--- title: prompt_jev() now works on images | daily.dev description: MotherDuck's prompt_jev() SQL function now accepts images, not just text, letting developers store images as BLOB columns and ask closed-ended questions... canonical: https://daily.dev/posts/prompt-jev-now-works-on-images-y3go2ruuj twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: prompt_jev() now works on images | daily.dev og:description: MotherDuck's prompt_jev() SQL function now accepts images, not just text, letting developers store images as BLOB columns and ask closed-ended questions... og:url: https://daily.dev/posts/prompt-jev-now-works-on-images-y3go2ruuj og:image: https://api.daily.dev/og/posts/y3Go2ruuJ.png og:image:alt: prompt_jev() now works on images og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # prompt_jev() now works on images **[MotherDuck](https://daily.dev/sources/motherduck)** · 5 min read · 4 upvotes · 0 comments ## Summary MotherDuck's prompt_jev() SQL function now accepts images, not just text, letting developers store images as BLOB columns and ask closed-ended questions directly in SQL with typed answers and confidence scores. The post shows image classification on Food-101 photos, benchmarks Jev against Gemini, GPT, and Claude on speed, accuracy and cost (Jev is fastest and cheapest at 90.2% accuracy vs 93-97% for the LLMs, but hits 97.7% accuracy when filtering for confidence >= 0.8), and demonstrates a large-scale use case analyzing 77k Metropolitan Museum artworks with 14 SQL-based visual questions per image, joined with catalog metadata via GROUP BY. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How accurate is MotherDuck's prompt_jev() compared to Gemini, GPT, and Claude for image classification? On the Food-101 dataset of 1,010 dish photos, prompt_jev() reached 90.2% accuracy at 44 rows/second for about $0.06 total. Gemini 3.8 Flash scored 96.8-97.0%, GPT 6.1 Sol scored 93.4%, and Claude Opus 5.5 scored 95.7%, but all three cost between $1.14 and $6.60 and ran at 7-10 rows per second. _Teams choosing a model for bulk image labeling can weigh these cost and accuracy trade-offs on daily.dev._ ### How do I filter prompt_jev() results by confidence score in MotherDuck SQL? Add a WHERE clause like WHERE jev.confidence >= 0.8 to keep only high-confidence answers. On the Food-101 benchmark, filtering to confidence scores of 0.8 or higher kept 80% of photos while raising accuracy from 90.2% to 97.7%, since the confidence score tracks how often the answer is actually correct. _Developers tuning SQL-based AI pipelines can track techniques like confidence filtering on daily.dev._ ### What image size does prompt_jev() in MotherDuck require? Each side of an image must be between 280 and 1,568 pixels for prompt_jev() to process it. Larger images need resizing first, which can be done using a MotherDuck Flight (Python running on MotherDuck compute) with the Pillow library; MotherDuck's docs provide a ready-to-use example that reads from S3, resizes oversized images, and writes them back as BLOBs. _Anyone wiring up image preprocessing for SQL-based AI functions can find setup details on daily.dev._ ## Similar posts on daily.dev - [5 Jev use cases for analytics, with real queries and datasets](https://daily.dev/posts/5-jev-use-cases-for-analytics-with-real-queries-and-datasets-xijgxxj63) · MotherDuck · 23 upvotes · 1 comments - [We need to talk about JEV \(in data pipelines\)](https://daily.dev/posts/we-need-to-talk-about-jev-in-data-pipelines--nr914jn81) · Medium · 3 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#sql](https://daily.dev/tags/sql), [#data-engineering](https://daily.dev/tags/data-engineering), [#duckdb](https://daily.dev/tags/duckdb) [View this post on daily.dev](https://daily.dev/posts/prompt-jev-now-works-on-images-y3go2ruuj) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"prompt_jev() now works on images","url":"https://daily.dev/posts/prompt-jev-now-works-on-images-y3go2ruuj","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/prompt-jev-now-works-on-images-y3go2ruuj"},"datePublished":"2026-10-16T15:48:18.118Z","dateModified":"2026-10-16T15:51:00.510Z","description":"MotherDuck's prompt_jev() SQL function now accepts images, not just text, letting developers store images as BLOB columns and ask closed-ended questions...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/20f9c16c0814159dc6c7dd22d84bb933?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/20f9c16c0814159dc6c7dd22d84bb933?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"MotherDuck","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"MotherDuck","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/f084d6835eef4c2b83f889f9a4e85ed6","url":"https://daily.dev/sources/motherduck"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/prompt-jev-now-works-on-images-y3go2ruuj","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,sql,data-engineering,duckdb","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"MotherDuck","item":"https://daily.dev/sources/motherduck"},{"@type":"ListItem","position":3,"name":"prompt_jev() now works on images"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/prompt-jev-now-works-on-images-y3go2ruuj#faq","mainEntity":[{"@type":"Question","name":"How accurate is MotherDuck's prompt_jev() compared to Gemini, GPT, and Claude for image classification?","acceptedAnswer":{"@type":"Answer","text":"On the Food-101 dataset of 1,010 dish photos, prompt_jev() reached 90.2% accuracy at 44 rows/second for about $0.06 total. Gemini 3.8 Flash scored 96.8-97.0%, GPT 6.1 Sol scored 93.4%, and Claude Opus 5.5 scored 95.7%, but all three cost between $1.14 and $6.60 and ran at 7-10 rows per second. Teams choosing a model for bulk image labeling can weigh these cost and accuracy trade-offs on daily.dev."}},{"@type":"Question","name":"How do I filter prompt_jev() results by confidence score in MotherDuck SQL?","acceptedAnswer":{"@type":"Answer","text":"Add a WHERE clause like WHERE jev.confidence >= 0.8 to keep only high-confidence answers. On the Food-101 benchmark, filtering to confidence scores of 0.8 or higher kept 80% of photos while raising accuracy from 90.2% to 97.7%, since the confidence score tracks how often the answer is actually correct. Developers tuning SQL-based AI pipelines can track techniques like confidence filtering on daily.dev."}},{"@type":"Question","name":"What image size does prompt_jev() in MotherDuck require?","acceptedAnswer":{"@type":"Answer","text":"Each side of an image must be between 280 and 1,568 pixels for prompt_jev() to process it. Larger images need resizing first, which can be done using a MotherDuck Flight (Python running on MotherDuck compute) with the Pillow library; MotherDuck's docs provide a ready-to-use example that reads from S3, resizes oversized images, and writes them back as BLOBs. Anyone wiring up image preprocessing for SQL-based AI functions can find setup details on daily.dev."}}]} ```