Jev Patterns and Cookbooks

Official docs split "how to build" into patterns and cookbooks. Patterns are the architecture. Cookbooks are end-to-end recipes with their own datasets. This hub explains the patterns. It does not re-run the cookbook tables.

Patterns

Pattern	Official page	Hub page
Speculative fan-out	fan-out	Decision layer and Ultrafast architecture
Confidence-gated routing	confidence-routing	Confidence-gated routing
Composite scoring	composite-scoring	Composite scoring
Intent routing	intent-routing	Intent detection and Model routing

Fan-out means one request can carry questions you might ignore. The questions share the state, run together, and cannot see each other. Your code throws away the branches that did not apply. The browser-agent loop uses that for "which control?" beside "which operation?"

Cookbooks worth opening next

The blurbs are the descriptions on TypeSafe's docs index, fetched 2026-09-30. Follow the link for the recipe.

Cookbook	What the official index says	Read beside it
Parallel questions	One call with 13 questions over a Wikipedia article, reported 12.2× cheaper and 10.0× faster than separate calls, with the same answers	API batching notes
Re-ranking	One question per query-candidate pair over BM25 shortlists	RAG filtering
Classifying RAG passages	Score each passage, then decide in code what reaches the answering model	RAG filtering
Citation check	One Choice on whether the quoted context supports the claim	Limitations on adversarial state
Guardrails	One request screens messages in and out of an LLM app	Content moderation
SDE cascade	A small extract, a verify step, then a reasoning model only when needed	Confidence-gated routing
Date extraction	Ask for the date parts named in the document, then resolve them in code	Limitations
Pre-parsed value extraction	Regex finds candidate spans. Jev selects one. Code copies it verbatim	Limitations on generation
Function calling	Map a request onto a function name and closed-set arguments	Agent tool selection
Skill suggestion	Pick at most one skill from a catalog, in two requests	Coding agents
Hierarchical classification	Beam search over Choice probabilities down a deep label tree	Classification
Classification using confidence	Report the leaf only when confidence holds; otherwise report the parent	Confidence-gated routing
Entity alignment	One Score plus companion Nouls on whether two records are the same product	Composite scoring
Structure recovery	Classify blocks and stitch hard-wrapped lines back into Markdown	Two requests, both closed label sets
Line-by-line search	Score line ids with a Choice, and use a Noul to ask whether the document contains an answer	Choice
Self-consistency	Noul and Choice cookbooks that keep uncertain cases visible for review	Noul
Feature discovery	Propose questions, turn them into features, fit a classical regressor	Composite scoring

The parallel-questions speedup is TypeSafe's measurement on that article and those 13 questions. Quote it with the cookbook, not as a property of every workload.

Dates, spans, and other things code should finish

Jev reads dates as text. The date cookbook's move matches the limitations page: ask which month, day, and year are written, including "not stated," then build the date in code and reject what does not validate.

The same split covers emails, amounts, and phone numbers. A regex proposes spans. A Choice picks the span that answers the question. Your code copies that span and normalizes it. The model never has to emit the characters.

Cascades

A cascade is confidence routing across models. A cheap typed call handles the clear cases. A generator, or a person, sees the rest. The SDE cookbook is TypeSafe's version for structured extraction. vercel-labs/jev-ai-sdk-form-router is a public form that keeps Jev's choice only when its confidence is at least 95% and otherwise calls a text model. That 95% is their application constant. It is not a TypeSafe default.