Getting Started with Jev and SerpApi: A practical fact-checker tutorial

# What Jev is and why it's different

Jev is TypeSafe's System One model designed for fast, structured decision-making. Unlike a large language model that generates free-form prose, Jev returns a decision in a fixed schema and attaches confidence probabilities. It accepts plain text, JSON objects, and arrays and supports three decision types: Choice, Score, and Noul.

# What the tutorial builds

The example project builds a fact-checker that evaluates a claim using fresh web evidence. Workflow in two API calls:

• Use SerpApi to fetch Google organic results for the user's query.
• Send the user input plus the returned titles, links, and snippets to Jev through OpenRouter and receive a verdict.

The verdict set for this tutorial is: supported, contradicted, mixed, or insufficient_evidence. For yes/no questions, supported maps to evidence for "yes," contradicted maps to evidence for "no." For general statements the verdict indicates whether the snippets support the statement.

# Practical workflow and data flow

1. 1Receive user claim or question. Example used in the tutorial: "Did Marie Curie win two Nobel Prizes?"
2. 2Call SerpApi's Google search endpoint to fetch organic results (titles, links, snippets). Keep the returned items alongside the user query.
3. 3Build a structured input that includes the question and the search snippets and send that payload to Jev via OpenRouter.
4. 4Jev responds with a Choice verdict and probabilities. Store or display the verdict with the accompanying search results.

This flow keeps the raw search evidence available for inspection while letting Jev produce a concise decision your application can act on.

# Setup and implementation notes

• Language and runtime: Python 3.10 or newer.
• Dependencies and commands: the tutorial references a repo on GitHub and a dependency install step (the text shows an install command that includes a sync/locked step). Follow the repository for exact commands and files.
• Accounts and keys: you need an active SerpApi account to call Google organic results and an OpenRouter API key with access to Jev.
• Integration: the tutorial uses Python and calls Jev through OpenRouter. The project demonstrates how to format the search results and the question for Jev's Choice decision type.

# Verdict meanings and how to use them in code

• supported: snippets support the claim (for yes/no, supports yes).
• contradicted: snippets argue against the claim (for yes/no, supports no).

• insufficient_evidence: the search snippets do not provide enough information to decide.

Use the returned probabilities to drive thresholds in your application (for example, only label strongly supported when probability exceeds your chosen cutoff). Keep the original snippets available so a human can inspect sources when needed.

# Where to find the code and next steps

The tutorial points to a GitHub repository containing the full code and example project. The implementation focuses on using Jev for decision outputs and SerpApi for fresh web evidence. From here, teams can adapt the input formatting, extend the set of verdict labels, or incorporate additional search sources supported by SerpApi.