CohereLabs/wikipedia-2023-11-embed-multilingual-v3-int8-binary

Multilingual Embeddings for Wikipedia in 300+ Languages (int8 & binary embeddings)

This dataset contains the wikimedia/wikipedia dataset dump from 2023-11-01 from Wikipedia in all 300+ languages. The embeddings are provided as int8 and ubinary that allow quick search and reduction of your vector index size up to 32. For more details, see Cohere int8 & binary Embeddings

The individual articles have been chunked and embedded with the state-of-the-art multilingual Cohere Embed V3 embedding model. This enables an easy way to semantically search across all of Wikipedia or to use it as a knowledge source for your RAG application. In total is it close to 250M paragraphs / embeddings.

You can also use the model to perform cross-lingual search: Enter your search query in any language and get the most relevant results back.

Loading the dataset

Loading the document embeddings

You can either load the dataset like this:

from datasets import load_dataset
lang = "simple" #Use the Simple English Wikipedia subset
docs = load_dataset("CohereLabs/wikipedia-2023-11-embed-multilingual-v3-int8-binary", lang, split="train")

Or you can also stream it without downloading it before:

from datasets import load_dataset
lang = "simple" #Use the Simple English Wikipedia subset
docs = load_dataset("CohereLabs/wikipedia-2023-11-embed-multilingual-v3-int8-binary", lang, split="train", streaming=True)
for doc in docs:
    doc_id = doc['_id']
    title = doc['title']
    text = doc['text']
    emb_int8 = doc['emb_int8']    #Int8 embeddings
    emb_ubinary = doc['emb_ubinary']  #(unsigned) binary embeddings

Note, depending on the language, the download can be quite large.

Search - int8 embeddings

int8 embeddings keep 99.7-100% of the search quality, while reducing your vector index 4x and speeding-up vector search up to 30%.

Below example using usearch to search on int8 embeddings.

from datasets import load_dataset
import numpy as np
import cohere
from usearch.index import Index

co = cohere.Client("<<YOUR_API_KEY>>")  # Add your cohere API key from www.cohere.com
lang = "simple"

#Load at max 1000 chunks + embeddings
max_docs = 1000
docs_stream = load_dataset(f"CohereLabs/wikipedia-2023-11-embed-multilingual-v3-int8-binary", lang, split="train", streaming=True)

docs = []
doc_embeddings = []
for doc in docs_stream:
    docs.append(doc)
    doc_embeddings.append(doc['emb_int8'])
    if len(docs) >= max_docs:
        break

doc_embeddings = np.asarray(doc_embeddings, dtype='int8')

print(doc_embeddings.shape, doc_embeddings.dtype)

#Create the usearch HNSW index on the int8 embeddings
num_dim = 1024 
index = Index(ndim=1024, metric='ip', dtype='i8')
index.add(np.arange(len(doc_embeddings)), doc_embeddings)

#Searching
query = 'Who was Alan Turing'
query_emb = co.embed(texts=[query], model="embed-multilingual-v3.0", input_type="search_query", embedding_types=["int8"]).embeddings
query_emb = np.asarray(query_emb.int8, dtype='int8')

#Search on the index and get top-10 results
matches = index.search(query_emb, 10)

# Print results
for match in matches:
    doc_id = match.key
    print(docs[doc_id]['title'])
    print(docs[doc_id]['text'])
    print("----")

Search - Binary embeddings

Binary embeddings keep 95% of the search quality, while reducing your vector index 32x and speeding-up vector search up to 40x.

Below example using faiss to search on binary embeddings.

#Run: pip install cohere datasets numpy faiss-cpu
from datasets import load_dataset
import numpy as np
import cohere
import faiss

co = cohere.Client("<<YOUR_API_KEY>>")  # Add your cohere API key from www.cohere.com
lang = "simple"

#Load at max 1000 chunks + embeddings
max_docs = 1000
docs_stream = load_dataset(f"CohereLabs/wikipedia-2023-11-embed-multilingual-v3-int8-binary", lang, split="train", streaming=True)

docs = []
doc_embeddings = []
for doc in docs_stream:
    docs.append(doc)
    doc_embeddings.append(doc['emb_ubinary'])
    if len(docs) >= max_docs:
        break

doc_embeddings = np.asarray(doc_embeddings, dtype='uint8')

#Create the faiss IndexBinaryFlat index
num_dim = 1024 
index = faiss.IndexBinaryFlat(num_dim)
index.add(doc_embeddings)

#Search
def search(index, query, top_k=10):
    # Make sure to set input_type="search_query"
    query_emb = co.embed(texts=[query], model="embed-multilingual-v3.0", input_type="search_query", embedding_types=["ubinary", "float"]).embeddings
    query_emb_bin = np.asarray(query_emb.ubinary, dtype='uint8')
    query_emb_float = np.asarray(query_emb.float, dtype="float32")

    # Phase I: Search on the index with a binary
    hits_scores, hits_doc_ids = index.search(query_emb_bin, k=min(10*top_k, index.ntotal))

    #Get the results in a list of hits
    hits = [{'doc_id': doc_id.item(), 'score_bin': score_bin} for doc_id, score_bin in zip(hits_doc_ids[0], hits_scores[0])]

    # Phase II: Do a re-scoring with the float query embedding
    binary_doc_emb = np.asarray([index.reconstruct(hit['doc_id']) for hit in hits])
    binary_doc_emb_unpacked = np.unpackbits(binary_doc_emb, axis=-1).astype("int")
    binary_doc_emb_unpacked = 2*binary_doc_emb_unpacked-1

    scores_cont = (query_emb_float[0] @ binary_doc_emb_unpacked.T)
    for idx in range(len(scores_cont)):
        hits[idx]['score_cont'] = scores_cont[idx]

    #Sort by largest score_cont
    hits.sort(key=lambda x: x['score_cont'], reverse=True)

    return hits[0:top_k]

query = 'Who was Alan Turing'
hits = search(index, query)

# Print results
print("Query:", query)
for hit in hits:
    doc_id = hit['doc_id']
    print(docs[doc_id]['title'])
    print(docs[doc_id]['text'])
    print(docs[doc_id]['url'], "\n")

Overview

The following table contains all language codes together with the total numbers of passages.

Language	#Docs
en	41,488,110
de	20,772,081
fr	17,813,768
ru	13,734,543
es	12,905,284
it	10,462,162
ceb	9,818,657
uk	6,901,192
ja	6,626,537
nl	6,101,353
pl	5,973,650
pt	5,637,930
sv	4,911,480
ca	4,156,889
ar	3,691,451
cs	3,118,524
he	2,948,882
hu	2,924,609
vi	2,835,049
zh	2,775,260
fi	2,427,097
id	2,358,335
no	2,211,270
sr	2,154,631
fa	2,073,154
tr	1,775,036
ro	1,770,527
el	1,599,770
ko	1,513,291
bg	1,455,765
hy	1,386,140
eu	1,327,579
da	1,224,982
eo	1,216,706
war	1,185,097
sh	1,139,309
tt	1,119,546
arz	1,089,164
gl	1,056,990
et	1,054,770
ce	1,013,217
ast	1,010,445
sl	984,855
hr	910,923
sk	874,014
ms	869,579
be	857,033
th	839,712
az	817,631
uz	811,028
mk	784,576
lt	770,069
bn	767,965
cy	762,338
ta	685,975
simple	646,424
te	634,778
kk	627,085
ka	595,401
hi	541,822
nn	530,590
lv	484,957
af	461,674
ba	434,939
ur	434,269
bs	396,692
sq	388,788
ml	384,795
min	373,156
la	340,521
pnb	335,958
be-x-old	314,600
kn	309,085
azb	294,248
oc	283,297
zh-min-nan	278,547
fy	248,075
my	241,025
lb	216,558
ky	216,344
als	206,387
mr	203,479
br	200,486
pa	188,091
is	177,272
mg	171,947
sw	171,650
ha	167,807
tl	166,907
nds	166,019
an	143,163
jv	142,104
ps	138,240
ig	132,250
new	128,696
tg	128,237
ga	125,456
lld	125,094
su	124,390
cv	122,671
ckb	120,886
si	119,223
mn	114,878
lmo	103,836
io	101,713
gu	99,450
vec	95,072
zh-yue	89,145
bar	88,238
sco	83,906
ne	83,598
ku	82,935
hyw	82,343
pms	77,834
as	76,093
km	74,177
sah	71,599
li	69,267
or	65,510
mt	64,038
szl	56,836
yi	55,375
ht	55,079
dag	53,343
sa	51,735
nv	49,355
bpy	47,757
vo	47,375
ug	44,764
sat	43,500
ia	42,012
bo	41,438
mwl	41,273
sd	40,395
bcl	39,967
mnw	39,578
hsb	39,560
avk	39,001
scn	38,359
rm	37,436
diq	34,743
vep	33,654
xmf	33,238
ban	32,320
wa	32,132
ilo	31,046
nds-nl	30,918
qu	30,529
so	29,936
mhr	29,619
vls	29,227
sc	28,977
fo	28,809
gd	28,149
rw	28,037
gom	27,792
yo	27,789
tum	26,743
wuu	26,532
frr	26,010
sn	25,941
tk	24,269
blk	24,194
mzn	23,837
co	23,065
szy	22,854
am	22,467
shn	22,432
skr	21,081
lfn	20,781
tyv	20,762
lij	20,553
ie	19,994
rue	19,916
crh	19,016
gor	18,146
ary	17,463
dv	16,941
lg	16,751
roa-tara	16,572
bjn	16,429
tw	16,304
bh	15,938
pam	15,134
os	15,096
myv	15,062
gn	14,983
lez	14,152
mai	13,806
kv	13,534
pcd	13,057
zh-classical	12,791
zea	12,528
lo	12,525
gv	12,074
stq	11,890
zu	11,680
smn	11,672
kw	11,539
bat-smg	11,240
hif	11,215
ext	10,967
ace	10,821
trv	10,546
ami	10,538
tcy	10,531
lad	10,386
alt	10,256
pap	10,187
kab	10,179
fur	10,148
nap	10,079
mrj	9,771
kaa	9,548
nqo	9,153
glk	9,120
pfl	8,790
fiu-vro	8,757
nso	8,635
jbo	8,577
bxr	8,549
wo	8,549
olo	8,530
map-bms	8,393
ksh	8,226
csb	8,085
av	7,873
mni	7,740
udm	7,730
mi	7,643
kbp	7,616
dsb	7,536
frp	7,294
om	7,045
ang	7,023
hak	6,866
gur	6,761
se	6,733
anp	6,704
tay	6,434
mdf	6,351
gcr	6,347
koi	6,300
krc	6,293
ay	5,985
cdo	5,917
nrm	5,786
xh	5,756
tn	5,712
tly	5,598
shi	5,179
pcm	5,076
fat	4,968
nia	4,795
dty	4,728
kbd	4,667
gpe	4,289
cbk-zam	4,224
ff	4,166
dz	4,117
guw	3,982
eml	3,979
ln	3,774
inh	3,768
nah	3,720
ab	3,465
ks	3,255
mad	3,236
haw	3,227
gag	3,076
tet	3,030
ny	2,933
pag	2,727
guc	2,454
roa-rup	2,409
jam	2,387
awa	2,242
pdc	2,239
to	2,165
za	2,132
st	2,051
ltg	2,005
atj	1,967
nov	1,916
ss	1,904
pwn	1,881
ee	1,819
sm	1,659
ts	1,645
gan	1,626
xal	1,619
kcg	1,555
cu	1,477
srn	1,395
got	1,280
fon	1,247
din	1,214
arc	1,167
fj	1,164
rmy	1,113
ady	1,040
rn	1,033
bm	1,017
tpi	957
ve	919
ki	798
pnt	796
chr	788
kl	770
lbe	766
bi	718
ti	706
kg	609
pih	606
ch	513
bug	429
ty	297
ik	275
iu	263
pi	260
sg	204
chy	57
cr	41
Total	247,154,006