Curated scope
Administrators choose the manuals, doctrine, policy, SOPs, and local references appropriate to each pack.
Source-grounded AI for defense teams
Answers grounded in the knowledge your mission approves.
Mimir uses retrieval-augmented generation—RAG—to connect a locally hosted model with controlled technical and operational libraries. Users receive a synthesized answer, the supporting passages, and a clear path back to the governing source.
Grounded does not mean infallible
Citations improve traceability; evaluation and human judgment remain essential.
RAG can retrieve the wrong passage, miss a relevant source, or synthesize imperfectly. Mimir is designed to make evidence visible, support abstention and escalation patterns, and keep current approved publications authoritative.
01 / Governed knowledge
The mission owner defines the source of truth.
A defense RAG system should not quietly combine a unit’s controlled material with the open internet. Mimir organizes customer-approved documents into versioned knowledge packs aligned to teams, platforms, or mission sets. The pack defines the body of material available to retrieval for that context.
This model helps administrators manage revisions deliberately. When a technical order or SOP changes, the revised source can be reviewed, processed, tested, and promoted as a new pack version. Users should be able to understand which body of knowledge supported an answer and whether a more current governing publication exists.
Versioned publication
Content changes can move through review and acceptance instead of silently altering the production corpus.
Access boundaries
Pack availability and administrative functions are aligned to authenticated roles and the deployed identity model.
Revision awareness
Metadata and release records help teams identify the source set in use and manage superseded material.
02 / Retrieval pipeline
Find, rank, synthesize, cite—locally.
When a user asks a question, Mimir searches the selected knowledge pack for semantically and lexically relevant passages. A reranking step can reduce noise before the most useful context is presented to the local model. The response is generated against that bounded context rather than relying only on what the model learned during pretraining.
The interface returns citations with the answer so the user can inspect the supporting material. This is especially important in maintenance, policy, and operational workflows where the current approved publication—not the generated prose—remains the authority.
Hybrid retrieval
Semantic and term-based signals can complement one another when exact nomenclature and conceptual matches both matter.
Reranking
A second relevance pass helps prioritize passages before they consume the local model’s context window.
Bounded generation
Prompt and application controls instruct the model to work from retrieved evidence and expose uncertainty.
Source inspection
Citations give the user a direct route to the evidence used, supporting verification and escalation.
03 / Evaluation
Measure the behavior that matters to the mission.
A polished demonstration is not enough. A useful evaluation set reflects the actual document collection, terminology, user questions, and consequences of a bad answer. The team should test retrieval coverage, citation relevance, synthesis quality, abstention, latency, and behavior when the answer is absent or ambiguous.
Evaluation also belongs in the content lifecycle. A new manual revision, embedding model, reranker, local language model, or prompt can change system behavior. Representative tests and expert review help teams decide whether a release improves the mission outcome without introducing unacceptable regressions.
Retrieval coverage
Does the pipeline surface the passages a qualified reviewer expects for representative questions?
Citation correctness
Does each reference actually support the nearby claim, and can the user locate it in the source?
Unsupported-answer rate
How often does the system answer beyond the available evidence instead of qualifying or abstaining?
Operational usefulness
Do intended users understand the answer, verify it efficiently, and apply the governing workflow correctly?
Mimir is decision support, not an authority that replaces current publications, qualified maintainers, commanders, or established approval chains.
Continue the mission thread
See how the pieces work together.
Related capability
Air-gapped AI
Understand the complete disconnected operating model, including updates, support, identity, and recovery.
Related capability
Maintenance support
See how source-cited retrieval can reduce manual-search burden while the current publication remains authoritative.
Related capability
Knowledge packs
Review the standard troubleshooting pack model for Red Hat, Cisco, Microsoft, and VMware environments.
Plan the deployment
Bring the mission and the constraints. We’ll map the system to both.
A Mimir briefing covers the operating boundary, approved knowledge, deployment pattern, security evidence, and the people the system is intended to support.