JianFeeeee
6c2039f5c9
feat(vector): pluggable multimodal vector space
核心暴露 MultimodalEmbedder 接口,两条路径共享同一套 L0/L2/L3
向量缓存、media.Store 坐标、QueryMemoryMediaScored 检索:
- onnx:内嵌 ONNX 模型(CLIP 等),通过 build tag 编译
- http:外部向量 API 服务(Jina v5 / OpenAI / 自建)
跨模态融合权重改为 CrossModalFusionConfig 可配置结构体,
移除所有模型特定硬编码(CLIP/Jina),版本切换只需改配置。
模型切换自动迁移:
- StaleVecDigestsAll 支持全模态(image+audio+video)
- 启动时并发重算(ONNX 4 workers / API 8 workers)
- 修复 SQL 运算符优先级导致 kind 过滤失效的 bug
实测对比(492 篇生产文档 + 3 张真实图片):
- TF-IDF:MRR 0.457(精确匹配快,语义差)
- fastText:MRR 0.530(语义中等,延迟 8ms)
- Jina v5-omni:MRR 0.900(全面领先,延迟 40ms)
- 中文文本→图片:Jina MRR 0.833 vs CLIP 0.611
See docs/embedding-comparison.md for full benchmark.
2026-09-09 17:38:34 +08:00
..
2026-08-05 09:59:33 +08:00
2026-08-14 00:48:40 +08:00
2026-07-29 14:48:23 +08:00
2026-09-09 17:38:34 +08:00
2026-07-25 11:17:31 +08:00
2026-09-04 20:53:32 +08:00
2026-09-09 17:38:34 +08:00
2026-09-05 12:03:30 +08:00
2026-09-06 09:51:31 +08:00
2026-08-18 09:07:21 +08:00
2026-09-05 12:08:33 +08:00
2026-09-06 09:51:31 +08:00
2026-09-06 09:51:31 +08:00
2026-07-28 11:42:29 +08:00
2026-09-05 12:03:30 +08:00
2026-09-04 22:00:03 +08:00
2026-09-09 17:38:34 +08:00
2026-09-04 20:53:32 +08:00
2026-09-09 17:38:34 +08:00
2026-09-04 06:25:51 +08:00
2026-09-04 06:25:51 +08:00
2026-08-31 12:16:39 +08:00
2026-07-16 12:11:16 +08:00
2026-08-18 09:07:21 +08:00
2026-09-04 20:53:32 +08:00
2026-08-26 16:10:02 +08:00
2026-08-18 11:49:57 +08:00
2026-09-03 12:37:58 +08:00
2026-07-24 14:49:08 +08:00
2026-09-03 12:37:58 +08:00
2026-09-03 15:38:31 +08:00
2026-08-25 10:50:37 +08:00
2026-08-26 16:10:02 +08:00
2026-07-27 15:26:23 +08:00
2026-09-06 09:51:31 +08:00
2026-09-09 17:38:34 +08:00
2026-09-06 09:51:31 +08:00