AI Bot
AI-500: Data, Knowledge & Memory Management Flashcards
Microsoft Multi-Agent AI Solutions Expert AI-500 Flashcards
Study our AI-500: Data, Knowledge & Memory Management flashcards for the Microsoft Multi-Agent AI Solutions Expert AI-500 exam with 46+ flashcards. View as flashcards, a searchable table, or as a fun matching game.
| Front | Back |
| Define retrieval augmented generation RAG | A technique that retrieves relevant documents then conditions generation on those documents to improve accuracy |
| Explain cosine similarity versus dot product | Cosine similarity measures angle similarity normalized by magnitude dot product measures raw magnitude overlap |
| Explain vector dimensionality trade off | Higher dimensions can capture nuance but increase storage and computation costs |
| How are embeddings generated | By passing text or data through a trained model that outputs fixed length vectors |
| How do embeddings integrate with Cognitive Search | Upload embeddings to a vector capable index to enable semantic similarity search |
| How to build a retrieval pipeline | Ingest clean data chunk text generate embeddings index vectors and implement similarity search then rank results |
| How to choose chunk size | Balance context preservation with embedding capacity often between 200 and 1000 tokens depending on model |
| How to evaluate retrieval quality | Use relevance metrics user satisfaction human evaluation and downstream task performance |
| How to handle multi turn context windows | Use truncation prioritization summarization and hierarchical state to fit context within token limits |
| How to handle sensitive data in embeddings | Apply differential privacy anonymization encryption and strict access controls |
| How to implement feedback loops for memory systems | Capture user corrections validate them and update memory with review processes |
| How to reduce token usage for context management | Use summarization selective recall and retrieval of only relevant chunks |
| How to secure knowledge bases | Use access controls encryption auditing and input sanitization |
| List benefits of retrieval augmented generation | Improves factual accuracy provides source grounding enables up to date responses |
| Name long term memory strategies | Selective storage summarization periodic review retrieval indexing |
| Name mitigation techniques for hallucinations | Use retrieval grounding chain of thought verification and constrained decoding |
| What are ANN algorithms | Approximate Nearest Neighbor algorithms that speed up similarity search with sublinear performance |
| What are best practices for indexing large corpora | Use incremental ingestion sharding batching and monitoring for index health |
| What are common challenges of RAG | Retrieval errors outdated sources context misalignment and hallucinations |
| What are embeddings | Numerical vector representations that capture semantic meaning of text or other modalities |
| What are knowledge graphs | Graph structures that represent entities and relationships to enable reasoning and rich queries |
| What are structured semi structured and unstructured data | Structured data follows a schema semi structured has tags or keys and unstructured is freeform like text or images |
| What is a knowledge base | A curated repository of facts documents and relations used to support retrieval and reasoning |
| What is a vector store | A database optimized for storing embeddings and performing similarity search |
| What is Azure Cognitive Search | A managed search service that provides indexing semantic search and enriched document ingestion |
| What is Azure Cosmos DB | A globally distributed multi model database service for low latency and high availability |
| What is chunking in document processing | Splitting large documents into smaller chunks for embedding and retrieval |
| What is context management in conversational AI | Managing the information that determines conversation state and influences model responses across turns |
| What is data modeling in AI systems | The process of structuring and organizing data for AI applications using schemas types and relationships |
| What is embedding drift | When embeddings for similar content change over time due to model updates or data shifts leading to inconsistency |
| What is Faiss | An open source library by Facebook for efficient similarity search and clustering of dense vectors |
| What is fine tuning versus retrieval | Fine tuning updates model weights retrieval supplies external relevant content without changing model weights |
| What is hallucination in LLMs | When a model generates plausible but incorrect or fabricated information |
| What is hybrid search | Combining keyword based filtering with vector based semantic similarity to improve relevance |
| What is index latency | Time taken to update or query an index affecting freshness and responsiveness |
| What is long term memory in AI agents | Persistent storage of user preferences facts and events that the agent can recall across sessions |
| What is Milvus | An open source vector database designed for scalable similarity search |
| What is semantic retrieval | Retrieving items based on meaning rather than exact keyword overlap |
| What is session affinity | A technique to route requests from the same user to the same backend to preserve session state |
| What is session state | Transient data that tracks user interactions and context within a conversation session |
| What is the cold start problem in memory systems | When there is insufficient stored data to make effective personalized or retrieval based decisions |
| What is vector quantization | A technique to compress vectors by approximating them with limited codebooks to reduce storage |
| What is versioning for knowledge bases | Tracking updates and history of documents indexes and schemas to enable rollbacks and audits |
| What metrics are used for vector search | Recall at K precision at K mean reciprocal rank and normalized discounted cumulative gain |
| When use vector search versus keyword search | Use vector search for semantic similarity and keyword search for exact matches and boolean logic |
| Why use Cosmos DB for AI workloads | Supports scalable storage fast reads global distribution and multiple data models |