Agentic RAG Consulting

Agentic RAG is retrieval that can rewrite the query, inspect what came back, retrieve again, and optionally call tools before generating an answer. Linear RAG retrieves once and generates. Agentic is worth it when multi-part questions fail a strong linear baseline. It is not a license to skip citations, query-time ACL, or refuse-when-empty. We still deliver in four weeks in your cloud when the hop graph is scoped.

When the golden set shows linear hybrid retrieval missing multi-hop questions — compare two policies, follow a citation to a second document, or join an identifier across sources. If linear RAG already hits, extra hops add latency and inference cost you pay. We measure the delta before we build a loop. Agentic is a scoped exception, not the 2026 default.

Only if each hop is hybrid, ACL-filtered, cited, and allowed to stop empty. An unconstrained loop hallucinates more because it keeps fetching until it can produce a fluent paragraph. We put refuse-when-empty on the last hop and score faithfulness in CI. Agentic without those controls is worse than linear RAG, not better.

A written hop budget from discovery, typically small enough that latency stays inside the product constraint and inference cost stays predictable. The budget is a control, not a suggestion. Traces show every hop. CI fails if extra hops do not raise hit rate on the golden set. You pay inference, so wasted hops are your money.

No. Agentic RAG still answers from documents. An AI agent also takes actions — tickets, CRM writes, refunds. We keep write tools off unless discovery names them. Mixing unbounded tools into a knowledge path is how RAG projects become un-auditable. If you need acting agents, that is a separate scope with different evals.

Standard implementation is four weeks when the hop graph is fixed in week 1. Financial services typically 8–12 weeks. Healthcare typically 10–14. Ecommerce typically 6–10. Thirty days on-call after handover. If the diagnosis is that linear RAG is enough, we say so and do not sell hops.

The client. Work runs in the client cloud. You own IP: prompts, hop graph, indexes, evals, runbooks. No shared training. Model-agnostic. You pay inference. Fixed-scope implementation fee plus platform fee, quoted in writing. We are not a hosted agent product.

Yes, if every retrieval hop applies per-user ACL at query time. A rewritten query must not become a way to search documents the user cannot open. We treat query rewrite as untrusted input to the retriever, not as a privilege escalation. Empty filtered retrieval still refuses.