moebiusT7/mmv-wiki-index-ja

MOBIUS MMV — Wikipedia Index (Japanese, multilingual-e5-large)

Pre-computed FAISS vector index + cleaned text chunks over Japanese Wikipedia, embedded with intfloat/multilingual-e5-large (1024-dim). A sibling of moebiusT7/mmv-wiki-index (English).

Built for cross-lingual retrieval / RAG. No translation step needed — the ME5 embedding space is shared across languages.

Contents

File	Description
wiki_index_ivfpq_me5.faiss	FAISS IndexIVFPQ (d=1024, nlist=4096, m=64, nbits=8, INNER_PRODUCT), 1,550,503 vectors
wiki_chunks.jsonl.gz	Cleaned chunks (JSONL, gzip). Line i ↔ FAISS vector i. Fields: title, url, text, chunk_index, license, chunk_id
line_offsets.npy	int64 uncompressed byte-offset per line → O(1) chunk lookup by id
wiki_manifest.json	Build manifest (model, params, source, counts)

Embedding convention (E5)

• Passages (indexed): embedded with the "passage: " prefix
• Queries (search): must use the "query: " prefix
• Vectors are L2-normalized; the index metric is inner product (= cosine)

Usage

import faiss, gzip, json, numpy as np
from sentence_transformers import SentenceTransformer
from huggingface_hub import snapshot_download

d = snapshot_download("moebiusT7/mmv-wiki-index-ja", repo_type="dataset")
index = faiss.read_index(f"{d}/wiki_index_ivfpq_me5.faiss"); index.nprobe = 32
rows = [json.loads(l) for l in gzip.open(f"{d}/wiki_chunks.jsonl.gz", "rt", encoding="utf-8")]

m = SentenceTransformer("intfloat/multilingual-e5-large")
q = m.encode(["query: " + "..."], normalize_embeddings=True).astype("float32")
D, I = index.search(q, 5)
for s, i in zip(D[0], I[0]):
    print(s, rows[i]["title"], rows[i]["url"])

line_offsets.npy lets you fetch a single chunk without loading the whole file: seek to offsets[i] in the decompressed stream and read one line.

Provenance

• Source: wikipedia_ja_all_mini_2026-02.zim (Kiwix Wikipedia ZIM)
• Chunks: 1,550,503 (chars per chunk ≤ 1536, overlap 256; <script>/<style> blocks and CSS stripped)
• Embedding model: intfloat/multilingual-e5-large
• Index: IndexIVFPQ d=1024 nlist=4096 m=64 nbits=8 (INNER_PRODUCT)
• Built: 2026-06-19T05:04:19.342807+00:00

Notes

• Scores are modest in absolute value (~0.2–0.35) because IVFPQ product-quantization compresses 1024 floats → 64 bytes/vector; ranking is preserved. Use a IndexFlatIP rebuild from the chunks if you need exact scores.
• Minor cosmetic artifacts remain in some chunk text (repeated leading titles, undecoded HTML entities like  ).

License & attribution

Text is derived from Japanese Wikipedia and is licensed CC BY-SA 4.0. Attribution: Wikipedia contributors. Index/tooling © MOBIUS LLC (Taiko Toeda).