Use agentic retrieval only when adaptive evidence work justifies the added cost and control surface

Canonical RAG and agentic RAG solve different problems

A canonical RAG system follows a relatively stable pipeline: transform the query, retrieve from approved sources, rank evidence, construct context, and generate. Agentic RAG can decide which retrieval operation to run next, which source class to use, whether evidence is sufficient, and when to verify or escalate.

Research surveys describe agentic RAG along dimensions such as agent cardinality, control structure, autonomy, and knowledge representation. [S1]

Use adaptive retrieval for a measurable reason

Useful reasons include multi-hop evidence, heterogeneous repositories, ambiguous questions that require clarification, source-specific tools, dynamic gap analysis, or a need to verify claims against independent evidence. Avoid agentic retrieval when the corpus and question class are stable enough for deterministic search and ranking.

Reference architecture

A production design typically separates the user or case goal, planning state, retrieval policy, source registry, query tools, evidence store, ranking and deduplication, context assembly, generation, claim verification, human review, and trace evidence. The agent should not receive unrestricted access to every repository merely because it can formulate queries.

Source authority and tenant boundaries

Record which sources are authoritative for each claim type, which tenant or case may access them, how freshness is determined, and how conflicting versions are handled. Retrieved content remains untrusted input and can carry indirect instructions or poisoned claims.

Evaluate retrieval and generation separately

Measure whether the necessary evidence was found, whether irrelevant or prohibited evidence entered context, whether citations support claims, whether the answer is complete, and whether the system knew when evidence was insufficient. RAG evaluation research emphasizes the need to examine retrieval quality, factuality, safety, and efficiency rather than one aggregate score. [S3]

Evaluate the retrieval trajectory

For agentic RAG, inspect query reformulation, source choice, tool arguments, iteration count, evidence gaps, verification, and stop reason. A correct final answer can still be produced through an unauthorized or excessively expensive path.

Budget loops and non-progress

Set limits for retrieval attempts, source classes, tokens, latency, money, and repeated queries. Detect oscillation and retrieval that adds no new evidence. One experimental comparison found that agentic RAG could cost substantially more than simpler approaches in its tested setting, reinforcing the need for a baseline and cost evidence. [S2]

Security and privacy controls

Use source allowlists, task-scoped identity, result-size limits, content labeling, injection-resistant instruction handling, sensitive-data filters, and provenance. Keep policy and authorization outside retrieved text.

Stop conditions

Stop and ask for human input when the question is underspecified, authoritative sources conflict, access would exceed the case boundary, the evidence budget is exhausted, or the consequence requires qualified judgment. A refusal or unknown result is preferable to fabricated completeness.

Research boundary

Agentic RAG is an emerging architecture family with task-dependent evidence. The cited research does not establish universal superiority. Test the actual corpus, users, latency, cost, authority, and failure conditions.