SAGE-Index-ReasonIR

Dense indices for SAGE encoded with ReasonIR, 4096-d.

raw/<domain>/chunk256_embeddings.npy                    float32, (n_chunks, 4096)
augmented/<domain>/chunk256_embeddings.npy              float32, (n_chunks, 4096)

These back the reasonir_* rows of the paper's main tables.

Loading

import numpy as np, json
from huggingface_hub import hf_hub_download

emb = np.load(hf_hub_download(
    "HughieHu/SAGE-Index-ReasonIR", "raw/computer_science/chunk256_embeddings.npy",
    repo_type="dataset"), mmap_mode="r")          # (1,414,444, 4096)

meta = json.load(open(hf_hub_download(
    "HughieHu/SAGE-Index-BM25",
    "chunk_meta/computer_science/chunk256_merged_index.json", repo_type="dataset")))

q = encode(query)                                  # L2-normalised, 4096-d
top = np.argsort(-(emb @ q))[:10]
for i in top:
    print(meta[i]["paper_title"])

Vectors are float32 and L2-normalised, so inner product is cosine similarity. Casting to float16 halves the footprint and leaves the ranking unchanged if you are short on disk. Retrieval in the paper is paper-level: chunk hits are deduplicated by doc_id, keeping each paper's best-scoring chunk.

Chunking

Every index here is built over the same chunking of the SAGE corpus: 256-word chunks, 4,937,036 chunks over 187,712 papers.

Domain	Papers	Chunks
computer_science	50,048	1,414,444
healthcare	52,100	1,307,287
humanities	34,996	892,910
natural_science	50,568	1,322,395

Row i of an embedding matrix is entry i of chunk256_merged_index.json, which gives {doc_id, paper_title, chunk_idx, total_chunks, row_idx}; the chunk text is at the same position in chunk256_merged_texts.json. Both files live in SAGE-Index-BM25 under chunk_meta/ and are identical for every retriever and every corpus variant, so you only ever need one copy. The documents themselves are in SAGE-Corpus.

Corpus variants

• raw/ — chunks encoded as parsed from the PDF.
• augmented/ — chunks encoded with the paper's corpus augmentation: domain terms extracted per chunk by Qwen3-Next-80B-A3B-Instruct, appended twice, together with the paper's title and venue. The chunk text is unchanged — augmentation is applied at encoding time — which is why chunk_meta/ is shared across variants. This is the configuration the code calls term_keywords_meta_2x.

Licence

CC-BY-NC-4.0 for the indices. The underlying papers remain the copyright of their publishers; see SAGE-Corpus.