Host sentence-transformers/LaBSE model in LiftWing
Open, Needs TriagePublicFeature
Actions

Assigned To

None

Authored By

	santhosh
	Feb 6 2026, 7:00 AM

Description

As part of https://phabricator.wikimedia.org/T404183 task that LPL team is working in Q3 and Q4 of 2025-26FY, we want an embedding model available in LiftWing.

Details as follows:

Model: https://huggingface.co/sentence-transformers/LaBSE
Architecture: Sentence Transformer
capability: It can be used to map 109 languages to a shared vector space.
License: Apache 2.0
https://embed.toolforge.org/ hosts the LaBSE model with an OpenVINO backend. But it is quantized.
Internal KServer API is enough as we can connect it from cxserver production instance. For development purpose we can continue to use embed.toolforge.org or the model can be operated in local dev environment
Expected API endpoints: embeddings from :predict method
Expecting about 5 Requests/Second , latency: <300 ms.
Request will be a list of template parameter names - small strings, with list size under 50

Details

Related Changes in Gerrit:

	Subject	Repo	Branch	Lines +/-
	Add sentence-transformers/LaBSE model	machinelearning/liftwing/inference-services	main	+17 -0

Customize query in gerrit

Related Objects

Mentioned In: rMLISde50b794c450: Add sentence-transformers/LaBSE model
T403146: Explore and evaluate Section suggestions based on language agnostic embedding similarity
Mentioned Here: P88721 nodejs client for the liftwing LaBSE embeddings
T404183: Explore and evaluate template parameter alignment based on language agnostic embedding similarity

Event Timeline

santhosh created this task.Feb 6 2026, 7:00 AM

Restricted Application added a subscriber: Aklapper. · View Herald TranscriptFeb 6 2026, 7:00 AM

Nikerabbit moved this task from To triage to Blocked on others on the Language and Product Localization board.Feb 6 2026, 8:41 AM

Change #1237731 had a related patch set uploaded (by Santhosh; author: Santhosh):

[machinelearning/liftwing/inference-services@main] Add sentence-transformers/LaBSE model

https://gerrit.wikimedia.org/r/1237731

gerritbot added a project: Patch-For-Review.Feb 9 2026, 5:27 AM

Here is nodejs client code that can give cosine similarity based on the above service.
Pull the above patch, then

docker compose build labse-embeddings
docker compose up labse-embeddings

Get this code,

P88721 nodejs client for the liftwing LaBSE embeddings

1	#!/usr/bin/env node
2
3	const http = require("http");
4
5	/**
6	* Get embedding vector from the API
7	* @param {string[]} text - Text to get embedding for
8	* @returns {Promise<number[]>} - Embedding vector
9	*/
10	async function getEmbedding(texts) {
11	const postData = JSON.stringify({
12	input: texts,
13	});
14
15	const options = {
16	hostname: "localhost",
17	port: 8080,
18	path: "/v1/models/labse-embedding:predict",
19	method: "POST",
20	headers: {
21	"Content-Type": "application/json",
22	"Content-Length": Buffer.byteLength(postData),
23	},
24	};
25
26	return new Promise((resolve, reject) => {
27	const req = http.request(options, (res) => {
28	let data = "";
29
30	res.on("data", (chunk) => {
31	data += chunk;
32	});
33
34	res.on("end", () => {
35	try {
36	const response = JSON.parse(data);
37	// Adjust this based on the actual API response structure
38	const embeddings = response.data;
39	resolve(embeddings);
40	} catch (error) {
41	reject(new Error(`Failed to parse response: ${error.message}`));
42	}
43	});
44	});
45
46	req.on("error", (error) => {
47	reject(new Error(`API request failed: ${error.message}`));
48	});
49
50	req.write(postData);
51	req.end();
52	});
53	}
54	/**
55	* Calculate cosine similarity between two normalized vectors
56	* @param {number[]} vec1 - First normalized vector
57	* @param {number[]} vec2 - Second normalized vector
58	* @returns {number} - Similarity score between -1 and 1
59	*/
60	function cosineSimilarity(vec1, vec2) {
61	if (vec1.length !== vec2.length) {
62	throw new Error("Vectors must have the same length");
63	}
64
65	let dotProduct = 0;
66	for (let i = 0; i < vec1.length; i++) {
67	dotProduct += vec1[i] * vec2[i];
68	}
69
70	return dotProduct;
71	}
72
73	/**
74	* Compare two strings for similarity using embeddings
75	* @param {string} str1 - First string to compare
76	* @param {string} str2 - Second string to compare
77	* @returns {Promise<number>} - Similarity score between 0 and 1
78	*/
79	async function compareStrings(str1, str2) {
80	console.log("Getting embeddings ...");
81	const embeddings = await getEmbedding([str1, str2]);
82	console.log("Calculating similarity...");
83	const similarity = cosineSimilarity(
84	embeddings[0].embedding,
85	embeddings[1].embedding,
86	);
87
88	return similarity;
89	}
90
91	// Main execution
92	async function main() {
93	const args = process.argv.slice(2);
94
95	if (args.length < 2) {
96	console.error("Usage: node compare-strings.js <string1> <string2>");
97	console.error('Example: node compare-strings.js "Hello world" "Hi there"');
98	process.exit(1);
99	}
100
101	const string1 = args[0];
102	const string2 = args[1];
103
104	console.log(`\nComparing strings:`);
105	console.log(`String 1: "${string1}"`);
106	console.log(`String 2: "${string2}"\n`);
107
108	try {
109	const similarity = await compareStrings(string1, string2);
110	console.log(
111	`\nSimilarity score: ${similarity.toFixed(4)} (${(similarity * 100).toFixed(2)}%)`,
112	);
113	} catch (error) {
114	console.error(`Error: ${error.message}`);
115	process.exit(1);
116	}
117	}
118
119	main();

Then

node compare-strings.js "cat" "gato"


Comparing strings:
String 1: "cat"
String 2: "gato"

Getting embeddings ...
Calculating similarity...

Similarity score: 0.9522 (95.22%)

Notice that the cosine similarity uses a simple dot product of the vectors since the embeddings are normalized in the liftwing modelserver. Euclidean distance over normalized vectors is a linear transform of the cosine distance.

santhosh mentioned this in T403146: Explore and evaluate Section suggestions based on language agnostic embedding similarity.Feb 9 2026, 5:40 AM

kevinbazira subscribed.Feb 9 2026, 6:27 AM

santhosh updated the task description. (Show Details)Feb 9 2026, 6:57 AM

Change #1237731 merged by jenkins-bot:

[machinelearning/liftwing/inference-services@main] Add sentence-transformers/LaBSE model

https://gerrit.wikimedia.org/r/1237731

santhosh mentioned this in rMLISde50b794c450: Add sentence-transformers/LaBSE model.Feb 9 2026, 7:45 AM

Maintenance_bot removed a project: Patch-For-Review.Feb 9 2026, 8:31 AM