把 Jev 接进两个真实产品：文档库与家庭病历夹

Josh Pigford @Shpigford原推发布于 2026.09.22

在 X 上打开原推

Lots of cool experimental Jev (@typesafeai) stuff getting posted lately, but what about using it in an existing product? Here are dozens of ways I'm using it now in two apps (http://granite.co and http://keptwell.org) Granite (document vault) Ingest pipeline - Second-opinion on Gemini's document classification, flags low-confidence ones for review - Scores PDF text-layer quality and routes bad ones to OCR - Verifies each extracted field against the page text - Detects what a document asks you to do (pay, sign, renew, respond) and how urgent it is - Judges whether two near-duplicate documents are the same or a revised version Document page and library - "Needs review" banner with one-tap confirm of the document type - "Not confirmed" marker on extracted values Jev couldn't verify - Action chip next to the document type Collections - Plain-English filing rules ("anything to do with my taxes") that auto-file matching documents Life View and email digests - "Needs your attention" block listing documents that require action Ask - Routes each question to the right tool path instead of a regex - Checks the final answer is supported by the cited passages and hedges if not Entity graph - Tiebreaker on whether two fuzzy-matched names are the same entity Evernote import - Triages each note into keep, reference, scratch, or clutter before import KeptWell (family medical binder) Trust - Verify every extracted lab value, dose, diagnosis, and provider against the source page - Flag values it cannot confirm with a quiet "check this" marker - Catch diagnoses stated more precisely than the page says - Second-opinion the document type after extraction - Gate prompt PRs with a cheap eval-corpus canary Attention - Tag new documents: new diagnosis, out-of-range result, med change, follow-up, act-within-7-days, admin-only - Order the dashboard feed by importance, not recency - Decide push-now versus digest per notification - Pick push wording from the PHI-free string set - Decide which lab trends are worth an Insight before calling Opus Chat - Detect emergency or distress before the model streams a token - Route docs-only questions away from paid web search - Rerank retrieved chunks against the question - Pick between two contradicting family facts - Filter PHI-audit false positives ("Ray" in "x-ray") - Check the answer is grounded in the cited record Binder - Tag every document by body system, specialty, and care phase for filters - Flag near-duplicate uploads for review - Break ties on whether a PDF text layer is usable Recordings and journal - Label each recording utterance and build an action-item checklist - Score journal entries on a symptom rubric for trend charts - Flag entries that look like a medication side effect - Decide whether an undated entry describes a specific past day - Replace the async journal tag job with one sync call Terminology and imports - Auto-pick clinical codes above a confidence bar; queue the rest - Replace the Sonnet pick in disambiguation - Decide which FHIR observations are real lab results - Map vital types the LOINC table drops - Merge brand and generic med names ("Lipitor" and atorvastatin) - Pick the right NPI when the registry returns several - Classify severity for manually entered diagnoses - Catch allergy denials the regexes miss Cost gates - Skip the highlight call when nothing is worth highlighting - Skip reprocessing documents a prompt change would not affect - Route extraction to batch or sync by urgency - Tell a bulk import from a runaway loop at the spend cap - Flag uploads containing instructions aimed at an AI Guards - Veto preventive reminders the record contradicts - Suppress marketing emails during a hard week - Rank appointment prep context by relevance - Auto-resolve visit questions the visit log answers - Mark share comments that are waiting on a reply