rag25-doc-segmented-test-nist.arctic-embed-l.parquet.shard09.flat.onnx.template

# Anserini Regressions: TREC 2025 RAG Track Test Topics

**Model**: Snowflake's [Arctic-embed-l](https://huggingface.co/Snowflake/snowflake-arctic-embed-l) with flat indexes (using ONNX for on-the-fly query encoding)

This page describes regression experiments for ranking _on the segmented version_ of the MS MARCO V2.1 document corpus using the RAG 25 test topics (= narratives), which is integrated into Anserini's regression testing framework.
This corpus was derived from the MS MARCO V2 _segmented_ document corpus and re-used for the TREC 2025 RAG Track.
Instructions for downloading the corpus can be found [here](https://trec-rag.github.io/annoucements/2025-rag25-corpus/).

We build on embeddings made available by Snowflake on [Hugging Face Datasets](https://huggingface.co/datasets/Snowflake/msmarco-v2.1-snowflake-arctic-embed-l), which contains vectors already encoded by the Arctic-embed-l model.
The complete dataset comprises 60 parquet files (from `00` to `59`).
Due to its large size (472 GB), we have divided the vectors into ten shards, each comprised of six files:
for example `shard00` spans `00.parquet` to `05.parquet`; `shard01` spans the next six parquet files, etc.

This page documents experiments for `shard09`; we expect the corpus to be in `msmarco_v2.1_doc_segmented.arctic-embed-l/shard09` (relative to the base collection path).
In these experiments, we are performing query inference "on-the-fly" with ONNX, using flat vector indexes.

Evaluation uses qrels over 22 topics from the TREC 2025 RAG Track test set.
These qrels represent manual relevance judgments from NIST assessors, contrasted with automatically generated UMBRELA judgments.
More details can be found in the following paper:

> Shivani Upadhyay, Nandan Thakur, Ronak Pradeep, Nick Craswell, Daniel Campos and Jimmy Lin. [Overview of the TREC 2025 Retrieval Augmented Generation (RAG) Track.](https://arxiv.org/abs/2603.09891) _arXiv:2603.09891_, March 2026.

The exact configurations for these regressions are stored in [this YAML file](${yaml}).
Note that this page is automatically generated from [this template](${template}) as part of Anserini's regression pipeline, so do not modify this page directly; modify the template instead.

From one of our Waterloo servers (e.g., `orca`), the following command will perform the complete regression, end to end:

```bash
bin/run.sh io.anserini.reproduce.ReproduceFromDocumentCollection --index --verify --search --config ${test_name}
```

## Indexing

Typical indexing command:

```bash
${index_cmds}
```

The setting of `-input` should be a directory containing the compressed `jsonl` files that comprise the corpus.

For additional details, see explanation of [common indexing options](${root_path}/docs/common-indexing-options.md).

## Retrieval

Here, we are using all 22 test topics from the TREC 2025 RAG Track with (automatically generated) UMBRELA relevance judgments.
Topics and qrels are stored [here](https://github.com/castorini/anserini-tools/tree/master/topics-and-qrels), which is linked to the Anserini repo as a submodule.

After indexing has completed, you should be able to perform retrieval as follows:

```bash
${ranking_cmds}
```

Evaluation can be performed using `trec_eval`:

```bash
${eval_cmds}
```

## Effectiveness

With the above commands, you should be able to reproduce the following results:

${effectiveness}