LAYRA: Large Academic Visual RAG (v1.1)
This model entry represents the optimized version of the LAYRA retrieval system, featuring LLM-based reranking.
Performance Summary (v1.1)
- Strategy: Hybrid Retrieval (ColQwen + BM25) $ ightarrow$ LLM Reranking (Qwen3-VL-Cloud)
- Recall@20: 1.00 (Perfect retrieval on Kanna Gold Set)
- MRR@20: 0.740 (Significant improvement over v1.0 baseline of 0.365)
- Recall@1: 0.600 (Correct answer is #1 result 60% of the time)
System Configuration
- Visual Encoder:
vidore/colqwen2.5-v0.2 - Vector DB: Milvus 2.6
- Fusion: RRF (Candidate Pool: 500)
- Reranker:
qwen3-vl-235b-instruct-cloud(Listwise Ranking)
Evaluation
Metrics verified on 2025-12-31 using the SAINTHALF/layra-kanna-goldset repository.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Evaluation results
- Recall@20 on LAYRA Kanna Gold Settest set self-reported1.000
- MRR@20 (SOTA) on LAYRA Kanna Gold Settest set self-reported0.740
- Recall@1 on LAYRA Kanna Gold Settest set self-reported0.600
- Average Latency (s) on LAYRA Kanna Gold Settest set self-reported1.500