AI-103: Conversational AI & Agents Flashcards
Microsoft Azure AI App and Agent Developer Associate AI-103 Flashcards
| Front | Back |
| Action Space for Agents | Set of callable functions or tools available |
| Agent Planning | Generate multi step plan before executing tools |
| Chunking Strategy | Split long docs into semantically coherent pieces |
| Clarification Question | Ask when user intent or slots are ambiguous |
| Cold Start for RAG | Seed vector store with curated canonical documents |
| Connector | Adapter that integrates external data sources or services |
| Context Window Limit | Max tokens model can attend to in a single pass |
| Conversation Memory Types | Episodic Semantic and Short term memory |
| Define Dialog State | Representation of user intent context slots and history |
| Dialogue Policy | Rules or model that decides next action |
| Embedding vs Tokenization | Embeddings map text to vectors tokenization splits text |
| End to End Conversation Test | Validate flows from trigger to resolution |
| Entity Extraction | Identifying named items like dates locations names |
| Evaluation Metrics for Conversations | Success rate turns to resolution user satisfaction |
| FAISS Milvus Pinecone | Examples of vector databases |
| Fallback Strategy | Default reply escalation to human or clarification |
| Hallucination Mitigation | Ground answers in retrieved evidence or tools |
| Intent Recognition | Classifying user goal from utterance |
| Multi Modal Tooling | Use vision audio or other modalities within agents |
| OAuth Webhook Connector | Common pattern for authenticated external integrations |
| Orchestration Layer | Coordinates retrieval generation tools and responses |
| Passage Retrieval | Find relevant document chunks for RAG |
| Persona Design | Define assistant tone style and constraints |
| Prompt Template | Parameterized instruction for consistent outputs |
| Purpose of RAG | Use retrieved documents to ground model responses |
| Rate Limiting and Quotas | Protect backend services from excessive calls |
| ReAct Pattern | Interleave reasoning steps and tool actions |
| Retrieval Re ranking | Score candidate documents to pick best evidence |
| Sensitive Data Handling | Redact PII and enforce data retention policies |
| Session Management | Create expire and resume conversation sessions |
| Slot Filling | Collecting required pieces of information across turns |
| Stateful versus Stateless | Stateful tracks context across turns stateless does not |
| System Prompt Role | High level rules that steer assistant behavior |
| Tool Output Validation | Cross check tool responses before user return |
| Tool Safety Checks | Validate inputs outputs and enforce permissions |
| Tool Use Pattern | When to call an external API versus respond directly |
| Tree of Thought | Explore multiple reasoning branches to improve decisions |
| Update Knowledge Base | Re ingest or re index documents after changes |
| User Simulation for Testing | Automated agents that mimic user interactions |
| Vector Store | Database for embedding vectors and nearest neighbor search |
About the Flashcards
Flashcards for the Microsoft Azure AI App and Agent Developer Associate exam cover essential conversational AI concepts, including intent recognition, entity extraction, slot filling, dialog state, session management, dialogue policies, clarification questions, and fallback strategies. Students can review how assistants maintain context, manage memory, and guide conversations toward successful resolutions.
The deck also explores retrieval-augmented generation, embeddings, vector stores, document chunking, agent planning, tool integrations, prompt design, and hallucination mitigation. Additional cards reinforce orchestration, authentication, sensitive data handling, rate limits, tool safety, multimodal capabilities, and conversation testing and evaluation.
Topics covered in this flashcard deck:
- Conversational AI fundamentals
- RAG and vector retrieval
- Agent planning and tools
- Prompts and assistant behavior
- Safety and data handling
- Testing and evaluation