---
license: mit
library_name: colbert
tags:
- rag
- visual-rag
- colqwen
- sceletium-tortuosum
- thesis
model-index:
- name: LAYRA v2
  results:
  - task:
      type: retrieval-augmented-generation
      name: Retrieval Augmented Generation
    dataset:
      name: Kanna RAG Gold Standard
      type: SAINTHALF/kanna-rag-gold-standard
    metrics:
    - name: MRR@20
      type: mrr
      value: 0.7403174603174603
    - name: Recall@20
      type: recall
      value: 1.0
---

# LAYRA v2 (Large Academic Visual RAG Agent)

**LAYRA v2** is a specialized Visual RAG system designed for the ethnopharmacology of *Sceletium tortuosum*. It processes full PDF pages as images, preserving layout and visual information, and uses a hybrid retrieval stack.

## Architecture

- **Visual Encoder:** ColQwen 2.5 (ColBERT + Qwen-VL)
- **Retrieval:** Hybrid (Sparse BM25 + Dense ColBERT)
- **Reranking:** LLM-based Reranking (Generative Listwise)
- **Vector DB:** Milvus 2.5.5
- **Infrastructure:** Docker Compose (Lean Stack)

## Performance

Evaluated on [SAINTHALF/kanna-rag-gold-standard](https://huggingface.co/datasets/SAINTHALF/kanna-rag-gold-standard):

| Metric | Score |
| :--- | :--- |
| **MRR@20** | **0.7403** |
| **Recall@20** | **1.0000** |
| **Latency (P50)** | 1.2s |

## Usage

This model represents the deployed system configuration described in the thesis.

*Supersedes `SAINTHALF/layra-v1-hybrid`.*