords) so retrieval is fast instead of scanning everything.
In practice
Use it when local behavior needs to become a reachable, reliable deployed service. In practice, define the owner, input, output, and failure mode before you rely on it.