Snowflake Cortex¶

The database role required to access LLM function is snowflake.cortex_user which is granted to PUBLIC by default. If this is not desired, this privilege should be explicitly revoked
Certain LLMs may not be available in all regions; there is no egress cost for cross region LLM usage, but must be enabled

Cortex Analyst¶

allows LLM based search based on a semantic model
supports multi-turn conversation (asking follow-up questions)
- must include previous messages, since LLMs are stateless
Permissions needed: snowflake.cortex_user role, usage on the model (YAML), select on the tables
cost is based on the LLM model used, appears in snowflake.account_usage.cortex_analyst_usage_history
UDTF SNOWFLAKE.LOCAL.CORTEX_ANALYST_REQUESTS provides logs of analyst's questions and answers
LLM Model can be changed, but at the account level (Ref Sarathi @ PS Tech Summit)

Logical Table
- Dimension: categorical data that provides context to facts, such as product, customer, location
  - a time_dimension provides temporal context, enables tracking metrics over specific time interval
- Fact: are measurable, quantitative data
- Metric: a quantifiable measure that is calculated by aggregating multiple fact rows
- Filter: named condition that limits query results to specific data such as time, location, or category
Relationship: connect tables through shared key
Verified Query: a collection of questions and SQL queries that are verified to be correct
Custom instruction: optional, provide the unique context of your business to LLM

Used mainly as fuzzy search with RAG
advantages v/s competition
- Hybrid keyword and vector ranking
- declarative and incremental indexing
specifies
- a query that returns data
- a single column that has text to be indexed and searched
- one or more attributes that are returned when a search succeeds
the underlying table must have change capture enabled for incrementally building search index
the serving and indexing components can be independently suspended
owner rights. The service owner needs all rights, the users need usage privilege

Agentic AI	Generative AI
Proactive: Initiates actions and pursues goals independently	Reactive: Responds to a prompt with an output
Goal-Oriented: Achieves goals through a series of actions	Content-Focused: Creates novel text, images, or code
Planner: Breaks down complex goals and adapts its plans	Non-Planner: Does not engage in multi-step planning
Tool User: Utilizes external tools and APIs to take action	Self-Contained: A system focused solely on generating output
Learns Continuously: Adapts and learns from feedback in dynamic environments	Learns from Training: Primarily relies on its initial training data

Has access to various agents
agents can be connected to multiple tools
3 types of tools
1. Cortex Analyst (or SQL agent): can run queries again Snowflake
2. Cortex Search (or Document/Text search tool):
3. Custom tool: e.g. send email, lookup weather
builtin tools for every agent: chart generator, onboarding/who-ami-i tool (replies with capabilities of other tools)
Calls Cortex agent API to orchestrate and call different tools
- Currently SI makes uses to unpublished features of Cortex Agent API such as deep reasoning, multi-tool selection
Use Planning Instructions under Orchestration tab, to provide general instructions to alter the default behavior of an agent.
- E.g. to always use charts, Whenever answer can be visualized in charts, use chart tool regardless if the user requested it

Extracting a page using AI_EXTRACT always incurs 970 tokens plus prompt and output tokens
It may not be able to parse complex documents. Alternatively, use AI_COMPLETE using a capable LLM to extract data
- Using images instead of PDF is better because PDF has hidden structure and data that is hard to render by LLMs