Skip to content

Implement hybrid retrieval pipeline with Snowflake Arctic embeddings + BM25 + reranker #13

Open
chaman56 wants to merge 1 commit intodevrev:mainfrom
chaman56:main
Open

Implement hybrid retrieval pipeline with Snowflake Arctic embeddings + BM25 + reranker #13
chaman56 wants to merge 1 commit intodevrev:mainfrom
chaman56:main

Conversation

@chaman56
Copy link
Copy Markdown

@chaman56 chaman56 commented Mar 31, 2026

Overview

Optimized hybrid retrieval pipeline for DevRev Search challange.

Pipeline: Dense (Snowflake Arctic Embed L v2.0) + BM25 → Weighted RRF (BM25 2x weight) → Article-sibling expansion →
BGE-reranker-v2-m3 (top-150 candidates, 768-char context)

Results (291 annotated queries):

R@10 P@10 R@50 P@50 MRR@10
22.27% 15.70% 32.99% 6.05% 0.4378

Submission name: snowflake-arctic-embed-l-v2.0-bm25-bge-reranker-v2-m3

Ablation Results (291 queries)

Method R@10 P@10 R@50 P@50 MRR@10
Dense Only 4.85% 3.92% 9.16% 2.48% 0.096
BM25 Only 14.63% 10.17% 28.38% 4.65% 0.276
Dense + BM25 (RRF) 12.56% 9.45% 27.67% 4.77% 0.233
Dense + BM25 + Reranker 20.66% 14.81% 31.23% 5.88% 0.427

Systems Used

Component Type Open Source
Snowflake Arctic Embed L v2.0 Dense embedding (1024d) Yes, Apache 2.0
BM25 Okapi Sparse retrieval Yes, Apache 2.0
FAISS Vector index Yes, MIT
BGE-reranker-v2-m3 Cross-encoder reranker Yes, Apache 2.0

work-item: ISS-1

- snowflake_embedding.py: Embedding pipeline using snowflake-arctic-embed-l-v2.0
- hybrid_retrieval.ipynb: Dense + BM25 + RRF + reranker with ablation study
- optimized_retrieval.ipynb: Weighted RRF, sibling expansion, tuned params
- ablation_metrics.json: Evaluation results on 291 annotated queries
- Updated .gitignore for large binary files
@chaman56 chaman56 self-assigned this Mar 31, 2026
@chaman56 chaman56 requested a review from nimit2801 March 31, 2026 18:34
@chaman56
Copy link
Copy Markdown
Author

@nimit2801 @prakhar7651 , Made the submission, please help with evaluations, file to be evaluated: test_queries_results.json

@prakhar7651
Copy link
Copy Markdown
Contributor

Hey!
There were some errors in the evaluation pipeline. These are your scores.
Recall: 0.2513
Precision: 0.2000

@prakhar7651
Copy link
Copy Markdown
Contributor

prakhar7651 commented Apr 1, 2026

Looking at the quality of submissions and eagerness for folks to contribute, we're extending the deadline to April 7th. Evaluations would be still going on. Please keep contributing.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants