Evaluation

Swapping Embedding Models Without a Retrieval Regression

September 16, 2026 · TokenTailors

A new embedding model tops the leaderboard and someone proposes a re-index. Here is how to decide whether it helps your corpus, and how to run the migration behind a shadow index without a quality cliff.

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.

Bootstrapping an Eval Suite From Your First 50 Production Traces

May 18, 2026

You do not need a benchmark or a labeling team to start evaluating an LLM app. Fifty real traces, honest labels, and a runnable harness beat both.