Optimizing Llama Snowflake Cookbook

site/sfguides/src/optimizing-llama-snowflake-cookbook/chain_of_thought_prompt_recipe

@@ -0,0 +1,12 @@
+ SELECT snowflake.cortex.complete(
+  'llama-3-7b',
+  PARSE_JSON('[
+    {"role": "system", "content": "You are a support triage assistant. For each incoming support ticket, classify the issue category, urgency (Low, Medium, High), and recommend the best team to handle it."},
+    {"role": "user", "content": "Customer reports they cannot access their account after multiple password resets."},
+    {"role": "assistant", "content": "Category: Account Access\nUrgency: High\nAssign To: Tier 1 Support"},
+    {"role": "user", "content": "User reports slow internet speeds when using our VPN service."},
+    {"role": "assistant", "content": "Category: Network Performance\nUrgency: Medium\nAssign To: Network Support Team"},
+    {"role": "user", "content": "New ticket: Application crashes when uploading files larger than 10MB."},
+    {"role": "assistant", "content": ""}
+  ]')
+) AS triage_response;

site/sfguides/src/optimizing-llama-snowflake-cookbook/cortex_budgets

@@ -0,0 +1,73 @@
+-- This script creates a stored procedure `mini_budget_on_cortex` to monitor
+-- Snowflake Cortex functions credit usage. This is a temporary solution until the native
+-- Snowflake Budget feature supports monitoring for Cortex services.
+--
+--
+-- The procedure checks usage for the current account and sends a notification
+-- if the usage exceeds a specified limit within a given billing period (day, week, or month).
+-- It also creates a task `run_mini_budget_on_cortex` to run the check every 30 minutes.
+
+
+CREATE OR REPLACE PROCEDURE mini_budget_on_cortex(
+    billing_period VARCHAR,
+    alert_limit FLOAT,
+    ni_name VARCHAR
+)
+RETURNS VARCHAR
+LANGUAGE SQL
+AS
+$$
+DECLARE
+    start_of_period       DATE;
+    INVALID_PERIOD        EXCEPTION (-20001, 'Invalid billing_period specified. Must be DAY, WEEK, or MONTH.');
+    current_usage_credits FLOAT;
+    notification_body     VARCHAR;
+BEGIN
+    CASE UPPER(billing_period)
+        WHEN 'DAY' THEN
+            start_of_period := CURRENT_DATE();
+        WHEN 'WEEK' THEN
+            start_of_period := DATE_TRUNC('week', CURRENT_DATE());
+        WHEN 'MONTH' THEN
+            start_of_period := DATE_TRUNC('month', CURRENT_DATE());
+        ELSE
+            RAISE INVALID_PERIOD;
+    END CASE;
+
+    SELECT COALESCE(SUM(TOKEN_CREDITS), 0) INTO current_usage_credits
+    FROM SNOWFLAKE.ACCOUNT_USAGE.CORTEX_FUNCTIONS_USAGE_HISTORY
+    WHERE START_TIME >= :start_of_period;
+
+    notification_body := 'Account: ' || CURRENT_ACCOUNT() || ' (' || CURRENT_REGION() || ')'
+                      || ' - Cortex functions credit usage in this ' || billing_period
+                      || ' = ' || current_usage_credits;
+    IF (current_usage_credits >= alert_limit) THEN
+        IF (ni_name IS NOT NULL AND TRIM(ni_name) <> '') THEN
+            CALL SYSTEM$SEND_SNOWFLAKE_NOTIFICATION(
+                '{"text/plain":"' || :notification_body || '"}',
+                '{"' || :ni_name || '": {}}'
+            );
+        END IF;
+    END IF;
+
+    RETURN notification_body;
+END;
+$$;
+
+
+-- Make sure the notification is sent to the correct Slack channel with a low limit.
+call mini_budget_on_cortex('MONTH', 10, 'MY_SLACK_NI');
+
+
+
+-- Create a task to run the procedure every 30 minutes.
+create or replace task run_mini_budget_on_cortex
+schedule = '30 minute'
+as
+call mini_budget_on_cortex('MONTH', 1000, 'MY_SLACK_NI');
+
+
+
+-- Make sure the task is running.
+alter task run_mini_budget_on_cortex resume;
+

site/sfguides/src/optimizing-llama-snowflake-cookbook/few_shot_prompt_recipe

@@ -0,0 +1,17 @@
+#few shot prompting
+# "Customer cannot reset password using the forgot password link."
+#"Category: Account Access Issue\nUrgency: High\nResponse: Please verify your email and try resetting your password again."
+#"Internet connection keeps dropping every few minutes, affecting video calls."
+#"Category: Network Issue\nUrgency: Medium\nResponse: Please restart your router and check for service outages in your area."
+#"Application crashes with error code 500 when uploading large files."
+#"Category: Application Bug\nUrgency: High\nResponse: Could you please provide the file size and error logs? We will check."
+
+
+-- Create the fine-tuning job pointing at your training data table
+SELECT snowflake.cortex.finetune(
+  'CREATE',
+  'MY_DB.MY_SCHEMA.CUSTOMER_SUPPORT_FINETUNED',
+  'llama3.1-8b',
+  'SELECT prompt, completion FROM MY_DB.MY_SCHEMA.CUSTOMER_SUPPORT_TRAINING',
+  NULL   -- optional validation data query here
+);

site/sfguides/src/optimizing-llama-snowflake-cookbook/llama fine-tuning sample

@@ -0,0 +1,30 @@
+import snowflake.connector
+
+conn = snowflake.connector.connect(
+    user='YOUR_USER',
+    password='YOUR_PASSWORD',
+    account='YOUR_ACCOUNT',
+    warehouse='YOUR_WAREHOUSE',
+    database='YOUR_DB',
+    schema='YOUR_SCHEMA'
+)
+
+cs = conn.cursor()
+
+cs.execute("""
+    SELECT cortex_finetune.create_finetune_job(
+        model_name => 'llama3.1-8b',
+        training_table => 'your_schema.your_training_data',
+        prompt_col => 'prompt',
+        completion_col => 'completion',
+        output_model_name => 'custom_llama_model_v1',
+        max_epochs => 3,
+        batch_size => 8,
+        learning_rate => 0.0001,
+        peft_method => 'lora'
+    );
+""")
+
+print(cs.fetchone())
+cs.close()
+conn.close()