Word Encoding API

Encode words into dense embeddings using Sawtone v4, a remote-code multilingual word encoder that exposes phonetic, orthographic, and semantic word channels through a stable API.

Endpoint

text

POST https://api.sawalni.com/v1/language/encode

Request Body

Parameter	Type	Required	Description
`model`	string	Yes	`sawtone-v4`
`input`	string or array	Yes	Word(s) or phrase(s) to encode
`channel`	string	No	`composite` (544d, default), `phonetic` (192d), `ortho` (192d), or `semantic` (128d)

Channels

Channel	Dimensions	Description
`composite`	544	Sawtone v4 composite word embedding
`phonetic`	192	CNN-based sound-pattern channel
`ortho`	192	Orthographic byte-structure channel
`semantic`	128	Base semantic channel inferred from form

Multi-word Input

When input contains multiple words (separated by whitespace), each word is encoded individually and the results are max-pooled into a single vector.

Example — curl

bash

curl https://api.sawalni.com/v1/language/encode \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sawtone-v4",
    "input": ["hello", "café", "München"],
    "channel": "composite"
  }'

Example — Python

python

import requests

response = requests.post(
    "https://api.sawalni.com/v1/language/encode",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model": "sawtone-v4",
        "input": ["hello", "café", "مرحبا"],
        "channel": "phonetic"
    }
)

result = response.json()
for item in result["data"]:
    print(f"{item['input']} → {len(item['embedding'])}d embedding")

Example — JavaScript

javascript

const response = await fetch('https://api.sawalni.com/v1/language/encode', {
  method: 'POST',
  headers: {
    'Authorization': 'Bearer YOUR_API_KEY',
    'Content-Type': 'application/json'
  },
  body: JSON.stringify({
    model: 'sawtone-v4',
    input: ['hello', 'café', 'مرحبا'],
    channel: 'composite'
  })
});

const result = await response.json();
result.data.forEach(item => {
  console.log(item.input, '→', item.embedding.length + 'd');
});

Response

json

{
  "data": [
    {
      "input": "hello",
      "embedding": [0.023, -0.009, 0.015, ...]
    }
  ],
  "model": "sawtone-v4",
  "channel": "composite",
  "dimensions": 544,
  "usage": {
    "input_bytes": 5,
    "input_count": 1,
    "total_words": 1,
    "processing_ms": 3.2
  }
}

Batch Encoding

Pass an array of strings to encode multiple texts in one request (max 128):

python

response = requests.post(
    "https://api.sawalni.com/v1/language/encode",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model": "sawtone-v4",
        "input": ["word1", "word2", "word3"],
        "channel": "composite"
    }
)

Model Details

Property	Value
Model ID	`sawtone-v4`
Composite Dimension	544
Sentence Embeddings	768
Max Batch Size	128

Pricing

Word encoding is metered by input bytes (UTF-8 encoded size).

Metric	Rate
Input bytes	0.10 credits per 1M bytes

Use Cases

Cross-lingual word similarity — Compare words across languages via cosine similarity
Phonetic search — Find phonetically similar words using the phonetic channel
Script-agnostic clustering — Group words by sound/meaning regardless of script
Lexicon building — Build multilingual lexicons using semantic embeddings
Transliteration features — Use orthographic embeddings as features for transliteration models

Parameter

Type

Required

Description

model

string

Yes

sawtone-v4

input

string or array

Yes

Word(s) or phrase(s) to encode

channel

string

composite (544d, default), phonetic (192d), ortho (192d), or semantic (128d)

Channel

Dimensions

Description

composite

544

Sawtone v4 composite word embedding

phonetic

192

CNN-based sound-pattern channel

ortho

192

Orthographic byte-structure channel

semantic

128

Base semantic channel inferred from form

curl https://api.sawalni.com/v1/language/encode \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "sawtone-v4", "input": ["hello", "café", "München"], "channel": "composite" }'

Example — Python

python

import requests

response = requests.post(
    "https://api.sawalni.com/v1/language/encode",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model": "sawtone-v4",
        "input": ["hello", "café", "مرحبا"],
        "channel": "phonetic"
    }
)

result = response.json()
for item in result["data"]:
    print(f"{item['input']} → {len(item['embedding'])}d embedding")

Example — JavaScript

javascript

const response = await fetch('https://api.sawalni.com/v1/language/encode', {
  method: 'POST',
  headers: {
    'Authorization': 'Bearer YOUR_API_KEY',
    'Content-Type': 'application/json'
  },
  body: JSON.stringify({
    model: 'sawtone-v4',
    input: ['hello', 'café', 'مرحبا'],
    channel: 'composite'
  })
});

const result = await response.json();
result.data.forEach(item => {
  console.log(item.input, '→', item.embedding.length + 'd');
});

{ "data": [ { "input": "hello", "embedding": [0.023, -0.009, 0.015, ...] } ], "model": "sawtone-v4", "channel": "composite", "dimensions": 544, "usage": { "input_bytes": 5, "input_count": 1, "total_words": 1, "processing_ms": 3.2 } }

Batch Encoding

Pass an array of strings to encode multiple texts in one request (max 128):

python

response = requests.post(
    "https://api.sawalni.com/v1/language/encode",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model": "sawtone-v4",
        "input": ["word1", "word2", "word3"],
        "channel": "composite"
    }
)

Property

Value

Model ID

sawtone-v4

Composite Dimension

544

Sentence Embeddings

768

Max Batch Size

128

Metric

Rate

Input bytes

0.10 credits per 1M bytes

Use Cases

Cross-lingual word similarity — Compare words across languages via cosine similarity

Phonetic search — Find phonetically similar words using the phonetic channel

Script-agnostic clustering — Group words by sound/meaning regardless of script

Lexicon building — Build multilingual lexicons using semantic embeddings

Transliteration features — Use orthographic embeddings as features for transliteration models