Aggretriever is an encoder to aggregate both lexical and semantic text information into a single-vector dense vector for dense retrieval, which is finetued on MS MARCO corpus with BM25 negative sampling, following the approach described in [Aggretriever: A Simple Approach to Aggregate Textual Representation for Robust Dense Passage Retrieval](https://arxiv.org/abs/2208.00511).

<p align="center">
<img src="https://raw.githubusercontent.com/castorini/dhr/main/fig/aggretriever_teaser.png" width="600">
</p>

The associated GitHub repository for fine-tuning is available [here](https://github.com/castorini/dhr) and the reproduce from pyserini is [here]. The following variants are also available:

Model | Initialization | MARCO Dev | Encoder Path
|---|---|---|---
aggretriever-distilbert | distilbert-base-uncased | 34.1 | [castorini/aggretriever-distilbert](https://huggingface.co/castorini/aggretriever-distilbert)
aggretriever-cocondenser | Luyu/co-condenser-marco | 36.2 | [castorini/aggretriever-cocondenser](https://huggingface.co/castorini/aggretriever-cocondenser)

## Usage (HuggingFace Transformers)
Using the model directly available in HuggingFace transformers. We use the implemented Aggretriever from pyserini [here](https://github.com/castorini/pyserini/blob/master/pyserini/encode/_aggretriever.py).

```python
from pyserini.encode._aggretriever import AggretrieverQueryEncoder
from pyserini.encode._aggretriever import AggretrieverDocumentEncoder

model_name = '/store/scratch/s269lin/experiments/aggretriever/hf_model/aggretriever-cocondenser'
query_encoder = AggretrieverQueryEncoder(model_name, device='cpu')
context_encoder = AggretrieverDocumentEncoder(model_name, device='cpu')

query =  ["Where was Marie Curie born?"]
contexts = [
    "Maria Sklodowska, later known as Marie Curie, was born on November 7, 1867.",
    "Born in Paris on 15 May 1859, Pierre Curie was the son of Eugène Curie, a doctor of French Catholic origin from Alsace."
]

# Compute embeddings: take the last-layer hidden state of the [CLS] token
query_emb = query_encoder.encode(query)
ctx_emb = context_encoder.encode(contexts)
# Compute similarity scores using dot product
score1 = query_emb @ ctx_emb[0]  # 45.56658
score2 = query_emb @ ctx_emb[1]  # 45.81762
```