🧠 AI Agents Interview Questions

AI Agents are autonomous systems that can perceive their environment, make decisions, and take actions to achieve goals. They go beyond simple question-answering.

Key Differences:

Agent Components:

• Perception: Understand environment through inputs
• Reasoning: LLM-based decision making
• Planning: Break down goals into steps
• Memory: Store and retrieve context
• Action: Use tools to interact with world
• Learning: Improve from feedback

ReAct is a powerful agent pattern that combines reasoning (thinking) with acting (using tools). The agent iteratively thinks about what to do, takes action, and observes results.

LangChain and LangGraph are popular frameworks for building AI agents with pre-built components and orchestration capabilities.

Advanced agents require sophisticated planning, memory management, and self-reflection capabilities to handle complex tasks autonomously.

RAG (Retrieval-Augmented Generation) is a technique that enhances LLM responses by retrieving relevant information from external knowledge sources before generating answers.

Why RAG is Important:

• Reduces hallucinations: Grounds responses in actual data
• Up-to-date knowledge: Access information beyond training cutoff
• Domain-specific: Incorporate proprietary or specialized data
• Cost-effective: Cheaper than fine-tuning for knowledge updates
• Transparency: Can cite sources for generated content

RAG vs Fine-tuning:

Chunking is the process of splitting documents into smaller pieces for embedding and retrieval. The strategy significantly impacts RAG quality.

Common Chunking Strategies:

Chunking Best Practices:

• Chunk size: 500-1500 characters for most use cases
• Overlap: 10-20% of chunk size to preserve context
• Metadata: Include source, page number, section headers
• Semantic boundaries: Split at paragraphs, not mid-sentence
• Test and iterate: Evaluate retrieval quality with your data

Embedding models convert text into dense vectors that capture semantic meaning. Choosing the right model affects retrieval quality significantly.

Popular Embedding Models:

Selection Criteria:

• Accuracy: Check MTEB benchmark scores
• Cost: API costs vs self-hosted
• Latency: Embedding generation time
• Language: Multilingual support needed?
• Privacy: Can data leave your infrastructure?

Vector databases store and efficiently search high-dimensional embeddings using approximate nearest neighbor (ANN) algorithms.

Popular Vector Databases:

• Pinecone: Fully managed, easy to use, scalable
• Weaviate: Open source, hybrid search, GraphQL API
• Chroma: Lightweight, embedded, great for prototyping
• Milvus: Open source, highly scalable, cloud-native
• Qdrant: Rust-based, fast, filtering support
• pgvector: PostgreSQL extension, SQL integration

Retrieval Methods:

Hybrid search combines semantic (vector) search with traditional keyword (BM25/TF-IDF) search to get the best of both approaches.

Why Hybrid Search?

• Semantic search: Understands meaning, handles synonyms
• Keyword search: Exact matches, specific terms, acronyms
• Combined: Better recall and precision

Hybrid Search Best Practices:

• Weight tuning: Test different keyword/semantic ratios
• Re-ranking: Use cross-encoder for final ranking
• Query expansion: Add synonyms for better recall
• Evaluation: Measure with nDCG, MRR, recall@k

LangChain is a framework for building applications with LLMs. It provides modular components that can be composed together.

Core Components:

• Models: LLMs and Chat Models (OpenAI, Anthropic, etc.)
• Prompts: Templates for structuring LLM inputs
• Chains: Sequences of calls (LLM, tools, etc.)
• Memory: Persist state across chain runs
• Agents: Use LLMs to decide which tools to use
• Tools: Functions that agents can call
• Retrievers: Fetch relevant documents

LCEL (LangChain Expression Language) is a declarative way to compose chains using the pipe operator. It's the modern, recommended way to build LangChain applications.

Key Benefits:

• Streaming: First-class streaming support
• Async: Native async/await support
• Parallel: Automatic parallelization
• Retries: Built-in retry logic
• Tracing: LangSmith integration

Tools in LangChain are functions that agents can call. They require a name, description, and implementation.

Tool Design Best Practices:

• Clear descriptions: Help LLM understand when to use
• Error handling: Return helpful error messages
• Input validation: Use Pydantic for type safety
• Async support: Implement _arun for async agents

LangSmith is a platform for debugging, testing, evaluating, and monitoring LangChain applications. It's essential for production deployments.

Key Features:

• Tracing: See every step in your chain
• Debugging: Inspect inputs/outputs at each step
• Evaluation: Test chains against datasets
• Monitoring: Track latency, costs, errors
• Datasets: Create test sets for evaluation

Production Monitoring:

Understanding when to use each approach is crucial for building effective LangChain applications.

Quick Decision Guide:

• Simple Q&A or RAG: LCEL Chain
• Tool use with reasoning: Agent
• Multi-step with retries: LangGraph
• Multi-agent systems: LangGraph
• Production reliability: LangGraph (better control)