Two debriefs and a timestamp
A team is holding short of a compound it has visited twice before. There is a white flatbed truck parked against the wall and the team lead wants to know if it has been there before. He types the question into the mission workspace in plain language. What vehicle was at the compound during the last two visits. The answer comes back in seconds. A white flatbed, both visits, cited to two patrol debriefs with the date and time of each sighting.
He taps the first citation. The debrief opens to the exact paragraph where a different team logged the truck five weeks earlier. He is not trusting the AI. He is reading the mission record with the AI doing the finding. The distance between those two things is the distance between a hunch and a decision.
Ask a cloud chatbot the same question and you will also get an answer. It will be fluent and it will be confident and it will be a guess. That model was trained on the open internet, not on your unit's patrol debriefs. An AI answer without a source is a rumor with good grammar. And no operator stakes a decision on a rumor.
The vault goes in before the team goes out
EdgeLance mission RAG starts before anyone steps off. An analyst loads the mission's knowledge into a vault inside the mission workspace. Patrol debriefs. Imagery notes. Watchlists. Intel reports. The SOPs the team actually runs. The vault is indexed on the devices the team carries and it never leaves the mesh. There is no cloud call in the loop because there is no cloud in the loop. A team in a valley with zero connectivity has the same answers as a team parked next to a fiber line.
During the mission any operator can ask the vault a question in plain language. Retrieval pulls the passages that bear on the question. A local open-weight model composes the answer from those passages and from nothing else. Every claim in the answer is pinned to the source document it came from, down to the specific debrief and the timestamp of the sighting. We made the argument in an earlier post that an answer you cannot check is an answer you cannot use. Mission RAG is that argument built into the retrieval path.
The refusal is the feature
The hard rule is what happens when the vault does not contain the answer. The system says so. It does not improvise. A language model always has a fallback available because it absorbed a copy of the internet in training, and that fallback is exactly where fabrication lives. EdgeLance blocks it. If the mission record is silent on a question, the model is barred from substituting whatever it learned before it ever met your mission. We call this no parametric fallback. It is enforced in code, not requested in a prompt. A prompt instruction is a suggestion. A blocked path is a wall.
The refusal is what makes the answers worth trusting. An assistant that always answers teaches you nothing about when it is wrong. An assistant that says the vault has no record of that is telling you something true and immediately useful. You learn to trust the yes because you have watched it say no. Operators calibrate on machines the same way they calibrate on people. The teammate who admits what he does not know is the one you believe under fire.
A refusal is also a tasking cue. When the system says the vault has nothing on a named individual, that gap is now visible to the whole team. Someone can go collect against it. A confident fabrication would have papered over the same gap and nobody would have known to look.
The analyst feeds the machine
No parametric fallback changes the analyst's job. On a cloud chatbot a thin knowledge base produces confident fabrication and the analyst never sees the cost. On mission RAG a thin vault produces refusals, and every gap in curation shows up as the system saying it does not know in front of the operator who asked. So the analyst stops being a report writer whose products sit in a shared drive. The analyst becomes the person who feeds the AI. The quality of the vault is the ceiling on the quality of every answer the team gets downrange.
This matches how the fastest units already work. Chatham House's study of Ukraine's wartime innovation found that the units adapting quickest are the ones that close the loop between the people who collect and the people who act. CSIS made the same argument about autonomy. Software and the data behind it decide the next war before airframes do. Mission RAG puts the analyst inside that loop with a feedback signal nobody has had before. Load a debrief tonight and it is citable by every operator on the mesh tomorrow. Skip one and the gap announces itself as a refusal.
One search across the whole mission
The vault is one surface inside a wider one. Mission-wide search covers chat, vault, map markers and source documents in a single query. An operator does not have to remember whether the truck was mentioned in a formal debrief or dropped into a chat thread at 0200. The search finds it wherever it lives and the citation points to exactly where it was found. The mission record behaves like one body of knowledge because it is one body of knowledge.
Every answer also carries the standard EdgeLance provenance receipt. Which model produced it. Which policy routed the call. The receipt comes from the same inference gateway that stamps every model call on the platform, so a commander reviewing the mission afterward can trace any answer from the operator who read it back through the model that wrote it to the debrief that grounded it. Nothing in the chain rests on trust in the machine. Every link is a document you can open.
The team lead at the compound got a white flatbed, two debriefs and two timestamps. He read both entries, looked up at the truck against the wall and made his call with the mission record open in his hand.