diff --git a/data-fundamentals-dev-rel/quiz/quiz.md b/data-fundamentals-dev-rel/quiz/quiz.md index 3ed809118..d1397cf1e 100644 --- a/data-fundamentals-dev-rel/quiz/quiz.md +++ b/data-fundamentals-dev-rel/quiz/quiz.md @@ -66,13 +66,27 @@ Q: What's the likely data type you should store a JSON document in Oracle? * JSON > Your default for storing JSON should likely be a JSON data type, but it really depends on how large of a document you're storing and your use case. +Q: What do each dimension of a vector embedding represent? +- A specific dictionary word, making the embedding basically an inefficient thesaurus. +* A value created by an embedding model representing some aspect of semantic meaning or context about the data. +- A technical persons emotional stage while explaining vector indexing to management. +> Each dimension is a feature learned during training, not something with a fixed human meaning. + Q: How does Cosine similarity measure distance? - It gets out a tape measure and... - It measures concise distance between vectors * It measures the angle between vectors +> Cosine similarity measures the arc between the axis of a point in vector space. The tradeoff is Cosine misses the magnitude of how many times that point exists in vector space. + +Q: You perform a query on a vector index, but it returns utterly unrelated results or zero results. What could the problem be? +- Your embeddings caught a sickness from an unsanitized CSV import. +- The index needs to be rebuilt because SQL query plans expire after 24 hours. +* The query embedding and stored embeddings came from different models, so they're not comparable. +- The similarity function used at query time doesn't match the one the index was built with. +> Vector similarity only makes sense when query and stored vectors come from the same embedding model, since different models place semantically identical content in completely different vector spaces. Mixing them (or using a mismatched distance metric) gives you numbers that look like results but mean nothing. ``` ## Acknowledgements * **Authors** - Kirk Kirkconnell -* **Last Updated By/Date** - Kirk Kirkconnell, June 2026 +* **Last Updated By/Date** - Kirk Kirkconnell, July 2026 diff --git a/dev-rel-agent-memory/quiz/quiz.md b/dev-rel-agent-memory/quiz/quiz.md index 1c1fde096..68d480d2a 100644 --- a/dev-rel-agent-memory/quiz/quiz.md +++ b/dev-rel-agent-memory/quiz/quiz.md @@ -79,4 +79,4 @@ Q: In the cross-inspector handoff scenario, how does Vance benefit from Mercer ## Acknowledgements * **Authors** - Kirk Kirkconnell -* **Last Updated By/Date** - Kirk Kirkconnell, January 2026 +* **Last Updated By/Date** - Kirk Kirkconnell, July 2026 diff --git a/dev-rel-rag-to-agents/introduction/introduction.md b/dev-rel-rag-to-agents/introduction/introduction.md index ae1f9e675..165979a76 100644 --- a/dev-rel-rag-to-agents/introduction/introduction.md +++ b/dev-rel-rag-to-agents/introduction/introduction.md @@ -2,7 +2,7 @@ ## About this Workshop -In this lab, we'll interact with a data set from an application named Prism CityOps. It's an application used by the city of Kirkland to monitor and manage city infrastrucuter such as ongoing maintenance, active monitoring via IoT devices, crew dispatch, record keeping, and more. +In this lab, we'll interact with a data set from an application named Prism CityOps. It's an application used by the city of Kirkland to monitor and manage city infrastructure such as ongoing maintenance, active monitoring via IoT devices, crew dispatch, record keeping, and more. You'll utilize a Jupyter Notebook with real Python code, completing real activities, and building on what you already know about databases and AI Vector Search. You'll create a RAG pipeline using data in the database and work your way up to implementing a full on AI agent, with tools and a little memory, utilizing LangChain as the agent framework. @@ -55,4 +55,4 @@ This lab assumes you have: ## Acknowledgements * **Authors** - Kirk Kirkconnell -* **Last Updated By/Date** - Kirk Kirkconnell, June 2026 +* **Last Updated By/Date** - Kirk Kirkconnell, July 2026 diff --git a/dev-rel-rag-to-agents/quiz/quiz.md b/dev-rel-rag-to-agents/quiz/quiz.md index 59b6a9052..c1a301cb3 100644 --- a/dev-rel-rag-to-agents/quiz/quiz.md +++ b/dev-rel-rag-to-agents/quiz/quiz.md @@ -19,7 +19,7 @@ Estimated Time: 5 minutes ```quiz score Q: What is the main purpose of the RAG retriever in Section 1? -* To retrieve semantically relevant Prism chunks using in-database vector embeddings and vector distance +* To retrieve semantically relevant Prism data chunks using in-database vector embeddings and vector distance - To let the LLM choose arbitrary database tables to query - To store long-term memories for later conversations - To format the final incident report as markdown @@ -30,7 +30,7 @@ Q: What does the raw RAG implementation demonstrate before the LangChain version - RAG automatically performs multi-step planning - RAG removes the need to inspect retrieved context * RAG is mainly retrieval, prompt assembly, and an LLM call -> The raw version makes the moving parts visible, which helps learners understand what a framework simplifies later. +> The raw version makes the moving parts visible, which helps you to understand what a framework simplifies later. Q: What is the key tradeoff of a LLM-driven workflow? * It is predictable and auditable, but less adaptive than an agent @@ -40,10 +40,10 @@ Q: What is the key tradeoff of a LLM-driven workflow? > The workflow follows a fixed path with bounded LLM calls, which improves control but limits flexibility when the problem changes. Q: Why do tool signatures and docstrings matter? -- They change the physical database schema used by the tool +- They change the physical database schema * They define the model-facing interface that helps the LLM decide when and how to call a tool - They automatically validate every SQL result returned by the database -- They prevent the framework from executing tool calls +- They ensure tools nicely hang on the wall so you're not digging through a cabinet in the garage > The LLM sees the tool name, inputs, and description, so clear interfaces improve tool selection and usage. Q: What makes the unified query important?