run-demo

name	run-demo
description	Present, rehearse, or drive the LLM observability demo — pre-flight checks, fresh trace generation, and a guided walkthrough of what to show and say in Langfuse and LibreChat. Use when the user says "run the demo", "demo this", "walk me through the demo", "prep me for a customer demo", or asks what to show for a given audience or time slot.

Run the Demo

You are the demo co-pilot: verify the stack is demo-ready, make sure there is fresh data, then guide the presenter act by act — or drive the demo yourself by generating live traces while the presenter narrates.

1. Pick the right demo

Ask (or infer from context) the audience and time slot, then choose:

Demo	Time	Best for	Script
Langfuse platform demo	45 min	Platform/ML engineering teams evaluating LLM observability	docs/LANGFUSE_DEMO_RUNBOOK.md
Agentic RAG demo	25 min	Technical audiences interested in agents + ClickHouse-native vector search	docs/AGENTIC_RAG_DEMO_RUNBOOK.md
Quick tour	10–15 min	Execs, first conversations	Pick 3–4 paths from docs/USE_CASES.md

docs/SA_FIELD_GUIDE.md has the full selection matrix, talk-track framing, and objection handling — point the presenter at it.

2. Pre-flight (always, before any demo)

./setup.sh --status        # every "Demo Readiness" line must be ✓

If anything fails, fix it now (troubleshoot skill); if the stack isn't deployed at all, use the deploy-demo skill first. Then make sure there is fresh, varied data to show:

docker compose run --rm text-to-sql python main.py     # 3 SQL traces
docker compose run --rm vector-rag python main.py      # 3 RAG traces
docker compose --profile tools run --rm test-scenarios # 40 eval scenarios (if scores look sparse)

LLM-as-a-Judge and code-evaluator scores appear ~30–60s after traces arrive. The full pre-demo checklist (browser tabs, datasets, playground state) is at the top of each runbook — walk the presenter through it.

3. During the demo

• Follow the chosen runbook's acts in order; each step has a click path, timing, and a Say: block. Feed the presenter the next step on request ("what's next?").
• To show live tracing, generate a trace on cue: docker compose run --rm text-to-sql python main.py --interactive (type the customer's own question), then refresh Langfuse > Tracing > Traces.
• To answer data questions on the fly ("what's our avg latency?", "show me that trace"), use the langfuse skill / ./scripts/langfuse-cli.sh.
• If something breaks mid-demo, use the runbook's Fallback Plans section first — it is written for exactly this moment. Diagnose quietly with the troubleshoot skill; don't debug on the projector.

Demoing LibreChat traces well

LibreChat trace richness depends entirely on whether the agent calls a tool. A generic prompt ("what is ClickHouse?", "testing 123") produces a thin ~10-observation trace — one LLM generation wrapped in LangGraph scaffolding — and a trivial start → agent → end graph. That looks underwhelming on a projector. Always drive LibreChat with a prompt that forces tool use:

• LLM Observability Analyst → "What were the slowest traces in the last hour and why?" or "Show me insights about my agents."
• ClickHouse Data Analyst → "What are the top 5 repos by stars in the github dataset?"
• Agentic RAG Assistant → "How does RAG reduce hallucinations, and how does ClickHouse fit in?"

These produce 45–60 observation traces with TOOL spans and multi-step reasoning loops — and a graph that actually shows the agent↔tools cycle. The node labels are LibreChat's internal agent IDs (cryptic but harmless); don't apologize for them, just narrate the loop.

Conversation-title generation is disabled in librechat.yaml (endpoints.all.titleConvo: false) so the trace list stays clean. If a customer's instance has titling on, filter Trace Name = LibreChat in Langfuse to hide the TitleRun noise.

4. After the demo

Offer the audience the self-serve path: clone the repo, run ANTHROPIC_API_KEY=... ./setup.sh --seed — or point their own coding agent at the repo and say "deploy this demo" (AGENTS.md makes that work).