Connect the dots. Build with understanding.
Start with model fundamentals, move through retrieval and agent state, and work toward deployment. Follow a path or find a specific problem.
Models & APIs
Understand model limits, build a reliable interface, and choose with evidence.
Explore the seriesRAG & retrieval
Turn documents into retrievable evidence and evaluate the answers they support.
Explore the seriesAgent engineering
Design workflows, tool contracts, and durable state for tasks that can recover.
Explore the seriesShipping AI apps
Handle security, latency, and failure before putting an AI application into use.
Explore the seriesAll guides
12 articlesUnderstand context before your first model call
Separate tokens, context windows, and output budgets, then practice assembling model input with an offline example.
Streaming and structured output solve different problems
Separate incremental display from a complete business result and validate output with a local Python example.
Build a task evaluation set before choosing a model
Define success with actual tasks and compare quality, latency, and failure modes rather than a leaderboard alone.
The smallest RAG loop: retrieve, cite, answer
Separate retrieval from generation and use offline keyword retrieval to understand provenance and missing evidence.
Chunk documents for meaning before tuning an index
Understand chunking, overlap, metadata, and approximate vector indexes using a small windowing example.
Evaluate retrieval and answers separately in RAG
Use Recall@k, evidence support, and unanswerable cases to locate failures in a RAG pipeline.
When an agent helps and when a workflow is enough
Decide using control flow, task uncertainty, and failure cost, then separate model decisions from program rules.
Tool calling and MCP still need execution boundaries
Define input, output, error, and permission contracts and practice with an offline tool dispatcher.
Recoverable agents: state, retries, and idempotency
Use SQLite to practice durable task transitions and distinguish conversation history from external side effects.
Prompt injection: keep untrusted content inside data boundaries
Recognize indirect instructions in documents and tool results and enforce permissions in application code.
Observe AI apps: first-token latency, total time, and cost
Separate retrieval, model, and tool latency, calculate percentiles, and record request-level telemetry.
From demo to deployment: design the failure paths
Check timeouts, rate limits, missing evidence, and tool failures using a practical deployment matrix.
No matching articles. Try a shorter query or browse another series.
The next guide, in your feed reader.
Follow new guides through RSS and learn at your own pace.