snowflake-labs AI Function Demos
Interactive walkthroughs organized by feature — each demo highlights specific studio capabilities so you can jump straight to what interests you. Not every demo runs the full create → evaluate → optimize pipeline; some focus on a single capability.
When to Load
Load from main skill when user intent matches DEMO: "demo", "example", "walkthrough", "show me", "how does this work".
Prerequisites: Before routing to a demo, ensure the user has a database and schema with CREATE TABLE/FUNCTION privileges. If not already established, the demo sub-skill will prompt for location.
Workflow
Step 1: Select Demo
If environment == snowsight, ask user:
Which demo would you like to try?
1. **Pseudo-Labeling & Teacher-Student Distillation** (~10 min)
Start with unlabeled insurance claims, generate ground truth labels
with a strong teacher model (Claude Opus), then build a cheap student
model and measure how closely it reproduces teacher decisions.
Features: Synthetic Data (pseudo-label), Create, Evaluate
2. **Prompt Optimization** (~10 min)
Take a policy-conditioned ticket router where cheap models fail badly
on unfamiliar company vocabulary, then watch prompt optimization close
the accuracy gap through prompt evolution and Pareto cost/quality analysis.
Features: Create, Optimize, Pareto Analysis
---
The following demos are currently available in the CLI only.
They may be supported in Snowsight in a future update.
3. **Quick Start: Create & Evaluate** (~5 min)
Build a toxicity classifier and measure its accuracy — the fastest
way to experience the core workflow. Uses built-in exact_match metric.
Features: Create (direct mode), Evaluate
4. **Custom Evaluation Metrics** (~5 min)
Build a legal contract field extractor and create a weighted composite
metric that scores 4 fields independently — governing law, parties,
effective date, expiration date. See how custom metrics give richer
signal than simple exact_match.
Features: Create, Custom Composite Metrics, Evaluate
5. **Multimodal: Document Extraction** (~10 min)
Extract structured fields from SEC 10-K filing PDFs, build a custom
composite metric for multi-field scoring, and evaluate extraction
accuracy with per-field analysis.
Features: Multimodal (PDFs), Custom Metrics, Evaluate
6. **Agent Research: Pre/Post-Processing** (~10 min) [Experimental]
Build a PII redaction function using Agent Research mode — the agent
searches the web for state-of-the-art techniques, proposes multiple
SQL UDF architectures with pre- and post-processing around AI_COMPLETE,
and you pick the approach. Then optimize the entire function body —
prompts, model, and SQL logic.
Features: Create (Agent Research mode, web search, SQL pre/post-processing), Optimize (body mode)
If environment == cli (or not set), ask user:
Which demo would you like to try?
1. **Quick Start: Create & Evaluate** (~5 min)
Build a toxicity classifier and measure its accuracy — the fastest
way to experience the core workflow. Uses built-in exact_match metric.
Features: Create (direct mode), Evaluate
2. **Pseudo-Labeling & Teacher-Student Distillation** (~10 min)
Start with unlabeled insurance claims, generate ground truth labels
with a strong teacher model (Claude Opus), then build a cheap student
model and measure how closely it reproduces teacher decisions.
Features: Synthetic Data (pseudo-label), Create, Evaluate
3. **Custom Evaluation Metrics** (~5 min)
Build a legal contract field extractor and create a weighted composite
metric that scores 4 fields independently — governing law, parties,
effective date, expiration date. See how custom metrics give richer
signal than simple exact_match.
Features: Create, Custom Composite Metrics, Evaluate
4. **Prompt Optimization** (~10 min)
Take a policy-conditioned ticket router where cheap models fail badly
on unfamiliar company vocabulary, then watch prompt optimization close
the accuracy gap through prompt evolution and Pareto cost/quality analysis.
Features: Create, Optimize, Pareto Analysis
5. **Multimodal: Document Extraction** (~10 min)
Extract structured fields from SEC 10-K filing PDFs, build a custom
composite metric for multi-field scoring, and evaluate extraction
accuracy with per-field analysis.
Features: Multimodal (PDFs), Custom Metrics, Evaluate
6. **Agent Research: Pre/Post-Processing** (~10 min) [Experimental]
Build a PII redaction function using Agent Research mode — the agent
searches the web for state-of-the-art techniques, proposes multiple
SQL UDF architectures with pre- and post-processing around AI_COMPLETE,
and you pick the approach. Then optimize the entire function body —
prompts, model, and SQL logic.
Features: Create (Agent Research mode, web search, SQL pre/post-processing), Optimize (body mode)
⚠️ STOP: Wait for user selection before proceeding.
Step 2: Route
Route based on the demo name the user selected (numbering differs between Snowsight and CLI menus):
If Quick Start: Load classification/SKILL.md
If Pseudo-Labeling: Load insurance-claim-routing/SKILL.md
If Custom Metrics: Load legal-doc-extraction/SKILL.md
If Prompt Optimization: Load policy-conditioned-routing/SKILL.md
If Multimodal Documents: Load pdf-field-extraction/SKILL.md
If Agent Research: Load redaction/SKILL.md
What to Expect
Each demo will:
- Create sample data in your account (with
DEMO_prefix) - Walk through specific studio capabilities (see feature tags above)
- Offer to clean up demo objects when finished
Not every demo runs every step of the create → evaluate → optimize pipeline. Demos are designed to be fast and focused on the feature they showcase.
Note: Demos execute real SQL and create real objects in your Snowflake account. All demo objects use the DEMO_ prefix for easy identification and cleanup.
Stopping Points
- ✋ Step 1: After presenting demo options, wait for user selection
Output
Routed to specific demo sub-skill (e.g., classification/SKILL.md, insurance-claim-routing/SKILL.md)