传统 RAG 处理网页和 PDF，很多时候会先转成文本，但表格、图片、排版一丢，原本的信息结构也跟着没了。
This is a dev post classified by Jev as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
传统 RAG 处理网页和 PDF，很多时候会先转成文本，但表格、图片、排版一丢，原本的信息结构也跟着没了。 PixelRAG 换了个思路：不急着解析文本，而是直接把页面渲染成截图，再用视觉信息做检索。 这样表格、图表、复杂布局都能一起保留下来，AI 看到的更接近原始页面。 官方测试还给出了最高 18.1% 的准确率提升，并声称能把 Agent 的 Token 成本降低约 10 倍。 目前已经预建了 828 万个 Wikipedia 页面索引，还有免费 API 和 Claude 插件可以体验。 做视觉 RAG、网页 Agent 的，可以研究一下。 🔗 https://github.com/StarTrail-org/PixelRAG
Posted by lumxss (24.8k followers) 1 h ago · 8 likes · 484 views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: pixelrag.
More dev work like this
- Meta Closed Muse vs Open Muse. — @0x0SojalSec
- Setting up @executorsh for my coding agent was the best decision ever — @leodev
- Only two days remain to register for the Qoder AI Hackathon at AI Everything Abu Dhabi… — @alibaba_cloud
- Your AI workspace doesn’t need to sit between you and your model provider. — @DanKornas
- Opus 5.5 system prompt is ~0.5M tokens just by itself! — @zainhas
- Djev-Spark (GitHub) figured out how to make DiffusionGemma 26B-A4B behave like Jev. They… — @TeksEdge
- More model choice, delivered fast. GPT-6 Sol and Luna from @OpenAI are now available on… — @gitlab
- Multimodal data analysis involves analyzing data from multiple sources - like audio,… — @freeCodeCamp
Every post is read and classified by Jev (TypeSafe): what it is, which market it belongs to, and whether the link is a real tool. 21.7k posts from 5k X accounts over the last 21 days, 2.5k tools, 12 markets. Collected every 5 minutes, fully re-ranked every hour — last update 2026-09-23 00:57 UTC. Full method.