Cortex_Analyst_Demo

Snowflake Cortex Analyst Demo in < 10 mins

Setup this demo to see the power of Conversational Analytics against structured datasets.

Notice: This demo has an additional step of summarizing the resultset via LLM calls to show art of the possible.

It is using a smaller model that can halucinate. Summarization IS NOT PART OF CORTEX ANALYST functionality & can be disabled!

CONTOSO DEMO SETUP INSTRUCTIONS:

1. Run Script Part-1 from the /Files/Setup_Script.sql to create the necessary objects

2. DRAG & DROP the 4 files(3 Sample parquet files + ContosoDemo.yaml) into ANALYST_STAGE via Snowsight before running the next set of scripts

3. Run Script Part-2 from the /Files/Setup_Script.sql to ingest the data & create search service for Fuzzy searching Customer & Product Names

4. Open the "cortex_contoso_demo_sis.py" file with a text editor & Copy all of the code.

5. Create a new StreamLit app in Snowsight & Paste the code from "cortex_contoso_demo_sis.py" file replacing all the existing code,

6. Run the Streamlit App in Snowsight

SAMPLE TALK TRACK:

1. SHOW ME TOP 15 PRODUCTS SOLD IN TX IN TERMS OF REVENUE

   • "TX" does not match the data in the table where the real value is TEXAS. Show the query where it is filtering on TEXAS
   • Because we didnt specified a time range, LLM will use the entire time time period for revenue
   • Notice how results can be displayed as Table, Charts but also as a business summary & insights using CORTEX.COMPLETE() function via Mistral_Large2
   • It drives insights using product descriptions such as Brand or Product Type even if those are not actual data columns

2. WHAT ABOUT IN 2009?

   • Notice this is a follow up to first question where Analyst will apply the YEAR=2009 to the previous question

3. HOW ABOUT IN FL?

   • Another follow up where we are switching to Sales in Florida during 2009

4. Sales revenue for product categories & sub catergories sold in FL in 2009 & YOY % growth?

   • This is a pretty complex question.
   • Notice how YOY% was interpreted by LLM as Year-over-Year & it figured out to compare to 2008

5. Who are the top 5 customers who purchased cameras in FL

6. How much did Rachel Gordon spent in different product categories in Florida during 2009?

   • Notice the name format in the table is "Gordon, Rachel". CORTEX_SEARCH() Hybrid(Vector+Text) service was able to identify the correct name.
   • Look at the SQL: WHERE NAME = 'Gordon, Rachel'. Query is NOT using LIKE + % to find a partial match but uses Vector search to locate the actual name

7. How much did Rachal Gordn spent in different product categories in Florida during 2009? (Notice the bad spelling for Rachel Gordon!)

   • We just butchered both the first & last names
   • Notice the SQL WHERE NAME = 'Gordon, Rachel'.
   • Cortex_Search service allows users to make mistakes via Fuzzy Vector Search & not have to be precise with their search criteria & still get proper results

8. Open & View the Yaml file.

   • Explain Symantic Generator auto splits columns in ach table into Dimensions, Measures, Time_Dimensions
   • Each Column has Synonyms to allow additional business terminology to be recognized
   • Look at the JOINS section
   • Look at thge UNITPRICE measure in SALES table where the description was edited to tell Analyst how to properly calculate revenue W/o the additional description, Revenue questions were using NETPRICE column assuming it contained total sales amount but the data in table is wrong and actual calcualtion should be as follows: description: Unit Price of a Product sold where Sales Revenue = UNITPRICE * QUANTITY This is how easy it is to fix potential problems