You have a concrete problem to solve and need to know where to look. This table maps common NLP, ML, and AI engineering tasks to the specific chapters and framework tutorials where you will find the techniques, code, and architecture patterns to solve them. Each row includes a concrete example and recommended tools. Bookmark this page; you will come back to it often.
How to Use This Key
Scan the Task column for what you need to accomplish. The Example column gives a concrete scenario. The Tools column links to framework tutorials in the appendices. The Where to Look column links directly to relevant chapters. Most real-world projects combine multiple rows.
Text Understanding and Classification
Categorize, label, and understand text at the document or sentence level
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Sentiment analysis | "Is this product review positive, negative, or neutral?" | HuggingFace, DSPy | 11.1 Prompt Engineering, 12.1 LLM vs. Classical ML, 14.6 Fine-Tuning for Classification |
| Topic classification | "Route this support ticket to billing, technical, or account team." | HuggingFace, LangChain | 11.1 Prompt Engineering, 12.3 Hybrid Pipeline Patterns, 14.6 Fine-Tuning for Classification |
| Intent detection | "Does the user want to book a flight, check status, or cancel?" | LangChain, LlamaIndex | 21.4 Multi-Turn Dialogue, 11.1 Prompt Engineering |
| Spam and content moderation | "Flag toxic comments before they appear on the forum." | HuggingFace | 32.1 LLM Security Threats, 12.1 LLM vs. Classical ML, 31.3 Production Guardrails |
| Zero-shot and few-shot classification | "Classify emails into 50 categories with only 3 examples each." | DSPy, HuggingFace | 11.1 Prompt Engineering, 6.7 In-Context Learning Theory |
Information Extraction
Pull structured data from unstructured text
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Named entity recognition (NER) | "Extract all company names, dates, and dollar amounts from SEC filings." | HuggingFace | 12.5 Structured Information Extraction, 1.1 Intro to NLP, 14.6 Sequence Tasks |
| Relation extraction | "Which drug treats which disease in this clinical trial report?" | HuggingFace, LangChain | 12.5 Structured Extraction, 28.3 Healthcare AI |
| JSON/structured output extraction | "Parse this invoice into {vendor, amount, date, line_items}." | LangChain, Semantic Kernel | 10.2 Structured Output, 12.5 Structured Extraction |
| Table and document parsing | "Extract tables from scanned PDFs into a spreadsheet." | LlamaIndex | 27.3 Document Understanding & OCR, 19.5 Vision-Based Document Retrieval |
| Event extraction and timeline building | "Build a timeline of key events from this legal case file." | LangChain, LlamaIndex | 12.5 Structured Extraction, 25.4 Research Agents |
Text Generation and Summarization
Produce, condense, and transform text
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Abstractive summarization | "Summarize this 40-page research paper in 3 bullet points." | LangChain, LlamaIndex | 11.1 Prompt Engineering, 5.1 Decoding Strategies, 19.4 Document Chunking |
| Creative writing and content generation | "Write marketing copy for a product launch in three different tones." | LangChain | 5.2 Stochastic Sampling, 21.2 Personas & Creative Writing, 11.1 Prompt Engineering |
| Code generation | "Generate a Python REST API from this OpenAPI spec." | LangChain, Semantic Kernel | 25.1 Code Generation Agents, 28.1 Vibe-Coding, 10.2 Structured Output |
| Constrained and structured generation | "Generate valid SQL queries that only reference columns in my schema." | HuggingFace | 5.3 Advanced Decoding & Structured Generation, 10.2 Structured Output |
| Translation | "Translate user-facing UI strings into 12 languages." | HuggingFace | 7.4 Multilingual LLMs, 2.3 Multilingual Tokenization |
| Text-to-SQL | "Let business users query our data warehouse in plain English." | LangChain, LlamaIndex | 20.4 Text-to-SQL in RAG, 23.1 Function Calling |
| Synthetic data generation | "Generate 10,000 realistic customer support conversations for training." | DSPy, HuggingFace | Ch. 13 Synthetic Data, 14.2 Data Preparation |
Search, Retrieval, and Question Answering
Find and surface information from large document collections
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Semantic search | "Find all contracts that mention early termination clauses." | LlamaIndex, LangChain | 19.1 Embedding Models, 19.3 Vector Databases, 1.3 Word Embeddings |
| RAG (retrieval-augmented generation) | "Answer questions about our internal docs with source citations." | LlamaIndex, LangChain | 20.1 RAG Architecture, 19.4 Document Chunking, 29.3 RAG Evaluation |
| Knowledge base Q&A | "Let employees ask HR questions and get policy-grounded answers." | LlamaIndex, LangChain | 20.1 RAG Architecture, 21.3 Memory & Context, 23.5 Agentic RAG |
| Hybrid search (keyword + semantic) | "Combine BM25 and vector search for an e-commerce product catalog." | LlamaIndex | 19.2 Vector Index Algorithms, 20.1 RAG Architecture |
| Recommendation and personalization | "Recommend articles based on what similar readers enjoyed." | HuggingFace | 28.4 Recommendation & Search, 19.1 Embedding Models |
Conversational AI and Chatbots
Build interactive dialogue systems and assistants
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Customer support chatbot | "Handle L1 support tickets: answer FAQs, escalate complex issues." | LangChain, LlamaIndex | 21.1 Dialogue Architecture, 20.1 RAG, 23.1 Function Calling |
| Multi-turn conversation management | "Maintain context across a 20-message travel booking dialogue." | LangGraph, LangChain | 21.3 Memory & Context, 21.4 Multi-Turn Dialogue |
| Voice assistant integration | "Add speech-to-text and text-to-speech to our AI concierge." | HuggingFace | 21.5 Voice & Multimodal Interfaces, 27.2 Audio Generation |
| Persona and character design | "Create a friendly, knowledgeable AI tutor that never gives answers directly." | LangChain, Prompt Templates | 21.2 Personas & Companionship, 11.1 Prompt Engineering |
Agents and Automation
Build autonomous systems that reason, plan, and act
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Tool-using agent | "Let the AI call our APIs to look up orders, process refunds, update CRM." | LangGraph, Semantic Kernel | 22.1 Agent Paradigm, 23.1 Function Calling, 23.4 Custom Tool Design |
| Multi-agent orchestration | "Coordinate a researcher, a coder, and a reviewer to produce a report." | CrewAI, LangGraph | 24.1 Framework Landscape, 24.2 Architecture Patterns, 24.4 Orchestration |
| Web browsing and scraping agent | "Navigate competitor websites and extract pricing into a spreadsheet." | LangGraph | 25.2 Browser Agents, 25.3 Computer Use Agents |
| Deep research agent | "Investigate a technical topic across 50 papers and produce a synthesis." | LangGraph, CrewAI | 25.4 Research Agents, 22.3 Planning & Reasoning |
| MCP and A2A integration | "Connect our agent to Slack, GitHub, and Jira via standard protocols." | LangChain, Semantic Kernel | 23.2 Model Context Protocol, 23.3 Agent-to-Agent Protocol |
| Agent safety and sandboxing | "Prevent our coding agent from accessing production databases." | LangGraph | 26.1 Agent Safety, 26.2 Sandboxed Execution |
Training, Fine-Tuning, and Model Customization
Adapt, align, and build models for your domain, data, and quality bar
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Full fine-tuning | "Train a domain-specific model on 100K medical Q&A pairs." | HuggingFace, W&B/MLflow | 14.1 When to Fine-Tune, 14.3 Supervised Fine-Tuning |
| LoRA and parameter-efficient tuning | "Adapt a 70B model on a single GPU using LoRA adapters." | HuggingFace, W&B/MLflow | 15.1 PEFT Methods, 15.3 Training Platforms |
| RLHF and preference alignment | "Make our model refuse harmful requests while staying helpful." | HuggingFace, W&B/MLflow | 17.1 RLHF, 17.2 DPO, 17.3 Constitutional AI |
| Knowledge distillation | "Compress GPT-4 quality into a 7B model for on-device deployment." | HuggingFace | 16.1 Distillation, 16.2 Model Merging |
| Pretraining from scratch | "Build a 3B-parameter model on proprietary legal corpora." | HuggingFace, Distributed ML | 6.1 Language Model Pretraining, 6.4 Scaling Laws, 4.1 Transformer Architecture |
| Data curation and filtering | "Clean, deduplicate, and quality-filter 500GB of web-scraped text." | HuggingFace, Datasets | 6.2 Data Pipelines, 14.2 Data Preparation, Ch. 13 Synthetic Data |
| Continual and incremental learning | "Update our model monthly with new product data without catastrophic forgetting." | HuggingFace, W&B/MLflow | 14.4 Continual Fine-Tuning, 16.2 Model Merging |
| Embedding model training | "Train a custom embedding model for our legal document corpus." | HuggingFace | 19.1 Embedding Models, 14.5 Representation Learning |
Scaling, Infrastructure, and Optimization
Train faster, serve cheaper, and run models on any hardware
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Distributed training | "Train across 8 GPUs with data and tensor parallelism." | Distributed ML, HuggingFace | 6.3 Distributed Pretraining, 6.5 Parallelism Strategies |
| Quantization for deployment | "Reduce model size from 32-bit to 4-bit for edge serving." | HuggingFace, Inference Serving | 9.1 Model Quantization, 9.5 Pruning & Sparsity |
| Inference serving at scale | "Serve our model at 1,000 requests/second with p99 latency under 2s." | Inference Serving | 9.4 Serving Infrastructure, 31.1 Deployment Architecture, 9.2 KV-Cache & Batching |
| Hardware planning and GPU selection | "Choose the right GPU cluster for fine-tuning a 70B model." | Hardware Guide | 9.3 Hardware Landscape, 33.5 Compute Planning |
| KV-cache optimization and speculative decoding | "Cut time-to-first-token by 3x with speculative decoding." | Inference Serving | 9.2 KV-Cache & Batching, 8.3 Test-Time Compute |
| Model compression and pruning | "Prune 40% of weights while keeping 95% of benchmark accuracy." | HuggingFace | 9.5 Pruning & Sparsity, 16.1 Distillation |
Deployment and Operations
Ship, monitor, and maintain LLM systems in production
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Production deployment architecture | "Design the infra for a multi-tenant LLM service with rate limiting." | Inference Serving, Env Setup | 31.1 Deployment Architecture, 31.3 Scaling & Guardrails |
| Observability and monitoring | "Set up tracing, cost tracking, and quality dashboards for our LLM app." | W&B/MLflow | 30.1 Observability & Tracing, 30.2 Drift Detection, 26.3 Agent Cost Control |
| Cost optimization | "Cut our monthly LLM API spend from $40K to $15K without losing quality." | W&B/MLflow | 33.3 ROI & Value, 33.5 Compute Planning, 9.1 Quantization |
| A/B testing and experiment tracking | "Compare two prompt variants on 10% of production traffic." | W&B/MLflow, DSPy | 29.2 Experimental Design, 31.2 CI/CD for LLMs |
| Build vs. buy decision | "Should we fine-tune an open model or use GPT-4 via API?" | 33.4 Vendor Evaluation, 33.6 Build vs. Buy & TCO, 12.1 LLM vs. Classical ML |
Multimodal and Domain Applications
Work with images, audio, video, and industry-specific problems
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| Image understanding and VQA | "Describe the contents of product photos for accessibility." | HuggingFace | 27.1 Vision-Language Models, 10.4 Multimodal APIs |
| Image generation | "Generate product mockups from text descriptions." | HuggingFace | 27.1 Image Generation, 5.4 Diffusion Models |
| Financial NLP | "Extract sentiment signals from earnings call transcripts for trading." | LangChain | 28.2 LLMs in Finance, 12.5 Structured Extraction |
| Healthcare and biomedical AI | "Summarize patient records and flag medication interactions." | HuggingFace | 28.3 Healthcare AI, 32.3 Bias & Fairness |
| Legal document analysis | "Review contracts for non-standard clauses and flag risks." | LlamaIndex, LangChain | 28.4 Legal AI, 20.1 RAG Architecture |
| Education and tutoring | "Build a Socratic tutor that adapts to student knowledge level." | LangChain | 28.6 Education, 21.2 Personas |
| Cybersecurity applications | "Analyze log files to detect anomalous access patterns." | LangChain | 28.5 Cybersecurity, 32.8 Red Teaming |
| Robotics and embodied AI | "Use an LLM to plan robot navigation in a warehouse." | HuggingFace | 28.7 Robotics & Embodied AI, 22.3 Planning |
Safety, Evaluation, and Governance
Test, secure, audit, and regulate AI systems
| Task | Example | Tools | Where to Look |
|---|---|---|---|
| LLM evaluation and benchmarking | "Compare three models on our domain-specific test set." | W&B/MLflow, Benchmarks | 29.1 Evaluation Fundamentals, 29.2 Experimental Design |
| Hallucination detection and mitigation | "Ensure our medical chatbot never fabricates drug dosage information." | LangChain, LlamaIndex | 32.2 Hallucination & Reliability, 20.1 RAG (grounding), 29.3 RAG Evaluation |
| Prompt injection defense | "Prevent users from jailbreaking our customer-facing assistant." | 32.1 Security Threats, 26.1 Agent Safety | |
| Bias and fairness auditing | "Test our hiring assistant for demographic bias before launch." | HuggingFace | 32.3 Bias & Fairness, 32.5 Risk Governance |
| Regulatory compliance (EU AI Act) | "Classify our AI system's risk level and prepare documentation." | Model Cards | 32.4 Regulation, 32.9 EU AI Act |
| Interpretability and model understanding | "Explain why the model rejected this loan application." | HuggingFace | 18.1 Attention & Probing, 18.3 Practical Interpretability, 18.2 Mechanistic Interpretability |
| Red teaming | "Systematically probe our model for harmful outputs before launch." | 32.8 Red Teaming, 29.1 Evaluation Fundamentals |