---
license: apache-2.0
tags:
  - retrieval
  - search
  - lightgbm
  - cross-encoder
  - bge-m3
language:
  - en
  - ar
---

# Harrir Search Stack v1

Two-tower asymmetric GCN search for fashion retail (Harrir catalog ~28k SKUs),
multilingual EN+AR. Three artifacts, version-locked together.

## Layout

```
.
├── adapter/                      # LoRA adapter on BGE-M3 (LIVE)
├── tokenizer/                    # BGE-M3 tokenizer (LIVE)
├── gcn_head.pt                   # W_q / W_p / W_img projection heads (LIVE)
├── ltr_model.txt                 # LightGBM LambdaRank champion (LIVE)
├── ltr_idf.json                  # BM25/TF-IDF stats over catalog
├── ltr_subcat_*.{json,pt}        # subcategory embeddings (EN+AR)
├── ltr_spec_*.{json,pt}          # product spec embeddings
└── cross_encoder/                # BAAI/bge-reranker-base fine-tuned
                                  # NOT loaded by the app today (served from
                                  # Modal). Kept here for future bake-in.
```

## Usage

The Harrir app loads these at startup. Set this repo in `.env`:

```
GCN_HF_REPO=rdxtremity/search-stack-v1
GCN_ARTIFACTS_DIR=./models/gcn_stage2
```

Then `python download_models.py` snapshot-downloads into `./models/gcn_stage2/`.

Or manually:

```python
from huggingface_hub import snapshot_download
snapshot_download("rdxtremity/search-stack-v1", local_dir="./models/gcn_stage2")
```

## Eval

Baseline nDCG (golden_v1, 346 queries / 19,923 graded pairs): **0.7695**
Latest with LTR+CE rerank: see `Primary.GateRuns` audit log in the app.