Architecture
This is an interactive map of SWIRL Enterprise 5.0. Every box is a real part of a running installation, and every arrow is labelled with the transformation that happens along it, not just the direction data travels.
Drag to pan, scroll to zoom, and click any box to see what it does and what it replaces. The Traces panel replays four journeys through the system - a federated search, recall of a document already read, an AI answer, and an agent searching - one step at a time, with a short explanation of each step.
The short version. One question goes out to every connected system at the same moment, under the user's own credentials. Results come back on their own incompatible scales, are re-scored on a single scale, deduplicated, and grouped so the current version of a document leads. Nothing is copied anywhere to make this work.
Reading the map
Colour carries meaning rather than decoration. Each box is one of:
| Kind | What it means |
|---|---|
| Source of truth | Authoritative state. This is what you back up. |
| Derived index | Rebuildable from a source of truth. Restore is optional; rebuilding is enough. |
| Transport | Moves work or events between parts. Holds nothing permanently. |
| Processing | Work done in memory while a search is running. |
| Service | A long-running process in your deployment. |
| Your sources | Systems you already own. SWIRL queries them; it does not hold them. |
| Optional | A separately licensed capability. |
A dashed arrow is asynchronous: it happens in the background and never delays the answer a user is waiting for. In the animated traces, particles on those arrows sometimes fade out mid-flight, which is an honest picture of background work that may not complete on any given pass.
What the shape tells you
Your data does not move
Follow any arrow from Connectors out to Microsoft 365, a database or an existing search engine. It goes out and the results come back. There is no arrow that carries your content into a SWIRL index, because there is no such index to fill. That single property is what removes the crawl schedule, the re-index window, and the second copy of sensitive content that has to be secured and audited separately.
Permissions are enforced at the source
Every connector arrow carries the signed-in user's own credential. Microsoft, your database and your existing indexes each apply their own access rules before SWIRL sees a single row, so SWIRL never has to reproduce your permission model - and cannot accidentally get it wrong.
One scale is what makes a merged list a ranking
Each system scores its own results on its own scale, and those numbers mean nothing to each other. The Relevance scoring and Re-ranking stages put every result on one comparable scale. Without that step a federated result list is simply several lists stapled together.
Answers are built from documents, not snippets
Trace the An AI answer journey and note the step into Document fetch. A search result carries a few hundred characters of preview text. SWIRL retrieves the real document, extracts its full text, and writes the answer from that. This is the single largest factor in whether an AI answer is useful or vague.
Where to go next
- SearchProviders - connecting a source.
- AI Search Guide - configuring answers and models.
- Semantic Cache Guide - recall, version clustering and canonical election.
- MCP Server - giving agents the same federated search.
- Quick Start - standing an installation up.