Effective stock selection relies on having up-to-date data from diverse sources—research analyst reports, 10-K filings, investor call recordings, historical tick data, and real-time market feeds. To manage this efficiently, a unified platform is essential—one capable of storing and processing all data types, whether PDFs of analyst reports, MP3 investor recordings, or SEC documents.

This data must be readily accessible for analysis using the latest large language models (LLMs) such as Anthropic, Gemma, LLaMA, or DeepSeek. Ensuring the trustworthiness and security of generated summaries and insights is critical, especially when they inform high-stakes financial decisions.

In addition, developing Agentic AI capabilities allows for natural language question-answering tailored to investors and traders who need precise, real-time insights before executing trades.

In this hands-on lab, you'll learn how to build a Stock Selection Agent from the ground up using the Snowflake AI Data Cloud.

Learn how you can leverage the latest AI technologies right within the Snowflake platform. When AI is deeply embedded in your trusted data platform, the possibilities are endless. We will be exploring the processing of both Unstructured and Structured data which will then allow the application of a Cortex Agent to help discover insights by leveraging All Data.

This hands-on lab utilizes a comprehensive dataset spanning multiple data types and financial analysts to demonstrate real-world AI applications in financial services. You'll work with:

Analyst Research Reports

The lab features comprehensive analysis from specialized research firms created specifically for this hands-on experience:

Multi-Modal Data Integration

Throughout the lab, you'll process and analyze:

AI-Powered Analysis Journey

Your AI assistant will synthesize insights from all these sources to answer questions like:

This diverse dataset enables you to experience how modern AI can unify structured financial data with unstructured analyst opinions, creating comprehensive investment intelligence.

In this quickstart, you'll build a comprehensive AI-powered financial analysis platform called StockOne using Snowflake's Cortex AI capabilities. This end-to-end solution demonstrates how to:

No downloads or CLI tools required! Deploy directly from GitHub using Snowflake's Git integration.

By the end of this lab, you'll have deployed:

This quickstart deploys a complete multi-modal AI platform combining:

Unstructured Data → AI Processing → Structured Data → AI Services → Applications

Data Types:

AI Capabilities:

Latest AISQL Syntax: All examples use 2025 AI_* functions

Option A - Free Trial (Recommended):

1. Visit https://signup.snowflake.com/
2. Sign up for a free 30-day trial
3. Choose Enterprise edition
4. Select a cloud region (AWS, Azure, or GCP)
5. Verify your email
6. Log in to Snowsight (https://app.snowflake.com)

Option B - Existing Account:

Deploy directly from GitHub - No downloads or CLI tools needed!

1. In Snowsight, click on Projects
2. Select Workspaces
3. Add new SQL file
4. Copy and paste this script:

4. Click Run (or press Cmd/Ctrl + Enter)
5. Wait for completion (~30 seconds)
6. ✅ Git integration complete! You're now connected to GitHub

Now will create a workspace from the github repository

1. Create a Workspace from Github repository by clicking on My Workspace, then select the option From Git repository

   

2. When Prompted, populate the following screen

   

   Use the following URL in the Repository URL field:

3. Press Create

You should see the repository file structure:

1. Navigate to assets/sql/ - This is where the deployment scripts are
2. You'll see SQL files numbered 01-08 (plus optional 05b for GPU)

✅ You're now ready to deploy!

Option A: Execute all at once (in one worksheet):

Option B: Use Git Repositories UI (interactive):

1. Navigate: Projects → Git Repositories → SNOWFLAKE_QUICKSTART_REPOS.GIT_REPOS.ACCELERATE_AI_IN_FSI_REPO
2. Browse to: assets/sql/
3. Right-click each file (01-08) → "Open in new worksheet"
4. Execute each script in order

What gets deployed:

1. ✅ Database ACCELERATE_AI_IN_FSI with 3 schemas
2. ✅ Role ACCOUNTADMIN with CORTEX_USER privileges
3. ✅ 20+ tables with ~10,000 rows of data
4. ✅ 2 Cortex Analyst Semantic Views
5. ✅ 1 Snowflake Intelligence Agent
6. ✅ 2 Streamlit applications
7. ✅ 4-5 Notebooks (GPU notebook optional)
8. ✅ 1 Native App (SnowMail email viewer)
9. ✅ Stages with PDFs, audio files, and images

⚠️ Note: Search services are created later in Notebook 4, not during initial deployment.

Deployment time: 15-20 minutes

After deployment completes, verify in Snowflake UI:

⚠️ Note: Cortex Search Services will appear after you run Notebook 4.

All set? ✅ Continue to the next section!

After Git integration deployment completes, let's verify everything was created successfully.

Open Snowflake UI and navigate to Data → Databases → ACCELERATE_AI_IN_FSI

You should see:

Expected Results:

⚠️ Note: The following are created during the notebooks, not during initial deployment:

You should see:

Streamlit:

Notebooks:

All verified? ✅ Let's start using the AI features!

Before diving into the notebooks, let's explore the AI & ML Studio - your one-stop shop for trying out AI functions using a user-friendly UI.

Navigate to the AI & ML section in the Snowflake navigation bar.

The AI and ML Studio provides access to:

The Cortex LLM Playground lets you compare text completions across multiple large language models available in Cortex AI.

Try it now:

1. Click on Cortex Playground in the AI & ML Studio
2. Select a model (e.g., claude-4-sonnet, llama-3.1-70b, mistral-large2)
3. Try asking a financial question:

Example prompt:

What you'll see: The model will suggest various factors to consider:

Key insight: Notice how the model doesn't make a direct recommendation - this demonstrates responsible AI. All the information suggested (analyst reports, earnings calls, financial metrics) is available in the datasets you'll be working with in this lab.

Now let's explore how to extract text and data from documents. The Document Processing Playground provides a UI for testing AI_EXTRACT and AI_PARSE_DOCUMENT functions.

1. Click Document Processing Playground in the AI & ML Studio
2. Click Add from stage

3. Select the following:
   • Database: ACCELERATE_AI_IN_FSI
   • Schema: DOCUMENT_AI (contains sample PDF documents)
   • Stage: ANALYST_REPORTS (contains analyst report PDFs)
4. Choose 1-3 analyst report PDF documents

5. Click Open playground

Once your document is loaded, you'll see three tabs: Extraction, Markdown, and Text.

The Extraction tab is where you can ask questions to pull specific information from the document.

Try creating these key-value question pairs:

After entering each question, click Add Prompt to see the extracted results.

The playground provides three different views:

Click through each tab to see how the document is processed differently.

Once you've asked at least one question, the playground automatically generates SQL code:

1. Click Code Snippets in the top right corner

2. Review the generated SQL using AI_EXTRACT and AI_PARSE_DOCUMENT functions
3. Click Workspaces to open the code in a new worksheet

Example generated code:

This SQL code can be used to automate document processing at scale!

Let's extract structured tables from financial reports:

1. Close the playground and reopen Document Processing Playground
2. Within the same schema, select the stage FINANCIAL_REPORTS

3. Select 1-2 financial report PDFs
4. Open the playground
5. In the prompt area, select Extract table

Configure table extraction:

Why dynamic column names? Notice we used "Current Year" and "Previous Year" instead of specific years like "2024" and "2023". AI_EXTRACT uses the LLM to intelligently map these generic descriptions to the actual column headers in the document, making the extraction work across different documents with different year labels.

Why This Matters: These document processing capabilities are essential for the notebooks you'll run next. Understanding AI_EXTRACT and AI_PARSE_DOCUMENT helps you appreciate how the notebooks process financial reports, analyst documents, and unstructured data at scale.

Before proceeding to the notebooks, you now understand:

✅ Cortex Playground - How to test LLM capabilities and prompt engineering ✅ AI_PARSE_DOCUMENT - Extract text and structure from PDFs ✅ AI_EXTRACT - Pull specific fields from documents using questions ✅ Table Extraction - Extract structured tables with dynamic column mapping ✅ SQL Code Generation - Automating document processing at scale

Next: Apply these concepts in Notebook 1 to process hundreds of financial documents!

In this section, you'll process multiple types of unstructured financial documents using Cortex AI functions. You'll work with the AI capabilities you just explored in the Document Processing Playground, but now at scale across hundreds of documents.

What You'll Process:

Cortex AI Functions You'll Use:

Navigate to AI & ML Studio → Notebooks → 1_EXTRACT_DATA_FROM_DOCUMENTS

The notebook is organized into 4 parts:

Part 1: Analyst Reports

Part 2: Financial Reports

Part 3: Infographics from Images

Part 4: Email Analysis

After running this notebook, you'll have created:

💡 Note: These tables are not pre-loaded during initial deployment. The notebook demonstrates the full extraction process using Cortex AI.

The notebook contains detailed instructions and explanations for each step. Simply:

1. Click Start to begin the notebook session
2. Read through each cell's markdown explanations
3. Run All or execute cells one by one
4. Observe the outputs and interactive visualizations

Time: Plan for 10-15 minutes to run through all cells and explore the results.

Note: All data is pre-loaded, so the AI extractions happen quickly for demonstration purposes.

Process earnings call audio recordings using AI_TRANSCRIBE to convert speech to text, then analyze the content for sentiment and insights.

What You'll Process:

Navigate to AI & ML Studio → Notebooks → 2_ANALYSE_SOUND

AI_TRANSCRIBE - Converts MP3 audio to timestamped text:

AI_SENTIMENT - Analyzes emotional tone:

Vector Embeddings - Enables semantic search:

After running this notebook, you'll have created:

💡 Note: The call_embeds table (vector embeddings) is actually created in Notebook 4, not here.

The notebook includes:

Simply run through the cells to see how audio becomes searchable, analyzable data.

Time: 10-15 minutes to complete.

Build and deploy a machine learning model for stock price prediction using Snowflake ML. This notebook demonstrates GPU-accelerated model training and the Model Registry.

⚠️ Note: This notebook requires GPU compute pools which may not be available in all regions. If GPU is not available, a pre-trained model (STOCK_RETURN_PREDICTOR_GBM) is already deployed for you!

What You'll Build:

Navigate to AI & ML Studio → Notebooks → 3_BUILD_A_QUANTITIVE_MODEL

Feature Engineering:

Model Training:

Model Registry:

GPU Benefits:

If GPU is not available, the agent can still use:

Function: GET_TOP_BOTTOM_STOCK_PREDICTIONS(model_name, top_n)

This powers the agent's ability to answer:

1. Run through cells to see model training
2. Observe model registry integration
3. Test predictions with the SQL function

Time: 15-20 minutes if training from scratch (pre-trained model saves time!)

Create 5 intelligent search services that make your financial data instantly accessible using semantic search. This is the foundation for RAG (Retrieval Augmented Generation) applications.

Why Search Services Matter:

What You'll Build:

1. Analyst Sentiment Search - Search 92 earnings call transcripts
2. Full Earnings Calls - Semantic search across chunked Snowflake transcripts
3. Analyst Reports - Search 30 reports with ratings and targets
4. Infographics - Search 11 company earnings visuals
5. Email Content - Search 950+ analyst emails

Navigate to AI & ML Studio → Notebooks → 4_CREATE_SEARCH_SERVICE

Cortex Search provides:

Creation Approach:The notebook uses programmatic SQL creation (vs UI point-and-click) for:

Each search service targets a specific data type and use case:

1. dow_analysts_sentiment_analysis - Search transcripts by sentiment context
2. snow_full_earnings_calls - Search chunked call segments
3. ANALYST_REPORTS_SEARCH - Search reports by full text
4. INFOGRAPHICS_SEARCH - Search infographics and branding
5. EMAILS - Search emails by ticker, rating, or sentiment

Before creating the search services, the notebook creates supporting tables:

💡 Note: These tables are prerequisites for the search services and are NOT pre-loaded during initial deployment.

The notebook walks through:

After completion, your search services will be visible in AI & ML Studio → Cortex Search.

Time: 10-15 minutes to complete.

Next: These search services become tools for the Snowflake Intelligence Agent!

Cortex Analyst enables natural language querying of structured data using semantic views. Instead of writing SQL, users ask questions in plain English and Cortex Analyst generates the SQL automatically.

What You'll Explore:

Before running Notebook 5, you need to get the Snowflake Public Data (Paid) dataset from the Snowflake Marketplace:

1. Click on Data Products → Marketplace in the left navigation
2. Search for "Snowflake Public Data"
3. Find the listing from Snowflake Public Data Products (has 60-day unlimited trial)
4. Click on the listing and then click Get
5. Accept the terms and the data will be available as a shared database

Navigate to AI & ML Studio → Notebooks → 5_CORTEX_ANALYST

The notebook explores marketplace data and shows how to:

Time for notebook: 5-10 minutes

After completing the notebook, let's explore the semantic views that power the agent.

1. From the navigation bar, click AI & ML → Studio
2. Click on Cortex Analyst

You'll see 2 semantic views that have been set up:

Click on Snowflake Analysts View to see its structure:

Logical Tables:

Field Organization:

Each field has:

Click on Company Data 8 Core Featured Tickers:

Key Features:

Relationship Example:

Why Relationships Matter: They allow Cortex Analyst to automatically join tables when answering complex questions like "Show me sentiment from earnings calls for companies with negative free cash flow"

Looking at the Infographics for Search table, you'll notice where the primary key is defined.

Notice some dimensions have this annotation:

This connects the semantic view to your search services, enabling the agent to seamlessly query both structured and unstructured data!

Click Playground to test queries:

Example questions:

The playground shows the generated SQL and results.

✅ Semantic Views bridge business language and database schemas ✅ Relationships enable multi-table analysis automatically ✅ Search Integration connects structured SQL with unstructured search ✅ Verified Queries can be saved for future agent use ✅ Named Filters (like TICKER_SNOW) simplify common queries

Next: These semantic views become tools for the Snowflake Intelligence Agent, enabling it to answer complex questions about financial data!

Test the search services you created:

1. Navigate to AI & ML Studio → Cortex Search
2. Select EMAILS service
3. Try a search: "Tell me about companies with strong earnings"
4. Observe:
   • Semantic search results (not exact match)
   • Ranked by relevance
   • Attributes returned (TICKER, RATING, SENTIMENT)

Total: 1,400+ searchable documents across 5 services

Go to AI & ML Studio → Cortex Analyst

You should see 2 semantic views:

1. COMPANY_DATA_8_CORE_FEATURED_TICKERS
2. SNOWFLAKE_ANALYSTS_VIEW

Purpose: Query financial data and sentiment for 11 companies using natural language

Tables:

Example Questions:

Try asking in natural language:

1. "What is the revenue for SNOW in Q2 FY2025?"
2. "Show me the sentiment scores for all companies"
3. "Which companies have the highest net income?"
4. "Compare EPS across all 11 companies"

How It Works:

Purpose: Deep dive into Snowflake (SNOW ticker) data

Tables:

Named Filter:

Example Questions:

1. "What was the highest stock price for Snowflake in 2024?"
2. "Show me the price trend over the last 6 months"
3. "What did analysts say in the most recent earnings call?"
4. "Compare pre-market and post-market prices"

New Feature: STOCK_PRICES table with pivoted structure

1. Click on SNOWFLAKE_ANALYSTS_VIEW
2. Try: "What is the latest closing price of Snowflake stock?"
3. View the generated SQL
4. Observe the results

The One Ticker agent is your AI-powered financial analyst assistant that combines multiple data sources, search capabilities, and analytical tools to provide comprehensive insights about stocks and companies.

What the Agent Can Do:

This agent represents the culmination of all the previous work - it brings together the unstructured data processing, search services, semantic views, and ML models into one conversational interface.

1. Navigate to AI & ML Studio → Snowflake Intelligence in your Snowflake account
2. Select the One Ticker agent
3. The agent will automatically select the right tools based on your question

Location: SNOWFLAKE_INTELLIGENCE.AGENTS."One Ticker"

The One Ticker agent has access to 10 different tools that it automatically orchestrates based on your questions:

1. SENTIMENT_INFO_FROM_TRANSCRIPTS_AND_SUPPORTING_DATA

2. snowflake_data

3. Sentiment_Analysis

4. Analyst_Reports_Snowflake

5. Latest_infographics_8_core_stocks

6. Email_Content

7. snowflake_full_earnings_calls

8. STOCK_PERFORMANCE_PREDICTOR

9. SEND_EMAIL

10. WEB_SEARCH 🆕

Try asking:

The agent will:

Behind the scenes: Uses the STOCK_PERFORMANCE_PREDICTOR function with gradient boosting models trained on historical price data (r_1, r_5_1, r_10_5, r_21_10, r_63_21 momentum ratios).

Try asking:

The agent will:

Data sources: Searches across 92 earnings call transcripts from 11 companies with AI-extracted sentiment scores.

Try asking:

The agent will:

Demonstrates: PROP, GAME, and MKTG have minimal published materials (realistic for growth-stage companies).

Try asking:

The agent will:

Data coverage: Historical stock prices from 2020-present with daily granularity.

Try asking:

The agent will:

Research firms covered: Apex Analytics, Sterling Partners, Veridian Capital, Pinnacle Growth Investors, Momentum Metrics, Quant-Vestor.

Try asking:

The agent will:

Demonstrates: How AI agents can combine internal data with external web searches for comprehensive analysis.

Try asking:

The agent will:

Email features:

IMPORTANT: After calling SEND_EMAIL, the agent will display a clickable URL - right-click or CMD+Click to open the email in SnowMail!

Here's an example conversation demonstrating the agent's capabilities:

You: "Give me top 3 vs bottom 3 trade predictions for the next period."

Agent: Runs ML model, shows chart with top performers (e.g., VLTA, QRYQ, ICBG) and bottom performers (e.g., PROP, GAME, MKTG)

You: "What have these bottom 3 performers published themselves in terms of reports?"

Agent: Searches infographics and reports, explains PROP/GAME/MKTG have limited investor materials compared to core 8 companies

You: "Now can we focus on snowflake - are there any c level people at snowflake that are negative on earnings calls?"

Agent: Analyzes earnings transcripts by speaker, shows C-level sentiment scores

You: "I would like to see a trend visualising the SNOW stock performance over time"

Agent: Queries STOCK_PRICES with TICKER_SNOW filter, creates time-series chart showing price trends

You: "What do the other analysts say about snowflake?"

Agent: Searches analyst reports and emails, summarizes ratings distribution, shows consensus

You: "I would like to see what happened to neuro nector and how did this impact Snowflake. Present the information based on what we have - then fact check with any available data from the web."

Agent: Reviews NRNT bankruptcy timeline from internal data, searches web for external verification, shows SNOW recovery pattern

You: "Finally send me an email of your findings - as i really want to put neuro nector drama to bed!!!"

Agent: Generates comprehensive email summary, returns SnowMail URL for viewing

After using the agent, explore its configuration:

1. Navigate to AI & ML Studio → Agents
2. Click on One Ticker agent
3. Click Edit to view:
   • Sample Questions: The questions you just tried
   • Tools: All 10 tools available to the agent
   • Orchestration Instructions: How the agent decides which tools to use
   • Access Control: Who can use the agent

Core 8 (Full materials):

Bottom 3 (Limited materials):

✅ Good: "What was Snowflake's NRR in Q2 FY2025?" ❌ Too vague: "Tell me about Snowflake"

The agent automatically creates charts because of this instruction:

This is intentional for better data comprehension!

The agent maintains context, so you can:

1. Ask initial question
2. Drill deeper on specific findings
3. Request different visualizations
4. Ask for comparisons
5. Request email summary

For fact-checking or current events:

End your research session with:

You'll get a professional email with all insights accessible via SnowMail!

When querying stock prices, the agent can use the TICKER_SNOW filter to quickly focus on Snowflake data.

Synonyms supported:

This makes queries like "show me SNOW stock performance" work seamlessly.

The agent automatically:

Most responses include:

Scenario: Weekly portfolio review

Questions to ask:

1. "Give me top 3 vs bottom 3 trade predictions"
2. "Show me SNOW stock price trends for the last quarter"
3. "What's the analyst consensus on SNOW?"
4. "How do the bottom performers compare to top performers on NRR?"
5. "Send me an email summary of this week's insights"

Time saved: 2-3 hours of manual research → 5 minutes with agent

Scenario: Investigating potential red flags

Questions to ask:

1. "Which companies have negative free cash flow?"
2. "Show me sentiment from analyst earnings calls for companies with FCF < 0"
3. "What concerns do analysts mention most frequently?"
4. "Search web for recent news about these companies"
5. "Email me a risk assessment summary"

Benefit: Rapid identification of potential issues across 11 companies

Scenario: Deep dive on Snowflake

Questions to ask:

1. "What do analyst reports say about Snowflake's competitive position?"
2. "Show me earnings call sentiment trends over 3 quarters"
3. "How does SNOW compare to ICBG and QRYQ on key metrics?"
4. "Search web for latest Snowflake news"
5. "What was the Neuro-Nectar situation and how did it affect SNOW?"
6. "Send me a comprehensive Snowflake analysis email"

Output: Complete research report with internal data + external validation

The agent automatically generates visualizations - this is intentional!

The agent is:

This ensures comprehensive, validated answers.

What to do:

1. Right-click the URL
2. Select "Open in new tab" (or CMD+Click / CTRL+Click)
3. View your email in the SnowMail interface

Note: The email wasn't "sent" externally - it's saved for demo viewing.

This is a demonstration environment showing FSI AI capabilities. Do not make actual investment decisions based on this data!

In this quickstart, you've built a comprehensive AI-powered financial analysis platform using Snowflake's Cortex AI capabilities. You started by processing unstructured data—extracting insights from PDF analyst reports using Document AI, transcribing earnings call audio with AI_TRANSCRIBE, and analyzing sentiment across thousands of data points. This transformed raw, unstructured financial content into structured, queryable data that forms the foundation of intelligent applications.

You then created the AI services layer that makes this data accessible and actionable. Cortex Search services enable semantic search across analyst reports, earnings transcripts, emails, and infographics—finding relevant information based on meaning rather than keywords. Cortex Analyst semantic views allow natural language queries against structured financial data, automatically generating SQL from plain English questions. Together, these services power the intelligence layer that connects users to insights without requiring technical expertise.

Finally, you deployed and explored the Snowflake Intelligence Agent—the "One Ticker" stock analyst that orchestrates all of these capabilities through a conversational interface. The agent seamlessly combines search services, semantic views, ML predictions, and web search to answer complex financial questions, generate visualizations, and even send email summaries via SnowMail. This demonstrates how modern AI platforms can unify structured and unstructured data analysis, enabling faster decision-making and deeper insights across financial services use cases.

Ready to build AI-powered applications with Snowflake Cortex? Start experimenting with your own data today!