Claude certification RAG: one exam tests it, one excludes it, two never mention it

Claude certification RAG appears on exactly one of the four exams. Architect Professional names retrieval-augmented generation in its exam description and carries two task statements on it inside a 19% domain. Architect Foundations does the opposite and publishes embeddings and vector stores as an exclusion. Neither Foundations-level guide for the associate or developer tracks mentions retrieval at all. Study it on the wrong track and you are preparing for material that is not there.

Which exam tests it, and how we know

Two of those rows are worth stating carefully, because they are different facts. Architect Foundations publishes an exclusion: its guide has an Out-of-Scope Topics list and embedding models and vector database implementation details are on it. The associate and developer guides publish no such list — retrieval simply does not appear in what they describe. Absence from a syllabus and a stated exclusion are not the same thing, and only the first one is a promise.

What the Professional exam actually assesses

Two task statements, both inside Architect Professional's Integration domain, which is 19% of that paper and its heaviest. The first covers designing a pipeline with appropriate chunking and indexing. The second covers matching a retrieval strategy to the shape of the data and the pattern of the queries.

Both are design judgments. Neither asks how a particular vector store works internally, which model produced an embedding, or how similarity is computed. The guide's own published sample question in this area asks a candidate to identify which stage of a pipeline caused a fault — a diagnosis, not an implementation.

Chunking is a boundary decision, not a length setting

The recurring error is treating chunking as a number to tune. A cut made at a fixed character count lands wherever the arithmetic puts it, and on a structured document that is frequently halfway through a clause — a condition parted from the rule it qualifies, a figure parted from what it measures. What comes back is grammatically fine and useless.

Following the divisions the document itself provides leaves every piece able to stand alone. The companion judgment is what must accompany a piece for anyone to interpret it — a passage identified only by a heading three levels up is ambiguous the moment it is retrieved out of its surroundings, and the fix is at indexing time rather than at query time.

Fit the method to the question being asked

The second task statement is the more examinable one, and it comes down to recognising that similarity search is one instrument among several. It surfaces passages resembling the query — precisely the wrong behaviour when someone has typed a reference number and expects that record back.

The last row catches people out. Withdrawn guidance often scores higher than the guidance that replaced it, having been written on the same topic in the same vocabulary. When being current matters, that has to enter the retrieval step itself, not be inferred later from a passage that may never declare its own status.

Which layer actually broke

Retrieval failures look identical from the outside — a confident answer built on the wrong material — and have three distinct causes with three different repairs. Working out which layer failed before reaching for a fix is the skill the published sample question tests.

That structure is why the exam rewards diagnosis over remedy. Retrieving more passages raises the chance of catching a missing neighbour and does not fix the boundary that separated them. A heavier configuration applies sharper reasoning to material that was already wrong on arrival. Both are plausible, both cost money, and neither touches the fault.

For how this sits beside the rest of the Professional blueprint, see the published domain weights, and for the material Architect Foundations does test in its place, coordinator and subagent patterns. If you have not settled on a track, the certification comparison guide is the place to start. Official outlines are published on the Anthropic Partner Academy certification pages.

Key takeaways

• One exam of four assesses retrieval. Architect Professional, inside a 19% Integration domain, with two task statements.
• Architect Foundations excludes it in writing. Embedding models and vector database implementation details are on its published Out-of-Scope list.
• The other two guides never raise it. That is absence rather than exclusion — a weaker fact, but enough to spend your time elsewhere.
• What is assessed is design, not internals. Where to cut documents, what to index, and which method suits a query.
• Chunk on structure, not on length. A unit has to be answerable when it is read alone, out of its surroundings.
• Isolate the layer before fixing anything. Chunking, search and ordering fail the same way from outside and need different repairs.