The first generation of doctrine-aware tools in DoW was chat-with-PDF: embed every paragraph of FM 3-0 or JP 3-0, retrieve the top-K nearest neighbors, ask the model to synthesize. The output looked impressive in demos and failed in operational use. The failure mode wasn't hallucination in the usual sense; it was confident composition of statements that were never made together. Two valid paragraphs from different doctrine documents get fused into a wrong third statement.
What works in the field is agentic RAG. The system reasons about the query first: what kind of question is this, what doctrine surface does it touch, what's the right retrieval strategy. Then a retrieval agent runs that strategy, an extraction agent pulls the actual quoted passages, and a synthesis agent answers, but only using the extracted text, with citations back to paragraph and page.
The trick is forcing the synthesis agent to refuse cleanly when the doctrine is silent. We score every answer for citation density, and answers below threshold are flagged as 'doctrine does not cover this, operator judgment required.' That refusal class is more useful than another invented answer; planners can act on the absence of guidance.
Where it shines is cross-doctrine queries. 'How does JP 3-13 frame information operations differently from FM 3-13?' is a question retrieval RAG cannot answer well because the right answer compares two passages from two corpora. An agentic system plans the comparison explicitly, retrieves both passages, and synthesizes a side-by-side with citations.
Three operational notes from deployments so far. First, the retrieval strategy planner is more important than the embedding model; we get more lift from a smarter planner than from upgrading the embedder. Second, classification of the query type ('definitional', 'procedural', 'cross-reference', 'judgment') is the first step and worth its own eval. Third, doctrine changes; the system needs to know what date the underlying corpus was last refreshed and surface that in the answer.