Contextual Retrieval
Context ManagementDefinition
Anthropic's RAG technique of prepending a short, chunk-specific context blurb to each chunk before embedding and indexing it (as both contextual embeddings and contextual BM25), so an isolated chunk retains the surrounding document context it would otherwise lose. Combined with reranking, it substantially reduces retrieval-failure rates versus naive embedding-only RAG, and prompt caching keeps generating per-chunk context economical at scale.
Tested in: CCAR-P
Why It Matters for the CCA-F Exam
CCAR-P integration and evaluation items treat retrieval quality as an architecture choice. Contextual retrieval is the named Anthropic technique for making chunks understandable out of document context.
In Depth
Contextual retrieval is Anthropic's retrieval pattern: before you embed (and before you build a BM25 index), you prepend a short, chunk-specific context blurb to each chunk. The blurb answers "what document is this from, and what does this chunk mean on its own?" so a passage that originally said "the fee is $125" becomes "In the Claude Certified Architect Foundations exam guide, the fee is $125."
The problem it solves
Naive chunking rips sentences out of their heading, table, and document identity. Embedding that orphaned sentence retrieves it for the wrong queries, and the model then answers with a confident number that belonged to a different product. Contextual retrieval attacks that at index time, not with a cleverer prompt at query time.
Anthropic describes two complementary indexes on the contextualized chunk:
- Contextual embeddings — the vector is computed on blurb + chunk, so semantic search sees the situated meaning.
- Contextual BM25 — lexical search also runs on blurb + chunk, so exact tokens (exam codes, error names, IDs) still match, with the document identity attached.
At query time you typically fuse the two ranked lists (for example reciprocal rank fusion) and pass the top chunks to Claude. You can still add reranking; contextual retrieval is about the representation you retrieve, not a replacement for every other RAG knob.
What the exam is not asking
It is not asking you to invent a new embedding model. It is asking whether you notice that isolated chunks lose meaning, and whether you fix that with situating context rather than by stuffing the entire source document into the prompt (which is a context-window and cost problem) or by hoping the generator will "figure it out."
Prompt caching pairs well with a stable corpus: cache the system instructions and maybe a small always-on glossary, not the whole retrieval dump. See Prompt caching and Held-out eval set for measuring whether the blurbs actually improved answer quality.
Frequently Asked Questions
Does contextual retrieval replace embeddings?
No. It changes what you embed. You still build a vector index (and usually a BM25 index) on the contextualized chunk, then retrieve and optionally rerank as usual.
Where does the blurb come from?
It is generated at index time, typically by asking Claude to write a short situating prefix for each chunk given the source document. That generation is a batch/offline cost, not a per-query cost.