There is a quiet assumption inside most enterprise AI projects: that the model is the thing you're buying, and everything else is plumbing. Pick the best model, wire it up, and the quality of the answers follows.
It doesn't. A frontier model pointed at the wrong context will give you a confident, beautifully formatted wrong answer. The same model, handed the right passage from the right document, will give you something you can act on. The model is rarely the bottleneck. The retrieval is.
That is the whole idea behind how we talk about SWIRL now: SWIRL makes your chosen LLM better. Not "use our assistant." Not "switch models." Keep the LLM you already trust - including an on-prem one - and give it dramatically better context.
"Better context" is a pipeline, not a slogan
It's easy to say "better answers." Here's what actually produces them in SWIRL, and why each step matters.
It retrieves live, across everything. Most RAG setups answer from a vector index that was built last week. SWIRL queries your systems at the moment of the question - SharePoint, Salesforce, iManage, Box, databases, the web - in parallel, honoring the permissions each one already enforces. The context is current, and it's only ever what the user is allowed to see.
It ranks with judgment, not just similarity. Relevance in SWIRL runs in three passes, and both models run locally: keyword and BM25 first (quoted phrases and exact terms honored), then embedding re-ranking, then a cross-encoder that reads the query and the document together and scores real relevance - not vector distance. The passage that actually answers the question rises to the top.
It serves the answer your organization approved. A model's "best guess" and "the version legal ratified" are not the same thing. SWIRL lets teams pin the canonical result for a query, so people and agents get the endorsed answer, not whatever ranked highest today.
Feed all of that to your LLM and the output improves - because the input did. Same model, sharper input.
And it does it without a second copy of your data
The conventional way to get "better context" is to copy everything into a vector database. That's a second copy of your content to secure, govern, keep in sync, and explain to your security team - and for regulated organizations, it's often simply not allowed.
SWIRL skips it. The data stays where it lives; nothing is duplicated; the LLM gets its context from a live, permissioned federation of your real systems. You get the upgrade without the liability.
Why this framing matters
If the model is a commodity - and increasingly it is - then the differentiator isn't which LLM you picked. It's whether that LLM can reach the right, current, approved knowledge inside your walls. That layer is what SWIRL is. It's the reason the same model gives a better answer through SWIRL than it does on its own.
SWIRL 5 is in preview now and goes generally available on July 15. If you'd like to see your own LLM get better on your own stack, request preview access - it includes a 30-minute guided stand-up against a slice of your systems.