Define retrieval engineering around getting the right authorized information to the application at the right time. The work can include ingestion, indexing, relevance, access boundaries, freshness, and evaluation.
A vector database alone does not define the role or establish that retrieval works.
Map the retrieval responsibility
| Layer | Ownership question |
|---|---|
| Source ingestion | Which documents enter the system, and how are changes detected? |
| Representation | How is content divided and described for retrieval? |
| Search and ranking | How are relevant results selected and ordered? |
| Authorization | Which results may this user receive? |
| Freshness | How are updates and removals reflected? |
| Evaluation | How does the team know the retrieved evidence is useful? |
Clarify which layers the hire owns and which depend on existing platform or data teams.
Separate retrieval from answer generation
An incorrect answer may result from missing evidence, poor ranking, stale content, or the model’s use of the retrieved material. Ask candidates how they would distinguish these causes.
Elastic’s RAG documentation describes grounding model responses in retrieved sources and connecting retrieval to access controls. For hiring, use those as separate responsibilities to assess. Elastic RAG documentation.
Use a fictional evidence task
Provide synthetic documents, sample questions, and simple access rules. Include an outdated document, a plausible but irrelevant passage, and information that one fictional user must not receive.
Ask the candidate to explain a retrieval approach and how they would evaluate it. Do not require a full implementation or a specific vendor.
Look beyond the happy path
Ask what happens when no document supports an answer, when permissions change, or when two sources conflict. The candidate should not assume that returning something is always better than returning no supported result.
Explore how they would debug a user report without exposing restricted content through logs or test fixtures.
Decide whether the specialty is needed
A small product may be served by a broader application engineer with retrieval experience. A dedicated role becomes more coherent when the retrieval problems themselves require sustained specialist ownership.
Use the AI application role guide and data quality interview. Discuss an AI search with Refery.