Tutorials

Field notes on RAG, evals, agents, and LLM cost.

Structured Output That Holds in Production: Schema Design, Validation, and Repair

September 15, 2026 · TokenTailors

Constrained decoding stops malformed JSON. It does not stop wrong JSON. How to design extraction schemas, validate beyond the type system, and repair failures without a retry loop that hides them.

Hybrid Search and Reranking: Where RAG Recall Actually Comes From

September 14, 2026

Dense-only retrieval misses exact terms; BM25-only misses paraphrase. How to combine them, where a reranker earns its latency, and how to measure each stage separately.

Evaluating Agents: Score the Trajectory, Not Just the Final Answer

September 14, 2026 · TokenTailors

Final-answer pass rates hide the ways agents actually fail: wrong tool, right answer by luck, twelve steps where two would do. How to build trajectory evals for LangGraph agents.

Indirect Prompt Injection Is a Data Problem: Containing Untrusted Text in LangGraph Agents

September 14, 2026

Every retrieved document and tool response is untrusted input. Provenance tracking, capability limits, deterministic action gates, and the injection evals that tell you the defenses hold.

Context Engineering for Long-Running Agents: What to Keep, Compact, and Throw Away

September 11, 2026 · TokenTailors

An agent that works for five turns and degrades by thirty has a context problem, not a model problem. Budgets, tool-output pruning, compaction, and external memory in LangGraph — with the evals that tell you it worked.