1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
impl SentencePieceTokenizer {
/// Create a new `SentencePiece` tokenizer
///
/// # Arguments
///
/// * `vocab` - List of (token, score) pairs where score is log probability
/// * `unk_token` - Unknown token string
///
/// # Errors
///
/// Returns error if vocabulary is empty or unknown token not found
pub fn new(vocab: Vec<(String, f32)>, unk_token: &str) -> Result<Self> {
if vocab.is_empty() {
return Err(RealizarError::UnsupportedOperation {
operation: "create_sentencepiece_tokenizer".to_string(),
reason: "Vocabulary cannot be empty".to_string(),
});
}
let mut token_to_id = HashMap::new();
let mut id_to_token = HashMap::new();
let mut scores = HashMap::new();
for (id, (token, score)) in vocab.into_iter().enumerate() {
let id = u32::try_from(id).map_err(|_| RealizarError::UnsupportedOperation {
operation: "convert_token_id".to_string(),
reason: format!("Token ID {id} exceeds u32 limit"),
})?;
token_to_id.insert(token.clone(), id);
id_to_token.insert(id, token.clone());
scores.insert(token, score);
}
let unk_token_id =
*token_to_id
.get(unk_token)
.ok_or_else(|| RealizarError::UnsupportedOperation {
operation: "create_sentencepiece_tokenizer".to_string(),
reason: format!("Unknown token '{unk_token}' not in vocabulary"),
})?;
Ok(Self {
token_to_id,
id_to_token,
scores,
unk_token_id,
})
}
/// Encode text to token IDs using Viterbi algorithm
///
/// Finds the most likely segmentation based on token scores.
///
/// # Arguments
///
/// * `text` - Input text to encode
///
/// # Returns
///
/// Vector of token IDs
#[must_use]
pub fn encode(&self, text: &str) -> Vec<u32> {
if text.is_empty() {
return Vec::new();
}
let chars: Vec<char> = text.chars().collect();
let (_best_score, best_token) = self.viterbi_forward(&chars);
let tokens = Self::viterbi_backtrack(&chars, &best_token);
// Convert tokens to IDs
tokens
.into_iter()
.map(|t| {
self.token_to_id
.get(&t)
.copied()
.unwrap_or(self.unk_token_id)
})
.collect()
}
/// Viterbi forward pass: find best score and token at each position
///
/// # Arguments
///
/// * `chars` - Character array of input text
///
/// # Returns
///
/// Tuple of `(best_score, best_token)` vectors
fn viterbi_forward(&self, chars: &[char]) -> ViterbiState {
let n = chars.len();
let mut best_score = vec![f32::NEG_INFINITY; n + 1];
let mut best_token: Vec<Option<String>> = vec![None; n + 1];
best_score[0] = 0.0;
for end in 1..=n {
// Try all possible start positions for token ending at `end`
for start in 0..end {
let substr: String = chars[start..end].iter().collect();
if let Some(&score) = self.scores.get(&substr) {
let new_score = best_score[start] + score;
if new_score > best_score[end] {
best_score[end] = new_score;
best_token[end] = Some(substr);
}
}
}
// If no token found ending at this position, use single character as unknown
if best_token[end].is_none() && best_score[end - 1] > f32::NEG_INFINITY {
let char_str: String = chars[end - 1..end].iter().collect();
best_score[end] = best_score[end - 1] - 100.0; // Penalty for unknown
best_token[end] = Some(char_str);
}
}
(best_score, best_token)
}
/// Viterbi backtracking: reconstruct token sequence from `best_token`
///
/// # Arguments
///
/// * `chars` - Character array of input text
/// * `best_token` - Best token at each position (from forward pass)
///
/// # Returns
///
/// Vector of tokens in forward order
fn viterbi_backtrack(chars: &[char], best_token: &[Option<String>]) -> Vec<String> {
let n = chars.len();
let mut tokens = Vec::new();
let mut pos = n;
while pos > 0 {
if let Some(token) = &best_token[pos] {
tokens.push(token.clone());
pos -= token.chars().count();
} else {
// Fallback: single character
let char_str: String = chars[pos - 1..pos].iter().collect();
tokens.push(char_str);
pos -= 1;
}
}
tokens.reverse();
tokens
}
/// Decode token IDs to text
///
/// # Arguments
///
/// * `token_ids` - Token IDs to decode
///
/// # Errors
///
/// Returns error if any token ID is invalid
pub fn decode(&self, token_ids: &[u32]) -> Result<String> {
let mut result = String::new();
for &id in token_ids {
let token =
self.id_to_token
.get(&id)
.ok_or_else(|| RealizarError::UnsupportedOperation {
operation: "decode_sentencepiece_token".to_string(),
reason: format!("Invalid token ID: {id}"),
})?;
result.push_str(token);
}
Ok(result)
}
/// Get vocabulary size
#[must_use]
pub fn vocab_size(&self) -> usize {
self.token_to_id.len()
}
/// Get token ID for a token
#[must_use]
pub fn get_token_id(&self, token: &str) -> Option<u32> {
self.token_to_id.get(token).copied()
}
/// Get token for a token ID
#[must_use]
pub fn get_token(&self, id: u32) -> Option<&str> {
self.id_to_token.get(&id).map(String::as_str)
}
/// Get score for a token
#[must_use]
pub fn get_score(&self, token: &str) -> Option<f32> {
self.scores.get(token).copied()
}
}
impl Tokenizer {
/// Create a new tokenizer from vocabulary
///
/// # Arguments
///
/// * `vocab` - Vocabulary mapping
/// * `unk_token` - Unknown token (default: `"<unk>"`)
///
/// # Errors
///
/// Returns error if unknown token not in vocabulary
///
/// # Examples
///
/// ```rust,ignore
/// let vocab = Vocabulary::from_tokens(vec!["<unk>".to_string(), "hello".to_string()])?;
/// let tokenizer = Tokenizer::new(vocab, "<unk>")?;
/// ```
pub fn new(vocab: Vocabulary, unk_token: &str) -> Result<Self> {
let unk_token_id =
vocab
.get_id(unk_token)
.ok_or_else(|| RealizarError::UnsupportedOperation {
operation: "create_tokenizer".to_string(),
reason: format!("Unknown token '{unk_token}' not in vocabulary"),
})?;
Ok(Self {
vocab,
unk_token_id,
})
}
/// Encode text to token IDs (simple word-level tokenization)
///
/// # Arguments
///
/// * `text` - Input text to encode
///
/// # Returns
///
/// Vector of token IDs
///
/// # Examples
///
/// ```rust,ignore
/// let token_ids = tokenizer.encode("hello world")?;
/// assert_eq!(token_ids, vec![1, 2]);
/// ```
#[must_use]
pub fn encode(&self, text: &str) -> Vec<u32> {
text.split_whitespace()
.map(|word| self.vocab.get_id(word).unwrap_or(self.unk_token_id))
.collect()
}
/// Decode token IDs to text
///
/// # Arguments
///
/// * `token_ids` - Token IDs to decode
///
/// # Returns
///
/// Decoded text string
///
/// # Errors
///
/// Returns error if any token ID is invalid
///
/// # Examples
///
/// ```rust,ignore
/// let text = tokenizer.decode(&[1, 2])?;
/// assert_eq!(text, "hello world");
/// ```
pub fn decode(&self, token_ids: &[u32]) -> Result<String> {
let tokens: Result<Vec<&str>> = token_ids
.iter()
.map(|&id| {
self.vocab
.get_token(id)
.ok_or_else(|| RealizarError::UnsupportedOperation {
operation: "decode_token".to_string(),
reason: format!("Invalid token ID: {id}"),
})
})
.collect();
Ok(tokens?.join(" "))
}
/// Get vocabulary size
#[must_use]
pub fn vocab_size(&self) -> usize {
self.vocab.size()
}
}