docs/DEMO_SCRIPT.md

Cortex Demo Script

Link: https://sk-aiu-cortex.hf.space · Access PIN: from the demo owner · Format: live walkthrough, with a recording as backup

Before you start (10 minutes ahead)

1. Open the link. If you see "warming up", wait 1-3 minutes (the Space sleeps when idle).
2. Settings → Access PIN → Unlock. The header should show "Shared: 20 of 20 left today".
3. Ask sample question 1 once. The answer is then cached, so it's instant and $0 during the demo.
4. Keep a second tab on Scenarios as a fallback if the LLM is busy.

Executive track (10 minutes)

Min	Screen	Say	Show
0-1	Overview	"Most AI here finds similar text. Real questions need connected facts. Same 73 PDFs, same model; the only difference is the graph."	Header stats: 1M+ nodes, 0.4 ms query
1-4	Ask Cortex, question 1 (EU AI Act)	"Left is today's RAG. Right is GraphRAG on FalkorDB."	GraphRAG lists all 6 controls, 4 procedures and 2 open findings. Point at the sources and the tokenomics strip
4-5	Engineer view → Evaluation	"We measured it on 34 questions with known answers."	97% vs 70% recall · 31 vs 19 fully correct · multi-hop 97% vs 47% · same tokens per correct answer
5-6	Compare two roles (CRO vs Branch Teller)	"Security is in the graph. The teller never sees audit findings; the model never even receives them."	"N sources withheld"
6-8	Scenarios → Fraud rings (role: Compliance)	"Same database, beyond RAG: 60 hidden fraud rings in a million records, in under 2 seconds, no false alarms."	Then switch to Branch Teller: access denied
8-9	Scenarios → Regulatory impact / Risk roll-up	"A new regulation lands: here's everything it touches, instantly. And the bank's risk score, from one control up to the board."	Impact table; bank score 60/100
9-10	Close	"Free to run, open to inspect, and you can try it yourself today."	One-pager with the link

Engineer track (5 extra minutes)

1. Engineer view → Pipeline: LangChain loader and splitter, local embeddings, and pattern extraction (295/295 relationships correct, 0 tokens).
2. Last queries: the exact Cypher for vector and GraphRAG retrieval.
3. Performance panel: FalkorDB execution times on the 1M-node graph.
4. X-Ray: every layer of the stack with its source file, live FalkorDB internals (version, memory, graphs, labels, indexes), real GRAPH.EXPLAIN / GRAPH.PROFILE plans, and the life of a question with timings.
5. Upload: drop a PDF and ask about it. It goes into a private sandbox graph that's wiped after 24 hours.
6. Code: github.com/Cosmius-SK/cortex/tree/main/cortex. No Docker needed: python app.py runs it all locally.

Likely questions

Question	Answer
"Isn't this rigged for the graph?"	Same PDFs, chunks, embeddings, LLM and prompt. Simple questions tie at 100%, and we show that.
"Would Neo4j do the same?"	Out of scope for this demo; a comparison write-up can follow.
"Our documents don't have IDs."	Uploads use LLM extraction; try it on the Upload tab.
"What does it cost?"	$0 here (free tiers). Tokens per correct answer are about the same as RAG.
"Is the data real?"	No. Cortex Bank is 100% synthetic.
"Licensing?"	FalkorDB is SSPL; check with Legal before production use.

Recording checklist (backup video)

• 1080p browser window, notifications off, PIN already unlocked, questions pre-cached.
• Recorded videos: docs/video/cortex-exec.mp4 (executive highlights, ~2 min) and docs/video/cortex-full.mp4 (full tour of every menu, ~3 min).
• Re-record with python docs/record_demo.py --mode exec|full after pre-caching the questions.