SnowPro Specialty Gen AI GES-C01Demo

SNOWFLAKE.CORTEX.SPLIT_TEXT_RECURSIVE_CHARACTER recursively splits text for embedding or search-indexing preprocessing and returns an array of strings containing the extracted text chunks.

Cortex Analyst enables business users to ask natural-language questions about structured data and receive direct answers without writing SQL. It provides managed text-to-SQL capabilities, generating and executing the required SQL on the user’s behalf.

Snowflake states that customer data is not available to other customers or model developers, and that access to Snowflake AI features and models can be governed through familiar role-based access control (RBAC).

The SNOWFLAKE.CORTEX_USER database role is granted to PUBLIC by default, and PUBLIC is available to all users and roles. Restricting default Cortex access requires revoking that database role from PUBLIC and revoking imported privileges on the SNOWFLAKE database from PUBLIC; access can then be selectively granted to intended roles.

Manhattan distance is the L1, or taxicab, distance, calculated by summing the absolute differences of corresponding vector coordinates. VECTOR_L1_DISTANCE performs this calculation; L2 distance is Euclidean distance.

A document RAG pipeline extracts document content, stores embeddings for that content, retrieves the most relevant content through vector similarity, and supplies the retrieved context to an LLM for analysis. PARSE_DOCUMENT extracts document content, EMBED_TEXT_768 produces vector embeddings, and COMPLETE can analyze the retrieved context.

CORTEX_SEARCH_DAILY_USAGE_HISTORY records daily Cortex Search consumption for serving and text-embedding tokens (as well as batch usage). Serving compute is tracked as SERVING, and embedding usage is tracked as EMBED_TEXT_TOKENS. Virtual warehouse compute is a Cortex Search cost but is not included in this view.

Creating a Snowflake Document AI model build requires the CREATE SNOWFLAKE.ML.DOCUMENT_INTELLIGENCE privilege on the schema in which the model build is created.

Snowflake Document AI uses Snowflake Arctic-TILT (Text Image Layout Transformer), Snowflake’s proprietary large language model for extracting data from documents.

Cortex Analyst lets applications answer natural-language questions about structured Snowflake data by generating and executing SQL from a semantic model. It is available through a REST API for integration into applications.

temperature controls the randomness of generated output, while top_p controls randomness and diversity by restricting the candidate-token probability mass used for sampling.

Cortex Search provides a managed hybrid retrieval engine for RAG. It handles embedding-related infrastructure and uses vector search to retrieve semantically similar documents, consolidating work that otherwise requires separate embedding and similarity-search steps.

Cortex COMPLETE bills for processed tokens, including output tokens; max_tokens caps the number of output tokens generated and can therefore reduce cost. Choosing a smaller LLM with fewer parameters can also reduce the model’s inference cost. The max_tokens setting does not limit input tokens, and virtual warehouse sizing is not a direct control for Cortex COMPLETE inference charges.

Snowflake Model Registry stores models as schema-level objects, supports lifecycle management through multiple model versions, and requires a default version. Registered model methods can be invoked for inference from SQL, including with the MODEL(model_name)!method_name(...) syntax.

Without a Verified Query Repository (VQR), Cortex Analyst uses its underlying LLMs to generate up to three suggested questions. When a VQR is defined, it returns up to five questions from that repository, selected according to their similarity to the user’s input.

Snowflake Cortex exposes LLM completion inference through a REST endpoint. The REST Cortex Inference API performs text completion inference comparable to snowflake.cortex.complete, enabling programmatic access outside the SQL interface. Copilot and Cortex LLM Playground are user interfaces, and the SQL function is invoked within Snowflake SQL.

Cortex Analyst does not store state between requests. Multi-turn context must be supplied on every request as the complete, chronologically ordered conversation history in the messages field, including prior user messages and Analyst responses.

Cortex Search serving compute continues to incur cost while a service is available, even when it receives no queries. Suspending serving during idle periods prevents those serving costs; intermittent services can also use AUTO_SUSPEND after query inactivity.

SNOWFLAKE.CORTEX.COMPLETE supports a response_format JSON schema, allowing a completion to return a structured object whose three named properties can be projected into separate columns. SNOWFLAKE.CORTEX.EXTRACT_ANSWER returns a string answer to one specified question from a text document, so one invocation per requested value can populate three separate columns.

SNOWFLAKE.CORTEX.COMPLETE requires a model identifier and the prompt or conversation history. Its optional third options argument can contain generation settings such as sampling or output controls. Snowflake documentation

Document AI model inference using model_build_name!PREDICT returns a JSON result containing extracted fields, confidence scores, and document metadata, which is semi-structured data. A named stage created without an external URL is an internal Snowflake stage, and the numeric second argument selects the Document AI model version.

Snowflake Cortex Analyst answers natural-language questions about structured data by using a semantic model to generate the required SQL. A semantic model stored on a stage can be used in the Cortex Analyst experience in Snowsight or supplied to the Cortex Analyst API by a custom application.