Enterprise RAG Permission Assembly
Daniel Shimoni argues that permissions must be baked into how enterprise context gets assembled for RAG systems, not applied as post-hoc filters. The central claim: assembly is the last moment where refusing to include something prevents the model from ever seeing it. AWS Context and similar platform moves have converged on this architecture, but most enterprise software hasn't caught up—teams still rely on filtering retrieved context after the fact, which leaves temporal gaps (someone leaves the finance team at 9 a.m., the sync runs at 2 a.m., and for seventeen hours they can still access restricted documents) and fails under pressure from security review.
Transcript
Pippa Okay, so the claim is permissions should live in assembly context, not as filters after retrieval. Seems obvious, but apparently it's not.
Tyler It's architecturally obvious. It's also the kind of thing that sounds right in a whitepaper and then you actually try to build it and realize the hard problem was never the architecture—it was knowing who should have access to what in the first place.
Pippa Right.
Tyler Shimoni's core example is solid. Someone leaves finance at nine a.m. on a Monday, the sync runs at two a.m., and for seventeen hours they can still pull finance docs. That's real. But moving permissions into assembly doesn't fix that unless you have a permission system that's already live and accurate—and most teams don't.
Pippa But that's exactly why assembly matters. If permissions are a filter you apply after retrieval, you can pretend the problem is solved until security review asks you how you'd catch the case where the model summarizes the CEO's comp review for an intern who asked about salary bands. Then you have no answer.
Tyler Okay, fair point.
Pippa Assembly forces you to answer that before the model ever sees the context. You can't hide behind post-hoc filtering because the model never ingests it in the first place. It's not a better architecture because it's theoretically purer—it's better because it makes permission gaps impossible to hide.
Tyler I don't disagree with that. The structural argument holds. What I'm skeptical of is whether it actually ships. Shimoni's whole point is that vendors announced this—AWS Context in June, platform vendors converging on the same step—but the software most companies run hasn't caught up.
Pippa Oh interesting.
Tyler So you're right that assembly is the right place architecturally. But the reason teams are still shipping with post-hoc filters isn't because they don't understand assembly—it's because building the permission model to feed into assembly is miserable, and Lake Formation or equivalent is not something they're running yet.
Pippa That's fair. But then that's not a problem with assembly—that's a problem that assembly makes visible. Right now, teams can convince themselves filtering works because they never hit the edge case in a pilot. Assembly forces them to actually map their permissions before they ship to production. That's painful, but it's the right kind of painful.
Tyler Yeah, I think that's right. The pain is the point. Assembly doesn't solve the permission problem—it refuses to let you pretend it's solved until you've actually built it.
Pippa And Modus is betting that enough teams will feel that pain and decide it's worth building the permission model because assembly is table stakes. Which—given that AWS, the major vendors, everyone's converging on this step—seems like a reasonable bet.
Tyler Yeah. Though the question I'd ask is how long until that happens, because announced and shipped are two different things.
Pippa Fair. AWS announced Context in June. How many enterprises are actually wiring it into their RAG pipelines right now?
Tyler Probably fewer than the press release implied. The architecture is right, the principle is right, but the adoption curve for something that requires you to fix your permission model first is going to be long.
Pippa Unless security review gets aggressive enough that the cost of not doing it exceeds the cost of actually building it. Then adoption accelerates fast.
Tyler That's possible. Once compliance teams start asking how you guarantee the model never saw restricted context, filtering stops being an answer.
Pippa Exactly. That's the moment assembly goes from architecture to requirement.
Tyler I'd put that moment somewhere in the next six to nine months for teams actually touching sensitive data. Once one major enterprise gets asked that question and has no answer, the pattern spreads.
Pippa Yeah, that tracks. So the piece is right that assembly is the structural move. Whether it ships fast is still an open question, but the fact that vendors have converged on where the step goes suggests the answer is becoming obvious enough that it's going to happen.
Tyler Shimoni's timing on this is good. The argument in June would have been theoretical. In September, after the announcements, it's 'here's why the thing everyone just announced is the right answer, and here's why it hasn't shipped yet.'
Pippa Which is the most useful take—not 'this is right' but 'this is right and here's where it's going to get stuck.'
Tyler The stuck part is the permission model you have to build. But that's not a problem with assembly. That's the thing assembly refuses to let you skip anymore.
Pippa And honestly, that's the whole point of the piece. Assembly isn't a new thing—it's the moment where you finally have to stop pretending your permission problem doesn't exist.