ᯤpaper-sonar

README_ja

• A code that suggests papers similar to input keywords

Demo using Huggingface spaces

• You can try from here

Target

• Accepted papers from ICLR2025

Note

This can be repurposed for any conference where paper titles, abstracts, and URLs are available by simply replacing the target files

Process

• Concatenate title, first 600 characters of abstract, and URL in plain text
• Convert to vectorstore using embedding model
• Output best-n similar papers using approximate nearest neighbor search

Note

The embedding model was selected based on MTEB leaderboard

Implementation

1. Download accepted papers from this link

• Place in data directory

2. Convert to txt file

python src/build_db/json2txt.py

3. Convert to vectorstore

python src/build_db

HuggingFace demo

1. Create a new space on Huggingface account
2. Move the following files to the created space:

• app.py
• vectorstore files created in data/db folder
• requirements.txt

Example

query: β-calibration of Language Model Confidence Scores for Generative QA suggestion: