What Is Retrieval Corpus?

How a Retrieval Corpus Works

Every retrieval-augmented answer is built from a bounded set of source material, and that set is the retrieval corpus. Before any query arrives, the documents in the corpus are split into passages through chunking, each passage is turned into a vector embedding, and those embeddings are stored in a searchable index. At query time the retriever does not roam freely — it searches only within this prepared corpus, pulling the closest passages to hand to the model.

Google Cloud’s RAG Engine states the definition directly: a corpus is “a collection of documents or source of information,” and “the corpus can be queried to retrieve relevant contexts for response generation.” In its API the store is a first-class resource named RagCorpus, into which you upload or import files that become the eligible source material. Whatever is inside that corpus is retrievable; whatever is outside it simply cannot appear in an answer.

That boundary is the whole point. Once you accept that a retrieval-augmented generation system can only cite what its corpus contains, the corpus becomes a strategy surface. The step that fetches passages — chunk retrieval — is only as good as the corpus it searches. A corpus of clean, self-contained passages gives retrieval strong candidates to rank; a corpus padded with duplicates and context-dependent prose forces it to waste its limited slots.

A corpus is also a living thing, not a one-time upload. Google Cloud’s RAG Engine supports create, update, import, and delete operations on a corpus precisely because source material changes: documents are added, revised, or removed, and the index has to be refreshed to keep answers current. A stale corpus produces confidently wrong answers grounded in outdated passages, which is the same failure mode a search engine hits when its crawl of the web falls behind.

Two Kinds of Corpus

In practice a retrieval corpus takes one of two forms:

• A provided corpus — a private, curated set of documents you supply: a product manual, a support archive, an internal wiki. The model answers strictly from this fixed body, common in enterprise assistants.
• The live web as corpus — for web-grounded AI answers, the corpus is effectively a search index rebuilt continuously by crawling. Here your page is only in the corpus if an AI crawler can reach, read, and index it.

Example of a Retrieval Corpus

A concrete, documented example is Google Cloud’s RAG Engine corpus. Its documentation defines the corpus as “a collection of documents or source of information” that “can be queried to retrieve relevant contexts for response generation,” and exposes it through a RagCorpus API resource with standard create, import, list, and delete operations. A developer populates the corpus two ways the docs name explicitly — uploading files from local storage, or importing files from other storage locations — and only after that import does the material become retrievable. The corpus is the queryable repository the engine draws from; retrieval never reaches past it.

The lesson carries straight over to open-web AI search. There, the retrieval corpus is a live search index, and the equivalent of “importing a file” is having a page that an AI crawler can fetch and index. If your content is not in that corpus — blocked, unindexed, or unreachable — it is not a candidate for any answer, no matter how well written. Membership in the corpus is the precondition for every citation that follows, which is why crawlability and clean, grounding-ready passages are the foundation the rest of AI visibility sits on.

Frequently Asked Questions

The Bottom Line

A retrieval corpus is the closed world an AI answer is built from — the indexed body of documents a retriever is allowed to search. Curate it well and the system has clean evidence to cite; leave it thin or noisy and no downstream model can compensate. For open-web answers, earning a place in that corpus starts with being crawlable and indexed.