We have PDF of graphs, tiles ( dashboard elements ) which keeps changing
once in 24 hours (200 - 300 images). Have to figure out a solution so that we can do similarity
matching `**text-to-image**` matching and then pass that along with query so that
**VLM** hopefully will be able to understand and able to answer the query.
| qwen + jina-clip | Colpali by vidore. | Visualized-BGE by BAAI | ImageBind by MetaAI | |
|---|---|---|---|---|
| Memory requirement | 223M model can run over CPU but old just suitable for smaller text matching tasks ( similar to CLIP arch) | 3B model requires > 40GB ram in case of CPU | <4 GB ram required about 300mb model | 10GB ram required |
| Performance | equivalent performance as of imagebind. | Almost perfect | slightly worse sometimes works sometimes doesn’t | working most of times compared to BGE. |
https://culinda-my.sharepoint.com/:u:/p/somesh/EXIRYlSZ8f9Is2zg33X2RLcBB90QG1IX35drj54EGEGaKw?e=5c7aI3
jina-clip-v1 )