After the basics · Worked guides
From concepts to working designs
After the eight beginner lessons, choose a task. Each guide connects question design, input data, application logic, and evaluation in one workflow.
Reliability and routing
Compose decisions, then validate uncertain cases.
Jev parallel questions: one state, several judgments, explicit routing
Use a repair ticket to separate independent questions from decisions that require new facts, then build a routing policy with a review path.
02 / 14 minutesJev composite scoring: rubrics, normalization, and explainable ranking
Rank tutorial candidates with explicit Score levels, normalization, weights, hard constraints, and missing-value handling.
03 / 16 minutesJev self-consistency and confidence: stable is not necessarily correct
Design repeat tests and abstention for ambiguous subscription requests without mistaking a majority vote for certainty.
Retrieval and evidence
Check candidate retrieval, ranking, and claim support separately.
Jev for RAG: relevance, evidence support, and citation checks
Follow a warranty-policy question through retrieval, filtering, generation, and claim-level verification. Relevance alone is not evidence.
02 / 17 minutesJev reranking and line search: candidates before answer locations
Use an offline PDF-export query to separate retrieval, reranking, line selection, and answer availability.
Structured data
Preserve source text and candidates while resolving structure, roles, and labels.
Jev structure recovery: preserve text, rebuild Markdown
Restore a damaged event notice using line joins, block labels, provenance, and safe rendering.
02 / 16 minutesJev entity alignment: aliases, names, and conflicting specifications
Compare multilingual product records without treating similar products as identical entities.
03 / 18 minutesJev date and candidate extraction: roles before normalization
Select spans from an email with several dates and amounts, then handle locale, reference time, and missing values in code.
04 / 17 minutesJev hierarchical classification: paths, alternatives, and parent fallback
Classify tutorials with a versioned taxonomy, greedy search, limited beam exploration, and coarse-label fallback.
Agents and risk controls
Understand tool gates before designing skill selection and input/output checks.
Jev agent tool calls: selection, arguments, permissions, and fallback
Build a calendar-assistant policy that separates user intent, proposed actions, authorization, and execution.
02 / 15 minutesJev agent skill routing: retrieve, verify, or load nothing
Distinguish creating a slide deck from editing an existing one using candidate retrieval and applicability checks.
03 / 17 minutesJev input/output guardrails: risk signals into explicit policy
Design pass, review, block, and failure behavior for a support assistant’s inputs, retrieved text, and outputs.
Cascades and research experiments
After evaluation basics, try field verification and feature iteration.
Jev structured extraction cascades: extract, verify, escalate
Validate event-registration fields before accepting small-model output or escalating to a stronger process.
02 / 20 minutesJev text feature discovery: semantic questions into testable experiments
Build versioned features for issue priority while controlling baselines, grouped splits, iteration, and target leakage.
Where to learn the 18 recipe topics
This map connects original guides here with official recipes and community translations. It tracks topic coverage, not notebook reproduction or execution. Check source recipes for their methods and results.
Sources and further reading
Inspired by official patterns and Datawhale practice topics. Explanations, examples, and exercises are independently written. These are teaching designs, not live API runs or benchmarks.
Datawhale Jev Cookbook