Skip to main content

Retrieval and grounding

Retrieval brings external evidence into a task; grounding ties a claim or action to that evidence. They are related but distinct: finding a relevant document does not establish that it supports the answer. The original RAG paper demonstrates combining a learned generator with retrieved non-parametric information in specific NLP tasks. Modern retrieval designs need not reproduce that architecture. rag

Selecting a retrieval strategy​

An indexed retriever provides repeatable access to a curated corpus, with index maintenance and access filtering obligations. Iterative tool-driven search lets the agent revise queries based on what it finds, adding latency and opportunities to pursue irrelevant leads. Anthropic describes selective runtime loading and hybrid approaches. context-engineering

As an engineering choice, use exact lookup for known identifiers, lexical search for distinctive strings, and semantic retrieval where paraphrase matters. Compare strategies on representative missing-evidence and ambiguous-query cases rather than assuming one retrieval mechanism is universally sufficient.

Evidence handling​

Preserve source identity, retrieval time and the passage or artifact supporting a material claim. Distinguish an original source from another page repeating it. Apply access controls before exposing results to the model, and treat retrieved content as untrusted data.

For example, a policy answer requires the applicable version and jurisdiction or business scope; a highly similar obsolete policy is still the wrong evidence. When sources disagree, narrow the claim, expose the uncertainty, or obtain additional evidence. Do not resolve contradictions solely from retrieval rank.

Evaluate retrieval and answer quality separately​

Check whether needed evidence was found, whether it was available to the model, and whether the final claim follows from it. A citation's presence is not proof of entailment. BrowseComp measures persistent search for difficult but verifiable answers while explicitly leaving out parts of realistic user-query handling. Its results cannot alone establish the quality of a research report. browsecomp

The established principle is external evidence access with attribution. The best mix of indexing and agentic exploration remains task- and corpus-dependent.

References​