Demo guide

Technical requirements checklist (before the demo)#

• Snowflake account with Cortex Code access, connection configured
• cortex --version succeeds
• make install run, make check passes
• A Snowflake database with sample tables available (e.g.

  SNOWFLAKE_SAMPLE_DATA.TPCH_SF1)

• Terminal and browser windows arranged side by side
• Dry-run once, timed, within the last 24 hours

Opening hook (30 sec — the problem)#

Live walkthrough (7 min — step-by-step with talking points)#

Step 1 — First impression (2 min)#

1. Show examples/chat_app.py — point out it is a short, readable file

   built from four calls: check_environment(), render_start_gate(), get_or_create_session(), panel() + chat_input_bar().

2. Run make chat.
3. Browser opens → show the panel UI (start gate, then the empty transcript

   and status line once started).

4. Type: "What tables are available in SNOWFLAKE_SAMPLE_DATA.TPCH_SF1?"
5. Watch the streaming transcript render in real time; point out the tool

   card that appears (Glob/SQL).

Step 2 — Approval gates (3 min)#

1. Type: *"Show me the top 5 customers by total order amount. Write the SQL

   and run it."*

2. Agent proposes SQL → approval banner fires. Walk through the full SQL

   shown, click Approve once → results appear.

3. Type: "Now create a summary table called TOP_CUSTOMERS in my schema."
4. Approval gate fires again — this time click Deny with a reason

   ("let's not create tables in the demo"). Agent acknowledges and suggests an alternative.

Step 3 — Structured output (1.5 min)#

1. Switch to make structured (examples/structured_output.py).
2. Show the output_schema in the code.
3. Type: "Give me a breakdown of orders by status."
4. Agent returns structured JSON → app renders it as a chart, not raw text.

Buffer (0.5 min)#

Land on the close from wherever you are in Step 3. If there is time left in the room after the scripted close, two things are worth showing but are not part of the timed 10-minute script:

• make headless (examples/headless_pipeline.py) — the same query() API

  with no Streamlit import, for CI and batch jobs.

• make bi-semantic or make backlog — copilot_rail(), showing CoCo as a

  right column in a product app rather than a full-page chat.

Closing (2.5 min — value recap + CTA)#

Summary points to land on screen or verbally:

• pip install "streamlit-coco[sdk]" — a working app in a handful of lines
• Approval gates on by default for destructive tools
• One library covers interactive panel, structured output, and headless use
• Open source (Apache-2.0), built for Snowflake

Call to action: "Try it against your own Snowflake account this week — the quickstart in the README gets you to a working app quickly. Reach out to the asset owner if you hit setup issues."

If they ask… (FAQ)#

Cleanup#