Follow a single search through SWIRL: how the question is prepared for each source, how the sources answer on their own incompatible scales, and how SWIRL puts every result on one scale, removes duplicates, leads with the approved version, and feeds the top results to one cited answer.
A person types a question into SWIRL, or an AI agent asks over MCP. Plain words, quoted phrases, AND, OR, NOT and source tags are all understood.
Why it matters One search box reaches every system a person is allowed to see.
SWIRL rewrites the question for each connected source based on that source's own settings. Exclusions are rewritten or dropped for sources that cannot handle them, a tag can route part of the query to one source, and optional rules add synonyms, correct spelling or let a language model rephrase.
Why it matters Each system is asked in a way it understands, so recall is not lost to syntax.
The prepared queries go to every selected source at the same moment. Each runs under the signed-in user's own credential, so every source applies its own access rules before SWIRL sees a single row. Nothing is crawled, copied or indexed.
Why it matters No second copy of sensitive content to secure, and no permission model to reproduce.
Each source answers with its own ranking and, sometimes, its own score. A 7.12 from one system and a 0.63 from another mean nothing to each other. SWIRL keeps the source rank and score beside every result and never compares them directly.
Why it matters Without a common scale a federated list is several lists stapled together.
SWIRL reads the query and each result together. It matches word stems and highlights every hit, compares the meaning of the query with the title, the body and the text around each match, gives title matches more weight, rewards exact phrases, and normalizes for result length and query length so short and long results compete fairly. The source's own rank adds a small boost; results that contain an excluded term are down-weighted.
Why it matters Scores from different systems become comparable, so the best result wins wherever it lives.
SWIRL Enterprise adds a third pass: a cross-encoder reads the query and each top result together and re-scores them. Every result then carries a confidence score from 0 to 1 that is comparable across searches, and the Galaxy UI stars results above a configurable threshold.
Why it matters One threshold you set once, and can trust on every query.
Exact duplicates are removed and near-duplicates are grouped behind one result. With the separately licensed Semantic Cache, drafts, finals and copies of the same document collapse into one version cluster. SWIRL auto-selects the best version, favoring authoritative sources, final naming, recency and format, and a version a curator has pinned for the query leads the list.
Why it matters One entry per document, and the approved version is the one people open.
The final list is ordered by confidence, with ties broken by date and source rank, and every match highlighted. When RAG is on, SWIRL takes only the results above the confidence threshold, fetches the full documents with the user's credentials, reads the most relevant passages, and writes one answer with citations that link back to each source.
Why it matters Answers are built from documents, not snippets, and every claim has a source.
SWIRL queries your systems at search time. There is no crawl schedule, no re-index window and no second copy of content.
Each source enforces its own access rules on the user's own credential. SWIRL cannot show what the source would not.
Every result carries its source, the query as sent to that source, the source's own rank, and an explanation of its SWIRL score.
Based on SWIRL public documentation and the open source Community edition.docs.swirlaiconnect.comswirlaiconnect.com