Jev job filtering

Use Jev to evaluate job descriptions and let code produce a shortlist. The workflow runs locally or on your server using your TypeSafe account. TrueFetch collection is in development, so start with a job JSON file you already have.

Get the files

• Python workflow — complete validation, evaluation and routing.
• Sample input — three fictional records.
• TypeScript API example — one prepared request using Node 24+.
• Full run instructions — input limits, output fields and failure handling.

Save them in one folder. Python 3.11+ needs no additional dependencies.

Prepare your data

python filter_jobs.py prepare jobs.json --output requests.jsonl

This validates input and writes the exact requests without calling Jev. Supply title, sourceUrl and description for each job. location provides work-arrangement context. Missing descriptions go to review. The file accepts 1–25 records; a collection envelope with status: "partial" requires --allow-partial.

Call Jev

Set TYPESAFE_API_KEY in your process environment using your secret manager, then run:

python filter_jobs.py run jobs.json --output decisions.jsonl --target-role "Python backend engineer" --work-mode remote

The four questions share one request per eligible listing. The script pins jev-1.13.0 and records the returned model version. Calls consume your TypeSafe allowance, separately from TrueFetch credits. Never expose the key in browser JavaScript.

For a minimal TypeScript call using the prepared file:

node evaluate.ts requests.jsonl answer.json

This evaluates the first eligible request only and saves raw answers. The Python command provides the complete filtering workflow.

Routing rules

Check	Decision
Possible evaluator instructions in the job text: Noul ≥ 0.2	Review
Evidence of actual duties: Noul < 0.8	Review
Role confidence < 0.8	Review
Role score ≤ 0.5 on a 0–3 scale	Skip
Requested work arrangement is unknown or confidence < 0.8	Review
Clear mismatch with requested work arrangement	Skip
Role score ≥ 2.5, with earlier checks passed	Shortlist
Any remaining case	Review

Rules run from top to bottom. --work-mode any bypasses the arrangement check. Thresholds live in THRESHOLDS near the top of the Python file. They are starting values to evaluate against labeled examples from your own sources. Confidence is a statistic of the model’s distribution, not an accuracy guarantee.

Remote status does not establish country, visa or timezone eligibility. Those restrictions need additional checks. This workflow filters listings; it does not assess people applying for work.

Trace results and handle failures

Each output row retains its source URL and collection status. Successful evaluations include answers, probabilities, thresholds, model, request ID when available, and token usage. Missing usage remains unknown. Read the source before acting on an uncertain result.

Malformed provider output goes to review and causes exit code 2. Configuration or input errors return 1. Existing output files are never overwritten. A failed run can leave completed rows, so avoid resubmitting them automatically. Retries are bounded to 429 and 529; timeouts are not automatically replayed.

Model limits

Jev takes text, not images, audio or video. Keep state focused. Perform arithmetic, exact date comparisons and URL handling in code. The model can be influenced by adversarial content; the injection question helps route suspicious text but cannot guarantee detection.

See the TypeSafe API reference, confidence guide, model limits and known Jev limitations. Continue with the worked tutorial or the Job Scraper contract.