Skip to content

rag25-doc-segmented-test-umbrela2.arctic-embed-l.parquet.shard08.flat.onnx.md

Latest commit

 

History

History
84 lines (61 loc) · 5.79 KB

rag25-doc-segmented-test-umbrela2.arctic-embed-l.parquet.shard08.flat.onnx.md

File metadata and controls

84 lines (61 loc) · 5.79 KB

Anserini Regressions: TREC 2025 RAG Track Test Topics

Model: Snowflake's 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.

We build on embeddings made available by Snowflake on Hugging Face Datasets, 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 shard08; we expect the corpus to be in msmarco_v2.1_doc_segmented.arctic-embed-l/shard08 (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. 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. arXiv:2603.09891, March 2026.

The exact configurations for these regressions are stored in this YAML file. Note that this page is automatically generated from this 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:

bin/run.sh io.anserini.reproduce.ReproduceFromDocumentCollection --index --verify --search --config rag25-doc-segmented-test-umbrela2.arctic-embed-l.parquet.shard08.flat.onnx

Indexing

Typical indexing command:

bin/run.sh io.anserini.index.IndexFlatDenseVectors \
  -threads 6 \
  -collection ParquetDenseVectorCollection \
  -input /path/to/msmarco-v2.1-doc-segmented-shard08.arctic-embed-l \
  -generator DenseVectorDocumentGenerator \
  -index indexes/lucene-flat.msmarco-v2.1-doc-segmented-shard08.arctic-embed-l \
  -docidField doc_id -vectorField embedding -normalizeVectors \
  >& logs/log.msmarco-v2.1-doc-segmented-shard08.arctic-embed-l &

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.

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, which is linked to the Anserini repo as a submodule.

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

bin/run.sh io.anserini.search.SearchFlatDenseVectors \
  -index indexes/lucene-flat.msmarco-v2.1-doc-segmented-shard08.arctic-embed-l \
  -topics tools/topics-and-qrels/topics.rag25.test.jsonl \
  -topicReader JsonString \
  -output runs/run.msmarco-v2.1-doc-segmented-shard08.arctic-embed-l.arctic-embed-l-flat-onnx.topics.rag25.test.jsonl.txt \
  -topics rag25.test -topicReader JsonString -topicField title -encoder ArcticEmbedLEncoder &

Evaluation can be performed using trec_eval:

bin/trec_eval -c -m ndcg_cut.30 tools/topics-and-qrels/qrels.rag25.test-umbrela2.txt runs/run.msmarco-v2.1-doc-segmented-shard08.arctic-embed-l.arctic-embed-l-flat-onnx.topics.rag25.test.jsonl.txt
bin/trec_eval -c -m ndcg_cut.100 tools/topics-and-qrels/qrels.rag25.test-umbrela2.txt runs/run.msmarco-v2.1-doc-segmented-shard08.arctic-embed-l.arctic-embed-l-flat-onnx.topics.rag25.test.jsonl.txt
bin/trec_eval -c -m recall.100 tools/topics-and-qrels/qrels.rag25.test-umbrela2.txt runs/run.msmarco-v2.1-doc-segmented-shard08.arctic-embed-l.arctic-embed-l-flat-onnx.topics.rag25.test.jsonl.txt

Effectiveness

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

nDCG@30 ArcticEmbedL
RAG 25: Test queries 0.2362
nDCG@100 ArcticEmbedL
RAG 25: Test queries 0.1526
R@100 ArcticEmbedL
RAG 25: Test queries 0.0450