What to learn next

Your roadmap ordered by prerequisite structure and your own practice signal. Concepts whose prerequisites you already know rise to the top; the ones you are failing rise with them.

No practice history found on this device yet, so this is the cold-start path: concepts in pure prerequisite order, foundations first. Practice a few cards in Memorize and this list will re-order around your weak spots.

1
Foundational — a good place to start.Chunking strategies are the lossy compression step that sets the ceiling on everything downstream in a RAG pipeline, fundamentally bounding retrieval quality through a trade-off between retrievability (small, focused chunks) and sufficiency (large chunks with context).
2
Foundational — a good place to start.Evaluation fundamentals is the principled framework for measuring model quality through metric selection, dataset construction, and statistical rigor, validated as a measurement instrument whose proxy and uncertainty must be understood. It serves as the foundation under agent evaluation, benchmark design, and production-pattern gates, where every claim about model capability rests on this rigorous evaluation methodology.
3
Builds on Evaluation Fundamentals.Fine-tuning adapts a pre-trained model to a specific task by updating some or all of its weights on labeled data. It plays the role of reshaping behavior on a base you keep and is the most reliable method for adapting general-purpose models to domain-specific tasks, with full fine-tuning updating every parameter and feature extraction freezing the backbone.
Prerequisites:Evaluation Fundamentals
4
TokenizationStart here
Foundational — a good place to start.Tokenization is the critical interface between raw text and neural language models, acting as a fixed, lossy dictionary through which the model sees only token IDs rather than characters, and directly shaping model quality, inference speed, multilinguality, and task capabilities.
5
Builds on Tokenization.Embeddings are dense, low-dimensional vector representations that map discrete, high-dimensional data (words, sentences, images) into a continuous space where semantically similar items land near each other. They serve as the foundational primitive for downstream AI tasks such as search, retrieval-augmented generation, clustering, and classification, with the embedding quality setting the ceiling for retrieval quality.
Prerequisites:Tokenization
6
Builds on Tokenization.Prompt engineering is the primary interface between human intent and large language model capabilities, functioning as programming a probabilistic interpreter where the prompt is the program; it specifies behavior precisely to narrow the model’s output distribution toward the desired intent, and in production it is treated as a software artifact requiring version control, testing, iteration, and documentation.
Prerequisites:Tokenization
7
Builds on Tokenization, Embeddings, Prompt Engineering, and Evaluation Fundamentals.Context engineering is the systematic practice of designing, assembling, and managing everything that goes into a language model's context window at inference time, subsuming prompt engineering as it decides what information earns a place under a hard token budget, in what order, and with what structure. It serves as the defining discipline of applied AI engineering and the umbrella over dynamic context assembly, context window management, and context compression.
Prerequisites:TokenizationEmbeddingsPrompt EngineeringEvaluation Fundamentals
8
Builds on Prompt Engineering and Context Engineering.Agent architectures are control loops wrapped around a stateless reasoner that determine how language models reason, plan, and act in complex environments, with each named pattern (e.g., ReAct, Plan-and-Execute, Reflexion) defined by when the loop stops and how much structure constrains the next step.
Prerequisites:Prompt EngineeringContext Engineering
9
Builds on Embeddings.A vector database is an approximate nearest neighbor (ANN) index that trades exactness for speed with a tunable recall/latency knob, serving as the storage layer beneath retrieval-augmented generation and semantic search.
Prerequisites:Embeddings
10
Hybrid Search (BM25 and Dense Fusion)Later
Builds on Chunking Strategies and Vector Databases.Hybrid search combines BM25 sparse retrieval and dense vector retrieval through Reciprocal Rank Fusion to outperform either method alone, and the curriculum presents it as the production-recommended default for most systems.
Prerequisites:Chunking StrategiesVector Databases
11
Builds on Chunking Strategies, Vector Databases, and Hybrid Search (BM25 and Dense Fusion).Retrieval strategies are the architecture of a RAG pipeline's retrieval stage—a funnel that trades recall for precision as it narrows, where each stage (hybrid first-stage, reranking, query transformation) fixes a specific failure mode of the prior stage, and that architecture sets the ceiling on answer quality because an LLM cannot reason over information it never receives.
Prerequisites:Chunking StrategiesVector DatabasesHybrid Search (BM25 and Dense Fusion)
12
Builds on Tokenization, Context Engineering, Retrieval Strategies, and Prompt Engineering.Context window management is the engineering discipline of managing a fixed-size token budget allocated across competing claimants like system prompts, retrieved documents, conversation history, and tool definitions, with an eviction policy for overflow and position-aware placement to mitigate the lost-in-the-middle effect, ensuring effective use of the model’s finite context window.
Prerequisites:TokenizationContext EngineeringRetrieval StrategiesPrompt Engineering
13
Builds on Context Window Management, Embeddings, and Agent Architectures.Agent memory is a hierarchical system—short-term working memory (context window), long-term storage (vector databases), and episodic memory—that transforms a stateless language model into a persistent agent capable of remembering past interactions, learning from experience, and maintaining state across sessions.
Prerequisites:Context Window ManagementEmbeddingsAgent Architectures
14
Builds on Context Engineering, Agent Architectures, and Agent Memory.Agent orchestration is the discipline of coordinating multiple agents, models, or processing stages into a coherent system that accomplishes goals no single agent could handle, serving as the connective tissue between individual agent capabilities and the complex, multi-step workflows required by real applications.
Prerequisites:Context EngineeringAgent ArchitecturesAgent Memory
15
Builds on Agent Architectures, Agent Memory, and Agent Orchestration and Workflows.Multi-agent systems distribute complex tasks across specialized agents that collaborate, debate, and coordinate, functioning as a distributed system with non-deterministic nodes communicating in natural language, and their use is justified only when task decomposition outweighs coordination overhead.
Prerequisites:Agent ArchitecturesAgent MemoryAgent Orchestration and Workflows
16
Builds on Tokenization, Agent Architectures, Agent Orchestration and Workflows, and Multi-Agent Systems.LangGraph is a framework for building stateful, multi-actor applications with Large Language Models that models application logic as a directed graph, where nodes are computational steps and edges define control flow; it serves as the orchestration layer for advanced AI systems, providing infrastructure for sophisticated agent architectures that require memory, human-in-the-loop, and complex branching.
Prerequisites:TokenizationAgent ArchitecturesAgent Orchestration and WorkflowsMulti-Agent Systems
17
Reranking Retrieved ResultsLater
Builds on Chunking Strategies, Embeddings, Vector Databases, Retrieval Strategies, and Hybrid Search (BM25 and Dense Fusion).LLM-based re-ranking uses instruction-following LLMs to perform listwise re-ordering of retrieved passages by relevance, serving as a powerful second-stage method to improve search accuracy.
Prerequisites:Chunking StrategiesEmbeddingsVector DatabasesRetrieval StrategiesHybrid Search (BM25 and Dense Fusion)
18
Builds on Embeddings, Vector Databases, Reranking Retrieved Results, and Hybrid Search (BM25 and Dense Fusion).Retrieval-Augmented Generation (RAG) is an agentic AI pattern that feeds a model the specific knowledge it needs at inference time through a retrieval step before generation, turning the model into an open-world agent that consults sources rather than relying solely on pretrained weights.
Prerequisites:EmbeddingsVector DatabasesReranking Retrieved ResultsHybrid Search (BM25 and Dense Fusion)
19
Builds on Retrieval-Augmented Generation (RAG), Evaluation Fundamentals, Retrieval Strategies, and Chunking Strategies.RAG evaluation is a two-stage pipeline evaluation that separately measures retrieval quality (context precision/recall) and generation quality (faithfulness and answer relevance) in order to attribute each failure to a specific stage, with faithfulness being the most critical metric for production systems.
Prerequisites:Retrieval-Augmented Generation (RAG)Evaluation FundamentalsRetrieval StrategiesChunking Strategies
20
Builds on Retrieval Strategies, Retrieval-Augmented Generation (RAG), Chunking Strategies, Vector Databases, Hybrid Search (BM25 and Dense Fusion), Reranking Retrieved Results, and RAG Evaluation.Advanced RAG is a dynamic, adaptive system where an intelligent agent decides the retrieval strategy and combines patterns such as corrective retrieval, multi-hop decomposition, and graph traversal based on query complexity.
Prerequisites:Retrieval StrategiesRetrieval-Augmented Generation (RAG)Chunking StrategiesVector DatabasesHybrid Search (BM25 and Dense Fusion)Reranking Retrieved ResultsRAG Evaluation
21
Builds on Evaluation Fundamentals, Agent Architectures, Multi-Agent Systems, RAG Evaluation, and Agent Memory.Agent evaluation is the process of assessing an agent's full trajectory — its sequence of reasoning steps, tool calls, and side effects — rather than just its final output. It plays the role of determining deployability by measuring multiple dimensions including success rate, efficiency, cost, reliability, and safety, which a single accuracy metric cannot capture.
Prerequisites:Evaluation FundamentalsAgent ArchitecturesMulti-Agent SystemsRAG EvaluationAgent Memory
22
Builds on Evaluation Fundamentals, Retrieval-Augmented Generation (RAG), and RAG Evaluation.Observability for LLM systems extends traditional observability with token tracking, cost monitoring, quality evaluation, and prompt versioning, using distributed tracing across multi-step pipelines to reconstruct the full context of a request after the fact, which enables debugging of stochastic failures through replay and provides both operational signals like latency and cost as well as quality signals from sampled online evaluation.
Prerequisites:Evaluation FundamentalsRetrieval-Augmented Generation (RAG)RAG Evaluation
22 concepts · prerequisite edges grounded from the roadmap corpus. Green prerequisites are strong, amber ones need work, gray are not yet practiced.