AI and Analytics for Data Teams: What It Is, What’s Available, and How to Get Started

Leadership is asking for AI-powered reports and analytics. Business analysts are being asked to deliver insights faster. Data warehouse teams and data engineers are evaluating new platforms. DBAs are fielding questions about where the database fits in an AI strategy. Everyone is hearing the same words — vectors, embeddings, semantic search, RAG, Copilot — and not everyone has had the time to understand what they actually mean or what options are available.

This article is for all of those people. It starts with plain-language explanations that anyone on the team can read and understand. It covers the full landscape of options — Microsoft, Snowflake, Google, AWS, Databricks, and standalone tools — honestly and without assuming you have already picked a platform. And for the technical members of the team, it goes all the way through to working code examples. Every section is labeled so you can read what is relevant to your role and skip what is not.

1 Why Leadership Is Asking About AI and Analytics Right Now Everyone

For most organizations, data has always been abundant and insight has been scarce. You had the transactions, the orders, the customer records, the support tickets — gigabytes or terabytes of it sitting in databases and data warehouses. But turning that data into answers took time. An analyst would write a report, a developer would build a dashboard, and by the time the insight reached the decision-maker it was days or weeks old and answered a specific pre-defined question.

AI changes that relationship in two fundamental ways.

First, AI makes it possible to ask questions that were not pre-defined. Instead of a fixed report that shows sales by region for last quarter, an executive can ask “Why did the Northeast underperform in Q1 compared to the same period last year, and which product categories drove the gap?” and get an answer from the data without a developer building a new report first. This is what leadership means when they ask for AI-powered analytics.

Second, AI makes it possible to find things in your data that keyword search cannot. A customer types “I can’t get online” into your support system. A keyword search might miss that ticket unless the database contains exactly those words. Semantic AI search understands the meaning — it connects “can’t get online” to “internet connectivity issue,” “network not working,” and “WiFi not connecting” and surfaces all related tickets regardless of exact wording. This enables a quality of analysis that was simply not possible before.

2 What AI in Analytics Actually Means — Plain Language Everyone

When people say “AI analytics” they are usually describing one or more of these four things. Understanding which one you are dealing with matters because each has different implications for your data platform, your team, and your budget.

1. Natural Language to Data (Text-to-SQL)

A non-technical person types a question in plain English — “Show me the top ten customers by revenue this year” — and the system generates the correct query, runs it against your database, and returns the result as a chart or table. No SQL knowledge required. This is what Copilot in Power BI, Snowflake Cortex Analyst, and similar tools do. The business value is immediate: analysts and executives can explore data without waiting for a developer to write a report.

2. Semantic Search

Instead of finding records that match exact keywords, semantic search finds records that match the meaning of a query. Search for “affordable accommodation near the city center” and get results that include “budget hotel downtown,” “cheap apartment in the CBD,” and “low-cost hostel central location” — none of which contain your exact words. This is used in product catalogs, support knowledge bases, document retrieval, and any scenario where the data is text and exact matching is too rigid.

3. AI-Powered Classification and Enrichment

Run an AI model against your existing data to add new information automatically. Label customer support tickets as positive, negative, or neutral. Identify product categories from free-text descriptions. Extract entities from unstructured documents. Flag potentially fraudulent transactions. The AI reads each row, applies its understanding, and writes the result back — turning unstructured data into structured, queryable insights.

4. Retrieval Augmented Generation (RAG)

Connect a large language model to your organization’s private data so it can answer questions with accurate, current, specific information — not just general knowledge. Ask a chatbot “What is our return policy for electronics purchased online?” and it retrieves the relevant section of your actual policy document and uses it to generate an accurate answer. This is the foundation of enterprise AI assistants built on your own data.

3 The Core Concepts Explained Simply Everyone

These terms appear constantly in AI analytics conversations. Here they are in plain language — no mathematics required.

Vectors and Embeddings

An AI model can convert any piece of text, image, or data into a list of numbers that represents its meaning. That list of numbers is called an embedding (or a vector). Two pieces of text with similar meaning produce similar lists of numbers. Two pieces of text with unrelated meaning produce very different lists. The database stores these lists of numbers and can find similar ones very quickly — which is how semantic search works.

Think of it like a map. Every piece of content in your database gets a location on a very large map based on its meaning. When you search for something, you find your own location on that map and then find everything nearby — everything with a similar meaning — regardless of the exact words used.

Large Language Models (LLMs)

LLMs are AI systems trained on enormous amounts of text — billions of documents — that can understand and generate human language at a high level. GPT-4, Claude, Llama, and Gemini are examples. They can write, summarize, classify, explain, and answer questions. They do not have access to your private data unless you explicitly provide it to them as part of a query — which is what RAG does.

RAG — Retrieval Augmented Generation

RAG is the pattern that connects your organization’s data to an LLM. When someone asks a question, the system first searches your database for relevant information (retrieval), then provides that information to the LLM along with the question (augmentation), and the LLM uses it to generate an accurate answer (generation). The LLM does not need to have memorized your data — it just needs the relevant section retrieved and handed to it at the moment of the question.

The SQL Connection

All of the major data platforms are now building these AI capabilities directly into their database engines. This means your existing data — the data already in your SQL Server, Snowflake, BigQuery, or Redshift — can become the foundation of AI analytics without copying it to a separate AI system first. Your security policies, access controls, and governance rules all apply automatically because the AI is operating inside the database, not outside it.

4 How AI Changes What Reports and Analytics Can Do Everyone

Traditional analytics answers questions you knew to ask. You defined the dimensions, the metrics, and the filters ahead of time. An analyst or developer built the report. Leadership read the output. The gap between question and answer was measured in days or weeks.

AI-enabled analytics changes this in three practical ways that affect every role in the data team:

5 The Major Platforms and What Each One Offers Leadership / Architects

Every major data platform has launched or is actively building AI analytics capabilities. The right choice for your organization depends on where your data already lives, what your team already knows, your cloud strategy, and your budget. There is no universally correct answer — but there are clear patterns that make one platform a better fit than another for specific situations.

6 Microsoft: SQL Server 2025, Azure SQL, and Fabric Everyone

Microsoft’s AI analytics story in 2026 is built around three connected products. SQL Server 2025 brings AI natively into the database engine for on-premises and cloud VM environments. Azure SQL Database is the fully managed cloud version with the same AI capabilities, always up to date. Microsoft Fabric is the unified analytics platform that sits above both — connecting operational data to analytics, AI, and Power BI in one environment.

What Is New in SQL Server 2025

SQL Server 2025 is the most significant SQL Server release in a decade. The headline addition is native AI built directly into T-SQL — not a bolt-on, not an external service, but SQL functions that generate embeddings, perform semantic search, and call AI models using the same language DBAs and developers already know. Key additions include a native VECTOR data type for storing AI-generated embeddings, VECTOR_DISTANCE for semantic similarity queries, CREATE EXTERNAL MODEL to connect SQL Server to Azure OpenAI or any compatible AI model, AI_GENERATE_EMBEDDINGS to generate embeddings directly in T-SQL, and Change Event Streaming to feed real-time data changes into analytics pipelines.

Microsoft Fabric

Microsoft Fabric is a single platform that brings together data engineering, data warehousing, real-time analytics, Power BI, and AI. It stores everything in OneLake — a shared data lake — so data does not have to be copied between services. Fabric Mirroring continuously replicates your SQL Server databases into Fabric within seconds, enabling near-real-time Power BI reports and AI analytics on operational data without ETL batch jobs. As of March 2026, Fabric serves more than 31,000 customers and is the fastest-growing data platform in Microsoft’s history.

Copilot Across the Microsoft Stack

Copilot is Microsoft’s AI assistant integration. In Power BI it generates reports and visuals from natural language descriptions. In SSMS 21 it generates T-SQL from plain English questions about your data. In the Azure Portal it helps manage and query Azure SQL databases. In Fabric it assists across all workloads — notebooks, pipelines, reports, and more.

7 Snowflake Cortex AI Everyone

Snowflake Cortex AI is Snowflake’s fully managed AI and machine learning capability built directly into the Snowflake platform. It requires no infrastructure management, no model deployment, and no data movement — AI functions run against the data already in your Snowflake tables using standard SQL.

What Cortex AI Includes

Cortex LLM Functions — a set of SQL functions that call large language models directly from Snowflake queries. SNOWFLAKE.CORTEX.COMPLETE calls an LLM for text generation. SNOWFLAKE.CORTEX.SENTIMENT classifies text as positive, negative, or neutral. SNOWFLAKE.CORTEX.SUMMARIZE generates a concise summary of a long text field. SNOWFLAKE.CORTEX.TRANSLATE translates text between languages. SNOWFLAKE.CORTEX.EXTRACT_ANSWER answers a question based on a provided document. All of these run as SQL functions — no Python, no external API calls, no configuration.

Cortex Search — hybrid search that combines keyword and semantic search over your Snowflake data. Point it at a table, define which columns to search, and Cortex Search handles the embedding generation, indexing, and query execution automatically. Accessible via SQL or a REST API for application developers.

Cortex Analyst — a natural language interface for business analysts. Connect it to your Snowflake data and semantic model (a YAML file describing your business metrics and relationships), and users can ask questions in plain English through a chat interface. Cortex Analyst generates the SQL, runs it, and returns results. No SQL knowledge required from the end user.

Arctic and Partner Models — Snowflake hosts its own open-source LLM called Arctic, optimized for enterprise SQL generation. It also provides access to models from Mistral, Meta Llama, Google, and others through the Cortex functions, giving you model choice without managing infrastructure.

8 Databricks with Unity Catalog and Mosaic AI Intermediate

Databricks is a data intelligence platform built on Apache Spark with strong open-source foundations. Its AI capabilities are grouped under the Mosaic AI brand, and its governance layer is Unity Catalog — a unified catalog that manages data, models, and AI assets in one place.

What Databricks Mosaic AI Includes

Model Serving — deploy and serve AI models at production scale, with auto-scaling and monitoring. Supports open-source models (Llama, Mistral) as well as external models like Azure OpenAI and Anthropic.

Vector Search — a managed vector database built into Databricks. Create vector indexes on Delta Lake tables, run similarity queries, and build RAG pipelines — all within the same platform as your data engineering and data science workloads.

AI Playground and Prompt Engineering — experiment with models and prompts directly in the Databricks UI without writing application code. Useful for rapid prototyping before committing to a production implementation.

Unity Catalog for AI Governance — models, feature tables, and AI assets are governed by the same Unity Catalog that manages data access, lineage, and auditing. This is Databricks’ key differentiator — unified governance across data and AI.

9 Google BigQuery ML and Vertex AI Intermediate

Google’s AI analytics strategy combines BigQuery ML — machine learning directly in SQL — with Vertex AI, Google’s managed machine learning platform, and Gemini, Google’s large language model family.

BigQuery ML

BigQuery ML allows you to create and run ML models using standard SQL CREATE MODEL statements directly in BigQuery. Supported model types include linear and logistic regression, k-means clustering, time series forecasting with ARIMA_PLUS, deep neural networks, and imported TensorFlow or scikit-learn models. You can also call Vertex AI hosted models and Google’s embedding APIs directly from BigQuery SQL using ML.GENERATE_TEXT and ML.GENERATE_EMBEDDING functions — similar in concept to SQL Server 2025’s AI_GENERATE_EMBEDDINGS.

Gemini in BigQuery and Looker

Google’s Gemini model is integrated across the analytics stack. In BigQuery it assists with query writing and explanation. In Looker (Google’s BI platform) it enables natural language questions against your semantic model — equivalent to Microsoft Copilot in Power BI or Snowflake Cortex Analyst for business users.

10 AWS: Redshift ML and Amazon Bedrock Intermediate

AWS provides AI analytics capabilities through Redshift ML for in-database machine learning and Amazon Bedrock for LLM access, with Amazon SageMaker as the underlying ML platform.

Redshift ML

Redshift ML uses CREATE MODEL SQL statements to train machine learning models on data in Amazon Redshift, using Amazon SageMaker Autopilot to handle the training and optimization automatically. Once trained, models are exposed as SQL functions you can call in queries. This follows the same pattern as BigQuery ML — ML in SQL — but is more tightly integrated with the broader AWS SageMaker ecosystem for teams that already use it.

Amazon Bedrock

Bedrock is AWS’s managed service for accessing large language models from multiple providers — Anthropic Claude, Meta Llama, Mistral, Amazon Titan, and others — through a single API. Data engineers and developers can call Bedrock models from applications, Lambda functions, or SageMaker notebooks to build RAG patterns on top of Redshift or RDS data.

11 Standalone Vector Databases — Pinecone, Weaviate, Milvus, pgvector Intermediate

Before the major database platforms added native vector capabilities, organizations building AI applications had to use purpose-built vector databases. These tools still have a role, though the case for using them standalone has narrowed significantly now that SQL Server 2025, Snowflake, BigQuery, and Databricks all support vectors natively.

The Main Options

Pinecone — fully managed cloud vector database. Extremely fast approximate nearest-neighbor search, no infrastructure to manage, simple API. Best for applications that need dedicated high-performance vector search and are willing to pay for a separate managed service. Does not store your relational data — you still need your primary database alongside it.

Weaviate — open-source vector database that can be self-hosted or used as a managed service. Supports hybrid search combining vector and keyword results. Includes built-in modules for automatic vectorization using popular embedding models.

Milvus — open-source vector database designed for very large scale — billions of vectors. More complex to operate than Pinecone but no per-query costs. Used by organizations that need extreme scale with full control over infrastructure.

pgvector — an open-source extension for PostgreSQL that adds vector storage and similarity search. If your organization runs PostgreSQL, pgvector adds vector search without a new database system. Performance at very large scale is more limited than dedicated vector databases but entirely adequate for most enterprise use cases.

12 How to Choose — A Framework for Your Organization Leadership / Architects

Platform selection for AI analytics is not primarily a technical decision — it is an organizational one. The best platform is almost always the one your team already knows and your data already lives in. Migration cost, retraining cost, and integration complexity almost always outweigh the differences between mature platforms in the same tier.

Questions to Ask Before Committing

• Where does the data already live? Moving data to a new platform to access AI features is rarely worth the cost if your existing platform offers comparable capabilities.
• What does your team already know? A technically superior platform that your team cannot operate is not actually superior in practice.
• What does leadership need from AI analytics — self-service reporting, semantic search, customer-facing AI features, or internal knowledge management? The answer changes the optimal tool.
• What are the governance requirements? Regulated industries need to know where data goes when an AI model processes it and who has access to what. All the major platforms have answered this — but the answers differ.
• What is the total cost? LLM calls are metered per token. Embedding generation has a cost. Vector index storage has a cost. Model the usage before committing to a cost estimate.

13 What Changes for Report Teams and Business Analysts Analysts / Report Teams

The most immediate impact of AI analytics for report teams and business analysts is the shift from building reports to enabling questions. Instead of a business user submitting a request and waiting for a developer to build a new query or dashboard, an AI-enabled analytics layer allows them to ask questions directly and get answers from the data in real time.

This does not eliminate the report developer’s role — it changes it. Someone still needs to build and maintain the semantic layer that defines what “revenue” means, what the right date hierarchies are, which tables relate to which, and what the business rules are. That semantic layer is what enables the natural language query tools to generate correct SQL rather than plausible-sounding but wrong SQL. Building and governing the semantic layer becomes one of the most valuable things a report team can do.

What Business Analysts Gain

• Ability to ask new ad-hoc questions without waiting for a developer to build a query
• Semantic search over unstructured data — product descriptions, customer comments, support tickets — that keyword search could not previously reach
• Automatic classification and tagging of unstructured content, reducing manual categorization work
• Summarization of large text datasets — “What are the five most common themes in this year’s customer feedback?” answered in seconds

What Business Analysts Need to Understand

• Natural language query tools generate SQL — they do not always generate correct SQL. Results should be validated, especially for financial or compliance reporting.
• AI-generated answers are only as good as the data and the semantic model behind them. Garbage in, garbage out still applies.
• Semantic search returns relevant results, not exact results. For compliance or audit scenarios where you need complete and exact records, traditional structured queries are still appropriate.

14 Copilot in Power BI and Fabric — Natural Language to Reports Analysts / Report Teams

Microsoft Copilot is integrated across Power BI and Microsoft Fabric, providing AI assistance at every step of the analytics workflow. For business analysts and report teams, the most immediately useful capabilities are natural language report generation and the ability to ask questions about data without writing DAX or SQL.

What Copilot in Power BI Can Do

• Generate report pages from a description. Type “Create a sales performance report showing revenue by region, product category, and month with a comparison to last year” and Copilot builds the page with appropriate visuals, selecting measures and dimensions from your semantic model.
• Answer questions about your data in natural language. Ask “Which product category had the highest growth rate in Q1 2026?” and Copilot queries your model and returns an answer with the supporting visual.
• Summarize report insights. Generate a plain-language executive summary of what the data in a report page is showing — useful for stakeholder communications where not everyone will read the full dashboard.
• Explain visuals. Ask “Why is this bar chart showing a dip in February?” and Copilot analyzes the data to provide a narrative explanation.

15 Snowflake Cortex Analyst — Asking Questions in Plain English Analysts / Report Teams

Snowflake Cortex Analyst is Snowflake’s natural language to SQL interface for business users. You provide a semantic model — a YAML file that describes your business metrics, dimensions, and the relationships between tables — and Cortex Analyst uses it to convert plain English questions into accurate SQL queries against your Snowflake data.

The semantic model definition is the critical piece. A good semantic model defines revenue with the exact calculation your business uses, specifies that “customer” and “client” mean the same thing, and explains that orders are tied to customers through the order_customer_id foreign key. With that context, Cortex Analyst can answer “What was our revenue last quarter broken down by customer segment?” with a query that will return the right answer.

# Example Snowflake Cortex Analyst semantic model (simplified YAML)
# This is what a report developer creates to enable natural language queries

name: sales_analytics
description: Sales performance metrics for business analysis

tables:
  - name: fact_orders
    description: All customer orders
    base_table: SALES_DW.PUBLIC.FACT_ORDERS
    measures:
      - name: total_revenue
        description: Total order revenue after discounts
        expr: SUM(order_amount - discount_amount)
        data_type: number
      - name: order_count
        description: Number of orders placed
        expr: COUNT(DISTINCT order_id)
        data_type: number
    dimensions:
      - name: order_date
        description: Date the order was placed
        expr: order_date
        data_type: date
      - name: customer_segment
        description: Customer tier - Enterprise, Mid-Market, or SMB
        expr: c.segment
        data_type: text

# Once this model is deployed, business users can ask:
# "What was our total revenue last quarter by customer segment?"
# "Which product had the highest order count in January?"
# "Show me month-over-month revenue growth for the past year"
# Cortex Analyst generates the SQL, runs it, returns results

16 Microsoft Fabric Mirroring and Real-Time Analytics DBAs / Data Engineers

Microsoft Fabric Mirroring solves one of the longest-standing problems in analytics — keeping your reporting and analytics environment current with operational data without complex ETL pipelines and batch windows. It continuously replicates changes from your SQL Server 2025 databases into Fabric’s OneLake within seconds using the transaction log change feed.

Once data is in OneLake, every Fabric workload can use it: Power BI Direct Lake reports that reflect near-real-time operational state, Spark notebooks for data engineering and ML, the Fabric Data Warehouse for SQL analytics, and AI workloads including vector search and RAG patterns. All without copying data between services — OneLake is the single store that all workloads share.

Change Event Streaming for Real-Time AI Pipelines

Change Event Streaming (CES) is a newer capability that goes further than Mirroring for event-driven architectures. CES streams individual row-level changes from the SQL Server transaction log to Azure Event Hubs or Fabric Eventstream as they happen. This enables architectures where a product catalog change triggers an immediate re-embedding of the new description, updating the semantic search index in real time — no batch job, no overnight delay.

-- Confirm SQL Server 2025 is Arc-enabled (required for Fabric Mirroring on-premises)
-- Run on the SQL Server instance
SELECT SERVERPROPERTY('IsArcEnabled') AS [Arc Enabled];

-- Check recovery model (required: FULL)
SELECT name, recovery_model_desc
FROM sys.databases
WHERE name = 'YourDatabase';

-- Fabric Mirroring is configured in the Fabric portal, not in SQL Server T-SQL
-- Once mirrored, connect SSMS to the Fabric SQL Analytics Endpoint:
-- Server: yourworkspace-xxxx.sql.fabric.microsoft.com
-- Database: YourDatabase (the mirrored copy in Fabric)
-- Then query as normal T-SQL for analytics workloads

17 SQL Server 2025: Vector Search in T-SQL — Step by Step Advanced / DBAs

This is the complete implementation for semantic search using SQL Server 2025’s native vector capabilities. The pattern works for product catalogs, knowledge bases, support articles, or any text data where meaning-based search is more useful than keyword matching.

-- ============================================================
-- End-to-end semantic search in SQL Server 2025
-- Prerequisites: SQL Server 2025, Azure OpenAI resource
-- ============================================================

-- Step 1: Enable preview features (for DiskANN vector index)
ALTER DATABASE SCOPED CONFIGURATION SET PREVIEW_FEATURES = ON;
GO

-- Step 2: Create credential and external model
CREATE MASTER KEY ENCRYPTION BY PASSWORD = 'YourStrongPassword!';
GO

CREATE DATABASE SCOPED CREDENTIAL [AzureOpenAICred]
WITH IDENTITY = 'HTTPEndpointHeaders',
SECRET = '{"api-key": "your-azure-openai-api-key"}';
GO

CREATE EXTERNAL MODEL [EmbeddingModel]
WITH (
    LOCATION   = 'https://your-resource.openai.azure.com/openai/deployments/text-embedding-ada-002',
    API_FORMAT = 'Azure_OpenAI',
    MODEL_TYPE = EMBEDDINGS,
    CREDENTIAL = [AzureOpenAICred]
);
GO

-- Step 3: Create table with VECTOR column
-- Dimension 1536 matches text-embedding-ada-002 output
CREATE TABLE dbo.KnowledgeBase (
    ArticleId   INT IDENTITY PRIMARY KEY,
    Title       NVARCHAR(200) NOT NULL,
    Content     NVARCHAR(MAX),
    Category    NVARCHAR(50),
    IsPublished BIT DEFAULT 1,
    ContentVec  VECTOR(1536)
);
GO

-- Step 4: Insert content (embeddings generated in next step)
INSERT INTO dbo.KnowledgeBase (Title, Content, Category)
VALUES
('How to reset your password', 'To reset your password click Forgot Password on the login page...', 'Account'),
('Wifi connection troubleshooting', 'If you cannot connect to wifi check that your router is powered on...', 'Network'),
('Return and refund policy', 'Items may be returned within 30 days of purchase with original receipt...', 'Policy'),
('Setting up two-factor authentication', 'Two-factor authentication adds an extra layer of security to your account...', 'Security');
GO

-- Step 5: Generate embeddings for all articles
UPDATE dbo.KnowledgeBase
SET ContentVec = AI_GENERATE_EMBEDDINGS(Content USE MODEL EmbeddingModel)
WHERE ContentVec IS NULL;
GO

-- Step 6: Create DiskANN vector index for fast approximate search
CREATE VECTOR INDEX IX_KB_ContentVec
ON dbo.KnowledgeBase (ContentVec)
WITH (METRIC = 'cosine', TYPE = 'diskann');
GO

-- Step 7: Semantic search -- find articles relevant to a user query
-- The query "I forgot how to log in" should surface password reset article
-- even though it does not contain those exact words
DECLARE @UserQuery VECTOR(1536) =
    AI_GENERATE_EMBEDDINGS(N'I forgot how to log in' USE MODEL EmbeddingModel);

SELECT TOP(3)
    ArticleId, Title, Category,
    VECTOR_DISTANCE('cosine', @UserQuery, ContentVec) AS Relevance
FROM dbo.KnowledgeBase
WHERE IsPublished = 1
ORDER BY Relevance ASC;
-- Expected: password reset article ranks #1 despite no keyword overlap

-- Step 8: Hybrid search -- semantic + relational filters
DECLARE @NetQuery VECTOR(1536) =
    AI_GENERATE_EMBEDDINGS(N'cannot get online wireless' USE MODEL EmbeddingModel);

SELECT TOP(3)
    ArticleId, Title, Category,
    VECTOR_DISTANCE('cosine', @NetQuery, ContentVec) AS Relevance
FROM dbo.KnowledgeBase
WHERE Category = 'Network'           -- relational filter
AND IsPublished = 1
ORDER BY Relevance ASC;

18 Snowflake Cortex AI in SQL — Step by Step Advanced / Data Engineers

Snowflake Cortex AI functions run as standard SQL — no infrastructure to configure, no API keys to manage. The functions call Snowflake-hosted models with the costs billed to your Snowflake account per token used.

-- ============================================================
-- Snowflake Cortex AI functions -- run in Snowflake SQL
-- No setup required beyond standard Snowflake access
-- ============================================================

-- Sentiment analysis on customer reviews
SELECT
    review_id,
    customer_name,
    review_text,
    SNOWFLAKE.CORTEX.SENTIMENT(review_text)       AS sentiment_score,
    -- Returns -1 (very negative) to 1 (very positive)
    CASE
        WHEN SNOWFLAKE.CORTEX.SENTIMENT(review_text) > 0.3  THEN 'Positive'
        WHEN SNOWFLAKE.CORTEX.SENTIMENT(review_text) < -0.3 THEN 'Negative'
        ELSE 'Neutral'
    END                                           AS sentiment_label
FROM customer_reviews
WHERE review_date >= DATEADD('day', -30, CURRENT_DATE());

-- Summarize long customer support tickets
SELECT
    ticket_id,
    SNOWFLAKE.CORTEX.SUMMARIZE(ticket_description) AS summary
FROM support_tickets
WHERE LENGTH(ticket_description) > 500
AND created_date >= DATEADD('day', -7, CURRENT_DATE());

-- Extract answers from documents using LLM
SELECT
    doc_id,
    SNOWFLAKE.CORTEX.EXTRACT_ANSWER(
        document_text,
        'What is the warranty period for this product?'
    ) AS warranty_answer
FROM product_documents
WHERE product_category = 'Electronics';

-- Generate embeddings for semantic search (Cortex Search handles this automatically
-- but you can also generate manually for custom use cases)
SELECT
    article_id,
    SNOWFLAKE.CORTEX.EMBED_TEXT_1024(
        'snowflake-arctic-embed-m',   -- Snowflake's own embedding model
        article_content
    ) AS content_embedding
FROM knowledge_base_articles;

-- Translate content for multilingual analytics
SELECT
    feedback_id,
    original_language,
    feedback_text,
    SNOWFLAKE.CORTEX.TRANSLATE(
        feedback_text,
        'auto',   -- detect source language automatically
        'en'      -- translate to English
    ) AS feedback_english
FROM customer_feedback
WHERE original_language != 'en';

-- Use COMPLETE for custom prompts and classification
SELECT
    product_id,
    description,
    SNOWFLAKE.CORTEX.COMPLETE(
        'mistral-7b',    -- model choice
        CONCAT(
            'Classify this product description into one of these categories: ',
            'Electronics, Clothing, Home & Garden, Sports, Food & Beverage. ',
            'Return only the category name. Description: ',
            description
        )
    ) AS predicted_category
FROM unclassified_products
LIMIT 100;

19 Building a RAG Pattern — Grounding AI in Your Own Data Advanced / Data Engineers

RAG is the architecture that makes enterprise AI assistants actually useful. Without it, an LLM answers questions based only on its training data — it knows nothing about your organization’s policies, products, or procedures. With RAG, every answer is grounded in your actual data, retrieved at query time from your database.

The pattern is the same regardless of platform: embed your knowledge base, retrieve the most relevant context for each question using vector search, provide that context to the LLM alongside the question, get a grounded answer back. The implementation differs by platform.

-- ============================================================
-- RAG implementation in SQL Server 2025 T-SQL
-- A stored procedure that answers questions using your knowledge base
-- ============================================================
CREATE OR ALTER PROCEDURE dbo.AnswerFromKnowledgeBase
    @Question NVARCHAR(2000),
    @Answer   NVARCHAR(MAX) OUTPUT
AS
BEGIN
    SET NOCOUNT ON;

    -- Step 1: Embed the user's question
    DECLARE @QuestionVec VECTOR(1536) =
        AI_GENERATE_EMBEDDINGS(@Question USE MODEL EmbeddingModel);

    -- Step 2: Retrieve top 5 most relevant articles
    DECLARE @Context NVARCHAR(MAX) = '';
    SELECT @Context = @Context + CHAR(13) + Title + ': ' + Content + CHAR(13)
    FROM (
        SELECT TOP(5) Title, Content,
               VECTOR_DISTANCE('cosine', @QuestionVec, ContentVec) AS Dist
        FROM dbo.KnowledgeBase
        WHERE IsPublished = 1
        ORDER BY Dist ASC
    ) AS ctx;

    -- Step 3: Call GPT-4 with the retrieved context
    DECLARE @Payload NVARCHAR(MAX) = CONCAT(
        N'{"messages":[',
        N'{"role":"system","content":"Answer using only the provided context. If the answer is not in the context say so."},',
        N'{"role":"user","content":"Context:\n', REPLACE(@Context, '"', '\"'),
        N'\n\nQuestion: ', REPLACE(@Question, '"', '\"'), N'"}',
        N'],"max_tokens":400}'
    );

    DECLARE @Response NVARCHAR(MAX);
    DECLARE @Status   INT;

    EXEC sp_invoke_external_rest_endpoint
        @url         = N'https://your-resource.openai.azure.com/openai/deployments/gpt-4/chat/completions?api-version=2024-02-01',
        @method      = 'POST',
        @headers     = N'{"api-key":"your-key","Content-Type":"application/json"}',
        @payload     = @Payload,
        @response    = @Response OUTPUT,
        @status_code = @Status OUTPUT;

    SET @Answer = JSON_VALUE(@Response, '$.choices[0].message.content');
END;
GO

-- Test the RAG procedure
DECLARE @Result NVARCHAR(MAX);
EXEC dbo.AnswerFromKnowledgeBase
    @Question = 'How do I get a refund for a product I bought last week?',
    @Answer   = @Result OUTPUT;
SELECT @Result AS [Answer];

-- ============================================================
-- RAG pattern in Snowflake using Cortex Search + COMPLETE
-- Simpler implementation -- Cortex Search handles embedding automatically
-- ============================================================

-- First: create a Cortex Search service on your knowledge base table
-- (done once, in Snowflake UI or SQL)
CREATE OR REPLACE CORTEX SEARCH SERVICE knowledge_search
    ON content                          -- column to search
    ATTRIBUTES title, category          -- filterable metadata
    WAREHOUSE = compute_wh
    TARGET_LAG = '1 minute'            -- how often to refresh the index
AS (
    SELECT title, content, category
    FROM knowledge_base_articles
    WHERE is_published = TRUE
);

-- Then: build a RAG query using Cortex Search for retrieval
-- and COMPLETE for generation
SELECT
    SNOWFLAKE.CORTEX.COMPLETE(
        'mistral-large',
        CONCAT(
            'Answer this question using ONLY the context below. ',
            'If the answer is not in the context say you do not know.',
            CHR(10), 'Context:', CHR(10),
            -- Cortex Search returns JSON with matching results
            -- In a real implementation this is called via the Search API
            -- and the results are passed here as context
            retrieved_context,
            CHR(10), 'Question: ', @user_question
        )
    ) AS ai_answer
FROM retrieved_context_table;

20 What Is Coming Next Across All Platforms Everyone

The pace of change in AI analytics is faster than any other area of the data platform landscape. Here is what is confirmed or strongly indicated across the major platforms heading into the rest of 2026:

Microsoft

• VECTOR_SEARCH and DiskANN vector index moving from preview to GA in a near-term SQL Server 2025 cumulative update
• SSMS 22 replacing Azure OpenAI endpoint requirement with GitHub Copilot account for Copilot in SSMS
• Fabric migration assistant for SQL Server — AI-assisted migration from on-premises SQL Server into Microsoft Fabric
• dbt Fusion support in Fabric expected Q2 2026
• Fabric IQ — natural language analytics across all Fabric workloads powered by semantic models
• GPU acceleration for AI inference workloads in SQL Server 2025 on supported hardware

Snowflake

• Cortex Analyst expanding model choice and semantic model capabilities
• Snowflake Arctic model family updates for improved SQL generation accuracy
• Deeper integration between Cortex Search and Cortex Analyst for end-to-end RAG pipelines in SQL

Industry-Wide

• Every major platform is investing heavily in natural language to SQL — the quality of AI-generated queries will continue to improve as semantic models become richer and models become more capable
• Agentic AI — AI systems that take multi-step actions against databases rather than answering a single question — is the next frontier. Microsoft, Snowflake, and Databricks are all building agentic capabilities into their platforms
• Multimodal data — queries that combine text, images, and structured data in a single search — is moving from research into early production availability
• Open-source model quality continues to close the gap with commercial models. The ability to run capable AI models on-premises using Ollama or similar tools matters significantly for regulated industries and organizations with strict data residency requirements

References

• Microsoft Docs — Intelligent Applications and AI in SQL Server 2025
• Microsoft Docs — Vector Search and Vector Index in SQL Server 2025
• Microsoft SQL Server Blog — Announcing SQL Server 2025
• Azure SQL Dev Corner — SQL Server 2025 Vector and AI Features
• Microsoft Docs — What is Microsoft Fabric?
• Microsoft Azure Blog — FabCon and SQLCon 2026 Announcements
• Microsoft Fabric Blog — Change Event Streaming to Fabric Eventstream
• Snowflake Docs — Cortex Analyst
• Snowflake Docs — Cortex LLM Functions (COMPLETE, SENTIMENT, SUMMARIZE, TRANSLATE)
• Snowflake Docs — Cortex Search
• Databricks — Mosaic AI and Machine Learning
• Google Cloud Docs — BigQuery ML Introduction
• AWS Docs — Amazon Redshift ML
• AWS — Amazon Bedrock
• Pinecone Docs — Getting Started
• Weaviate Docs — Overview
• GitHub — pgvector: Open-Source Vector Similarity Search for PostgreSQL
• LangChain Docs — SQL Server Vector Store Integration
• Microsoft Docs — Semantic Kernel Overview
• SQLYARD — SQL Server Always On Availability Groups: The Complete DBA Guide