What needs the closest review in this change?
Token validation · retry behavior · rollback plan
What is Jev AI? Start with a structured decision
Learn how Jev uses state, typed questions, and probabilities; follow source-backed examples and understand the limits before a real model call.
119Guide topics · sources & verification kept
Practical Jev guides
Follow practical guides and open-source walkthroughs with local sample data.
How Jev AI turns a question into a decision your code can use
01 / STATE
Release change: the login endpoint adds token checks and changes its retry behavior.
Return one option with its probabilities
Code checks confidence, then routes automatically or sends the case for human review.
Illustrative workflow only, not a live call or model evaluation. Application code decides what action to take.
Three structured question types in Jev AI
Read the beginner guide →Choose from options you define
Example: Is this ticket about billing, shipping, or a technical issue?
Judge a yes-or-no question
Example: Does this request need human review?
Score against an ordered rubric
Example: Rate ticket urgency as low, medium, or high.
These types constrain the shape of an answer. You still need representative examples to evaluate judgment quality and code to enforce review and action boundaries.
For the model definition and current capability boundaries, refer to TypeSafe AI’s primary sources: Official introduction · Confidence guide
Where can Jev AI fit?
Browse all use cases →Start with a repeated decision and a clearly bounded answer. Review the examples and sources before deciding whether it fits your workflow.
From understanding Jev to your first walkthrough
What is Jev AI?
Understand its inputs, question types, probabilities, and where it fits alongside language models.
Read the five-question guide →Walk through a first decision
Use a fictional release change to follow state, model, questions, and human fallback.
Read the first-call guide →Find examples by use case
Browse open-source examples and exercises for ticket routing, data classification, and agent safeguards.
Browse all examples →Free CSV tools with worked examples
No login required. Files are processed locally in your browser. Each tool has downloadable teaching inputs and expected reports, without a model call.
Check CSV blanks, duplicates and columns
Find 5 issues in 7 sample records, then check the repaired fixture.
Open the CSV checker →Read the worked exampleCompare two classification CSVs
Match stable IDs to find changed, missing and added records.
Open the comparison tool →Learn how to interpret the differencesRoute low scores and missing labels
Route 5 results into 3 reviews and 2 spot checks, retaining input columns.
Open the review queue →Read threshold and handoff stepsLatest practical guides
Open all cases →
Original guideDesign atomic Jev questions for reviewable routing
Separate category, evidence sufficiency and impact for fictional product reports, then combine the answers in code. Includes a specification and edge cases.
Original guideChoose Jev confidence thresholds with a labeled CSV
Compare automatic coverage and accepted errors across candidate cutoffs, inspect high-confidence mistakes and reserve a validation set. Includes local calculation code.
Original guideWrite a reviewable Jev Score rubric for issue impact
Replace vague severity labels with observable anchors. Separate impact and evidence, retain fractional scores and human expectations, and produce reviewable priorities.
Original guideHandle Jev Python timeouts, retries and error states
Configure bounded SDK retries, distinguish authentication, timeout and response validation errors, retain request IDs and persist task states for refresh and duplicate submissions.
Original guideFind faulty CSV rows with DuckDB reject tables
Fix CSV types, collect rejected row locations and reasons with store_rejects, export usable data and errors separately, and reconcile counts with a malformed fixture.
Original guideExport PDF tables with Docling and retain provenance
Export candidate PDF tables as CSV and HTML, retain source hashes and table indices, and inspect merged cells, repeated headers and numeric columns using a checklist.
Original guidePause and resume LangGraph approval without duplicate actions
Use interrupt and Command for approval, rejection and resumption. Distinguish thread IDs, in-memory checkpoints and durable storage, with an audit-record checklist.
Original guideCheck saved outputs with Promptfoo CSV regression fixtures
Use __expected exact assertions and a local JavaScript provider to check saved outputs. An intentional failure verifies reporting without an LLM judge. Includes a complete fixture archive.
Original guideHandle n8n failures with error queues and bounded retries
Collect automatic failures with Error Trigger, classify credentials, rate limits and validation, and retain execution IDs and completed steps. Includes an error-register CSV.
Original guideDebug browser automation with Playwright traces
Record actions and DOM snapshots on a local fixture, inspect locators and waits in trace.zip, and write a minimal reproduction. Includes a local-page exercise archive.
How does Jev work with generative language models?
Handle bounded decisions
When software needs to classify, choose a route, score, or decide whether review is needed, Jev can return a decision in a predefined type with probabilities.
Create text and open-ended answers
Use a generative model for replies, explanations, code, or long-form text. A workflow can also use one to write while Jev handles bounded decisions inside it.
Keep rules, permissions, and actions in code
Keep arithmetic, permissions, and sending decisions in ordinary code and business rules. Route low-confidence results to people according to your workflow.
Jev handles a specific kind of judgment rather than replacing every AI tool. Choose based on whether the next step needs a bounded answer or newly generated content.
Jev AI frequently asked questions
Is Jev AI a chatbot?
Jev is designed for structured decisions inside software and returns results in predefined types. Generative models remain a better fit for writing, conversation, and long explanations.
Does Jev guarantee a correct decision?
No. A well-formed result is not necessarily semantically correct. Evaluate on representative examples and keep human review for uncertain or consequential cases.
Can I try the live Jev API on JevLog?
No. This site provides explanations, local examples, and workflow diagrams. Live API access requires authorization from TypeSafe AI.
Where should I start learning Jev?
Start with the five-question introduction, then follow the fictional release-change walkthrough. Browse examples by ticket routing, agent safeguards, or data classification next. Open the beginner guide →