SKAPIAN LAB◆ CASE STUDY · 08 — MULTI-RAG RETRIEVER
Ask anything. Cited back.
A single chat interface for a Premier League club — pulling answers from match databases, document libraries, and broadcast transcripts. No SQL. No BI tool. Every answer source-cited.
3
Retrieval sources fused
100%
Source-cited answers
1
Non-technical CEO unlocked
8 wks
MVP to live
— 01 / THE INTERFACE
One chat. Three retrievers.
The router decides which sources answer the question. Every answer is inline-cited — click through to the raw clip or page.
◆
Hi 👋 Ask me anything about match data, contracts, or broadcast.
What was attendance over the last 6 home matches?
→ routing to RELATIONAL DB
Attendance averaged 42,300 — up 8% vs prior season. Peak was Match 4 vs rival club, 47,800 through the gate.
📊 SOURCE · match-data-db · rows 1,204 – 1,210
Summarise the broadcast rights renewal terms.
→ routing to DOCUMENT STORE
Contract runs 3 years from Jan 2025. Opt-out at month 18. Revenue share: 62 / 38 in the club's favour. Auto-renew if not noticed by month 30.
📄 SOURCE · contracts/broadcast-renewal-2025.pdf · pg. 4
Find moments commentary mentioned "late penalty".
→ routing to TRANSCRIPTS
Found 4 references across 2 matches. Latest: Match 6, 87'"…and that's the third — a late penalty in the 87th."
🎙 SOURCE · match-06.mp4 · 01:27:14 · Whisper transcript
— 02 / THE THREE RETRIEVERS
Different shapes. One answer.
Each retriever indexes a different shape of data. The router decides which to invoke — sometimes all three on the same question.
RELATIONAL DB
Match Data · Demographics
Structured rows over years of match data — attendance, results, ticket revenue, segmented by demographic.
MATCHATTRES
vs Rival47,800W 2-1
vs Lions41,200D 1-1
vs Wolves43,440W 3-0
Postgres · indexed views
DOCUMENT STORE
Contracts · Reports · Manuals
PDFs, contracts, demographic reports, operations manuals — semantic-searched and chunked.
📄broadcast-renewal-2025.pdf
📄stadium-ops-manual-v4.pdf
📄demographic-report-Q4.pdf
📄ticketing-policies-2024.pdf
Azure AI Search · embeddings
VIDEO TRANSCRIPTS
Broadcast · Commentary
Match recordings transcribed via Whisper, timestamped, indexed — search the spoken word, jump to the clip.
01:27:14"…and that's the third — a late penalty…"
00:34:08"…breaks free down the left flank…"
Whisper · timestamped index
— 03 / THE ROUTER
Knows when to ask which source.
And — critically — knows when to fall back to "I don't know."The router is what made this a system instead of a demo.
— 04 / PIPELINE
Query to citation in under a second.
Query
Plain English
Router
Classify · pick source
Retrieve
Fan out · gather
Ground
LLM + sources
Cite
Inline link-out
— 05 / IN PRODUCTION
One CEO. Zero BI tools.
3
Sources fused live
100%
Source-cited responses
0
SQL learned
8 wks
MVP to production
"The CEO wanted to ask 'what was attendance last home match?' and get a number — not write SQL, not learn a BI tool, not ask an analyst at 11pm."
— The brief, faithfully
— 06 / BUILT WITH
The toolkit.
Azure OpenAI
Azure AI Search
Whisper
Postgres
React
— NEXT CASE STUDY