Snowflake AI & LLM Fundamentals: From LLMs to RAG

Understand how a large language model processes text by tokenizing input, mapping tokens to unique ids, predicting the next token, and iterating until the end of the token sequence.

Explore the landscape of machine learning models, from object detection and weather forecasting to stock price prediction and large language models, with LLMs standing out for self-supervision.

The LLM model gains popularity through self-supervised learning, overcoming labeling bottlenecks, by inferring labels from input data and predicting next tokens with context.

Explore how LLM architecture works under the hood, and how LLMs understand and respond to human input, answering key questions about processing and generating language.

Explore LLM architecture through approach one, building a static word dictionary with IDs and meanings, using a barcode analogy, but reveal how fixed meanings miss nuanced intent.

Explore word2vec embeddings as coordinates in a multi-dimensional space and how similar words cluster. Understand why fixed coordinates create semantic ambiguity in approach two.

Explore approach 3 using self-attention and transformers to generate context-aware word meanings, replacing static embeddings with dynamic, sentence-specific representations via query, key, and value.

Explore why embeddings add meaningful context to tokens by mapping words into a multi-dimensional space, where embedding vectors carry different features and meanings across dimensions beyond token ids.

Explore how humans derive meaning from word context and how transformer architecture enables machines to do the same, by showing how each word influences surrounding meanings.

Explore how word embeddings enable cross-word comparisons through matrix multiplication and cosine similarity, measuring similarity between words like cat, kitten, and car.

Focus on common embedding dimensions when comparing words, rather than comparing every dimension across all words. This approach uses shared dimensions to reveal how words relate.

The lecture introduces the query matrix and the key matrix as a big dictionary for word embeddings and shows how dot products between their extracted dimensions measure similarity.

Explore how embeddings are compared using query and key matrices to compute attention scores in large language models, with step-by-step math intuition from token embedding to attention.

Explore how softmax and masking drive attention, and what the value matrix represents, as embedding and positional information transform tokenized input into context-aware vectors for predicting the next token.

Convert raw attention scores with softmax into probabilities that sum to one across each row. Show how the sample input 'cat sat on the' yields 'on' receiving about 28% attention.

Explore masking in self-attention, where a token attends only to previous tokens, not future ones. Learn how masking and softmax produce the final attention score matrix in language models.

Explore how the value matrix shapes word meaning in embeddings by showing how adding rotten to apple shifts its embedding toward the rotten concept through attention and value extraction.

Explore the value matrix in self-attention, calculating raw attention scores with query, key, and value matrices, applying masking and softmax to form context-aware embeddings for next-token prediction.

Discover how Snowflake ai extract, an ai sql function built on cortex ai, uses a 128k token visual language model to read up to 125 pages.

Explore how AI Extract uses a visual language model to turn unstructured data, such as PDFs or images, into structured outputs by answering predefined questions within an ELT workflow.

Use AISQL AIExtract to derive first name and city from unstructured data, then parseJson to convert the result into a Snowflake json object for dot notation queries.

Upload a PDF to a Snowflake staging area, run AI extract in an ELT workflow, and obtain entities like inspection grade, ingestion unit, serial number, and name Emily Johnson.

Use AI extract to pull key metrics from an image file of a Snowflake table, exploring 2024 and 2025 ending paid and churn values.

Discover how a visual language model processes multi-page documents by creating mathematical models to answer questions, up to 125 pages with a 120k token context, and compare costs.

Evaluate AI-Extract's VLM function in scenarios such as reliance on training data, cross-reference conflicts, line traps, lying reports, misaligned grids, and zigzag schedules to reveal VLM performance and its limitations.

Analyze how Snowflake's AI extract visual language model handles contract documents to reveal liability terms, detect contradictions, and rely on document content rather than training data, using prompts and tests.

This lecture demonstrates testing the AI extract VLM for cross-reference conflicts in a multi-page lease agreement, using Snowflake to verify the late fee cap of 200 despite a 10% calculation.

Explore how a visual language model handles contradictions between text and charts in the lying report test. Examine visual grounding, text reliance, and gaslighting risks in VLM-powered summaries.

Explore how a misaligned grid in a Snowflake ELT test affects the VLM extract function, showing how firewall and database statuses can be misread as safe instead of critical danger.

Conduct a zigzag schedule test to evaluate AI extract and VLM model performance in Snowflake, revealing a spatial disconnect and a reading-order error, while the zigzag test passes.

Introduce aiParseDocument, built on snowflake-artic-tilt, a VLM-based document AI model that reads PDFs, outputs parsed JSON, stores it in a table, and powers AI extract for answers.

This AI parse demo shows parsing a document in Snowflake with the APARSE layout mode, producing a Markdown JSON, and querying inspection grade and injection unit status, noting OCR caveats.

Test the layout awareness of the AIPars document function by processing an 11-page tables document into a json object. Show impairment loss in 2010 is 5000; none for the date.

Learn how to embed text in Snowflake with the AI embed function, store vectors, and use AI similarity and vector cosine similarity to match queries with content.

Discover why vector embeddings use 768 concept dimensions to distinguish objects and capture language complexity, illustrating with laptops and coffee machines.

Analyze the limitations of using embeddings for all structured data and show when to rely on exact-value sql queries for precise error messages.

demonstrates building a cortex search service on transcript text in snowflake, using embedding-based cosine similarity and vector search to rank results by region.

Demonstrates using snowflake cortex search to concatenate ticket id and issue description into a searchable text, enabling semantic and keyword search, and previews the multi-index cortex search solution.

Discover how multi-index cortex search in Snowflake blends text and vector indexes to boost ticket search results, using create or replace cortex search service and issue description indexing.

Explore how search operates as a ranker, not a filter, returning a sorted list of possibilities from vector indexes and cortex search service, driven by keyword and semantic similarity.

Explore how concatenating description data in cortex basic search blurs distinctions and reveals results, then learn how multi-index search targets a specific column as a text index to improve precision.

Discover how Cortex search service powers retrieval augmented generation (RAC) to fetch relevant tickets, enrich results with an LLM, and generate human-like responses.

Apply column weights to multi-index searches in Snowflake, using text and vector weights to influence results through the weights config, while scores stay consistent.

Learn to implement a multi-column search in Snowflake by adding text and vector indexes across product code, product name, description, and category, then run multi-index queries with filtering.