Roadmap de Estudos
AI-Native 2026

Q/K/V como projeções lineares, Multi-Head vs Grouped-Query Attention, flashattention e os mecanismos de KV cache que tornam inferência eficiente em contextos longos

• Self-attention: Q/K/V projections, escala √d_k, operação softmax
• Multi-Head vs Multi-Query Attention vs Grouped-Query (GQA) trade-offs
• KV cache: acúmulo K/V por token, eviction policy, prefill vs decode phase
• Positional Encoding: RoPE (extrapolação suave), ALiBi (linear bias), NoPE
• FlashAttention 2/3: IO-aware attention, SRAM tiling, sub-quadratic memory
• Speculative decoding: drafter + verifier, token acceptance rate e speedup

Leis de Kaplan e Chinchilla, phase transitions para capacidades emergentes, tokenização BPE e arquiteturas pós-transformer MoE e SSM

• Kaplan (2020): potência entre compute, parâmetros e loss — overshooting anterior
• Chinchilla (2022): tokens = 20× parâmetros, optimal compute frontier
• Emergência: BIG-Bench phase transitions, unpredictable capability jumps
• Tokenização BPE & SentencePiece: byte-level, vocab size vs cobertura
• MoE (Mixture of Experts): routing, sparse activation, GPT expert claims
• SSM alternativas: Mamba (selective state space), RWKV hybrid approaches

Token budget management, RAPTOR e compression chains, prefix caching e políticas de sliding window para contextos de 1M+ tokens

• Budget allocation: static vs dynamic, per-component accounting, headroom policy
• "Lost in the Middle": posição relativa importa — primacy e recency bias
• RAPTOR: recursive abstractive processing, clustering semântico hierárquico
• KV cache reuse: prefix sharing, cache warming, TTL e invalidation triggers
• Sliding window + chunking: overlap, stride, span selection por relevância
• Context compression: entropy-weighted pruning, selective summarization chains

Vector store internos (HNSW, IVF, PQ), retrieved híbrido BM25+dense, reranking neural e arquiteturas de memória episódica para agentes de longa duração

• Dense retrieval: bi-encoders, cross-encoders, late interaction ColBERT
• HNSW vs IVF+PQ: recall@k, latência de busca, tamanho de índice trade-offs
• Hybrid search: BM25 + dense, RRF (Reciprocal Rank Fusion)
• Neural reranking: cross-encoder reranker, MonoT5, listwise rerankers
• Memória episódica vs semântica: MemGPT, Mem0, A-MEM consolidation
• Políticas de memória: TTL, importance scoring, forgetting curves

DSPy transforma engenharia de prompts em programação de módulos LM tipados — Signature, ChainOfThought, Retrieve e optimizadores MIPRO/BootstrapFewShot

• DSPy Signature: typed input/output spec que substitui string literals de prompt
• Módulos: dspy.Predict, ChainOfThought, ReAct, ProgramOfThought, Retrieve
• Optimizers: BootstrapFewShot, MIPRO v2, COPRO — otimização automática de prompts
• Assertions & Suggestions: constraints declarativos que deflect ou assert no output
• TypedPredictor: Pydantic models como output type, validação automática
• Pipeline end-to-end: compilação, traces, evals integradas ao optimizer

Engenharia de ontologias OWL/RDF, SPARQL para queries sobre grafos de conhecimento, GraphRAG e integração de lógica de primeira ordem com LLMs

• OWL 2: classes, propriedades de objeto, restrições axiomáticas (DL expressividade)
• RDF/SPARQL 1.1: grafos de triplas, SELECT/CONSTRUCT/ASK, property paths
• KG Embeddings: TransE, RotatE, ComplEx — representação em espaço latente
• GraphRAG & Subgraph-RAG: subgraph retrieval como contexto estruturado
• Constraint propagation: SAT, CSP solvers como validadores de saída LLM
• Logic programming + LLMs: Prolog, Datalog, Answer Set Programming (ASP)

Arquitetura Host/Client/Server, JSON-RPC 2.0 sobre stdio, HTTP+SSE e Streamable HTTP — lifecycle de sessão e negociação de capacidades

• JSON-RPC 2.0: request/response/notification, batch, error code taxonomy
• Transport stdio: framing newline-delimited, process lifecycle, sequência de init
• HTTP+SSE: SSE para server→client (GET), POST para client→server
• Streamable HTTP (spec 2025-11-25): session resumption, upgrade de SSE
• Capability negotiation: initialize handshake, protocol versioning, roots
• Tool annotations: readOnlyHint, destructiveHint, idempotentHint, openWorldHint

OAuth 2.1 com PKCE para servidores remotos, JSON Schema validation para ferramentas, Sampling schema e defesa contra prompt injection via MCP tools

• OAuth 2.1 + PKCE: authorization code flow, token rotation para MCP remoto
• Tool JSON Schema: input validation rigorosa, additionalProperties: false
• Sampling schema: temperature, top_p, stop sequences, max_tokens como contrato
• Prompt injection via MCP: attack vectors, tool result poisoning, mitigações
• Sandboxing: Docker isolado para tools destrutivas, read-only mounts
• A2A Protocol (Google): agente-para-agente via HTTP+JSON-RPC, agent cards

ReAct (reason+act), Tree-of-Thoughts com beam search e MCTS, Reflexion com memória verbal e Self-Consistency por amostragem múltipla

• ReAct: thought→action→observation loop, grounding externo no ambiente
• Chain-of-Thought (Wei et al.): zero-shot CoT, exemplar selection, processo passo-a-passo
• Tree-of-Thoughts: nós de raciocínio, beam search, BFS vs DFS vs MCTS
• Reflexion (Shinn et al.): estado episódico, self-eval criteria, memória verbal
• Self-Consistency: múltiplos paths de raciocínio, aggregation por votação
• Evaluator-Optimizer: generator + critic loop com critério externo definido

Orchestrator-Workers, Parallelization, comunicação estruturada inter-agente, estado compartilhado e padrões de Human-in-the-Loop com checkpoints

• Orchestrator-Workers: delegação dinâmica, routing por capacidade e especialização
• Parallelization: fan-out + join, rate limiting, concurrency control per tool
• Comunicação inter-agente: typed message contracts, schema validation
• Estado compartilhado: eventual consistency, conflict resolution, CRDT patterns
• Self-healing: diagnóstico automático, retry com backoff, circuit breaker
• HITL (Human-in-the-Loop): checkpoints, interrupt patterns, approval gates

Loop agêntico perceive→plan→act→reflect, subagents paralelos, CLAUDE.md como contrato com o agente e GitHub Copilot agent mode com MCP integration

• Claude Code: subagents, tarefas paralelas, extended thinking em code review
• CLAUDE.md: estrutura do projeto, comandos, boas práticas — contract com o agente
• AGENTS.md: multi-agent coordination, project map, agent skill routing
• Copilot agent mode: inline + sidebar + agent, tool calls, MCP servers
• .instructions.md: applyTo globs, scoped context, instruction layering
• Cursor: .cursor/rules vs .cursorrules, composer context, notepads

Design de SKILL.md, hooks de ciclo de vida (SessionStart, PostToolUse), injeção automática de contexto e o Neuro-Symbolic Context Engine como single source of truth

• SKILL.md: structure, trigger conditions, domain knowledge packaging
• Hooks: SessionStart (pre-load), PostToolUse (observe), pre-commit (validate)
• Context injection: auto-sync, workspace manifest, pre-loaded knowledge digests
• Neuro-Symbolic Context Engine: projectId, activity routing, depth levels
• Knowledge base: contexts, agents, shared infrastructure, auto-generation via MCP
• Self-healing protocol: implement → tsc → vitest → fix loop (max 3 ciclos)

RAGAS (context_precision, faithfulness, answer_relevancy), LLM-as-judge, Expected Calibration Error, detecção de alucinação e benchmarks técnicos

• RAGAS: context_precision, context_recall, faithfulness, answer_relevancy — métricas RAG
• LLM-as-judge: preference modeling, G-eval, scalable oversight para annotation
• Calibração: ECE (Expected Calibration Error), reliability diagrams, temperature scaling
• Detecção de alucinação: factuality scoring, entailment classifiers, SelfCheckGPT
• Benchmarks: MMLU, HELM, BIG-Bench, LMSYS Arena Elo, GAIA, SWE-bench
• Evals framework: promptfoo, LangFuse evals, custom harness com CI integration

LangSmith e Phoenix/Arize para tracing de LLMs, red-teaming adversarial, Constitutional AI, guardrails e estratégias de deployment cost-efficient

• Tracing: LangSmith, Phoenix/Arize — spans, traces, token accounting por request
• Métricas: P95/P99 latência, TTFT (Time-to-First-Token), throughput, tokens/s
• Red-teaming: jailbreaks, indirect injection, data poisoning, model inversion
• Constitutional AI: RLHF com feedback de princípios, harmlessness, helpful, honest
• Guardrails: NeMo Guardrails, Llama Guard 3, Rebuff prompt injection detector
• Deployment: serverless vs batch inference, cost/quality frontier, caching