-
Notifications
You must be signed in to change notification settings - Fork 19
/
lib.rs
726 lines (649 loc) · 24.5 KB
/
lib.rs
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
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
#![doc = include_str!("../README.md")]
#![cfg_attr(docsrs, feature(doc_auto_cfg, doc_cfg))]
use std::{
cmp::Ordering,
fmt,
iter::once,
ops::{Range, RangeFrom, RangeFull, RangeInclusive, RangeTo, RangeToInclusive},
};
use either::Either;
use itertools::Itertools;
mod characters;
#[cfg(feature = "tokenizers")]
mod huggingface;
#[cfg(feature = "markdown")]
mod markdown;
mod text;
#[cfg(feature = "tiktoken-rs")]
mod tiktoken;
pub use characters::Characters;
#[cfg(feature = "markdown")]
pub use markdown::MarkdownSplitter;
pub use text::TextSplitter;
/// Result returned from a `ChunkSizer`. Includes the size of the chunk, in units
/// determined by the sizer, as well as the max byte offset of the text that
/// would fit within the given `ChunkCapacity`.
#[derive(Copy, Clone, Debug, PartialEq)]
pub struct ChunkSize {
/// Whether or not the entire chunk fits within the `ChunkCapacity`
fits: Ordering,
/// max byte offset of the text that fit within the given `ChunkCapacity`.
max_chunk_size_offset: Option<usize>,
/// Size of the chunk, in units used by the sizer.
size: usize,
}
impl ChunkSize {
/// Generate a chunk size from a given size. Will not be able to compute the
/// max byte offset that fits within the capacity.
pub fn from_size(size: usize, capacity: &impl ChunkCapacity) -> Self {
Self {
fits: capacity.fits(size),
max_chunk_size_offset: None,
size,
}
}
/// Generate a chunk size from an iterator of byte ranges for each encoded
/// element in the chunk.
pub fn from_offsets(
offsets: impl Iterator<Item = Range<usize>>,
capacity: &impl ChunkCapacity,
) -> Self {
let mut chunk_size = offsets.fold(
Self {
fits: Ordering::Less,
max_chunk_size_offset: None,
size: 0,
},
|mut acc, range| {
acc.size += 1;
if acc.size <= capacity.end() {
acc.max_chunk_size_offset = Some(range.end);
}
acc
},
);
chunk_size.fits = capacity.fits(chunk_size.size);
chunk_size
}
/// Determine whether the chunk size fits within the capacity or not
#[must_use]
pub fn fits(&self) -> Ordering {
self.fits
}
}
/// Determines the size of a given chunk.
pub trait ChunkSizer {
/// Determine the size of a given chunk to use for validation
fn chunk_size(&self, chunk: &str, capacity: &impl ChunkCapacity) -> ChunkSize;
}
/// Describes the largest valid chunk size(s) that can be generated.
///
/// An `end` size is required, which is the maximum possible chunk size that
/// can be generated.
///
/// A `start` size is optional. By specifying `start` and `end` it means a
/// range of sizes will be considered valid. Once a chunk has reached a length
/// that falls between `start` and `end` it will be returned.
///
/// It is always possible that a chunk may be returned that is less than the
/// `start` value, as adding the next piece of text may have made it larger
/// than the `end` capacity.
pub trait ChunkCapacity {
/// An optional `start` value. If both `start` and `end` are specified, a
/// valid chunk can fall anywhere between the two values (inclusive).
fn start(&self) -> Option<usize> {
None
}
/// The maximum size that a chunk can be.
#[must_use]
fn end(&self) -> usize;
/// Validate if a given chunk fits within the capacity
///
/// - `Ordering::Less` indicates more could be added
/// - `Ordering::Equal` indicates the chunk is within the capacity range
/// - `Ordering::Greater` indicates the chunk is larger than the capacity
fn fits(&self, chunk_size: usize) -> Ordering {
let end = self.end();
match self.start() {
Some(start) => {
if chunk_size < start {
Ordering::Less
} else if chunk_size > end {
Ordering::Greater
} else {
Ordering::Equal
}
}
None => chunk_size.cmp(&end),
}
}
}
impl ChunkCapacity for usize {
fn end(&self) -> usize {
*self
}
}
impl ChunkCapacity for Range<usize> {
fn start(&self) -> Option<usize> {
Some(self.start)
}
fn end(&self) -> usize {
self.end.saturating_sub(1).max(self.start)
}
}
impl ChunkCapacity for RangeFrom<usize> {
fn start(&self) -> Option<usize> {
Some(self.start)
}
fn end(&self) -> usize {
usize::MAX
}
}
impl ChunkCapacity for RangeFull {
fn start(&self) -> Option<usize> {
Some(usize::MIN)
}
fn end(&self) -> usize {
usize::MAX
}
}
impl ChunkCapacity for RangeInclusive<usize> {
fn start(&self) -> Option<usize> {
Some(*self.start())
}
fn end(&self) -> usize {
*self.end()
}
}
impl ChunkCapacity for RangeTo<usize> {
fn start(&self) -> Option<usize> {
Some(usize::MIN)
}
fn end(&self) -> usize {
self.end.saturating_sub(1)
}
}
impl ChunkCapacity for RangeToInclusive<usize> {
fn start(&self) -> Option<usize> {
Some(usize::MIN)
}
fn end(&self) -> usize {
self.end
}
}
/// How a particular semantic level relates to surrounding text elements.
#[allow(dead_code)]
#[derive(Copy, Clone, Debug, Eq, Ord, PartialEq, PartialOrd)]
enum SemanticSplitPosition {
/// The semantic level should be included in the previous chunk.
Prev,
/// The semantic level should be treated as its own chunk.
Own,
/// The semantic level should be included in the next chunk.
Next,
}
/// Information required by generic Semantic Levels
trait Level: fmt::Debug {
fn split_position(&self) -> SemanticSplitPosition;
/// Whether or not when splitting ranges, whitespace should be included as previous.
fn treat_whitespace_as_previous(&self) -> bool {
false
}
}
/// Implementation that dictates the semantic split points available.
/// For plain text, this goes from characters, to grapheme clusters, to words,
/// to sentences, to linebreaks.
/// For something like Markdown, this would also include things like headers,
/// lists, and code blocks.
trait SemanticSplit {
/// Internal type used to represent the level of semantic splitting.
type Level: Copy + Level + Ord + PartialOrd + 'static;
/// Levels that are always considered in splitting text, because they are always present.
const PERSISTENT_LEVELS: &'static [Self::Level];
/// Generate a new instance from a given text.
fn new(text: &str) -> Self;
/// Retrieve ranges for each semantic level in the entire text
fn ranges(&self) -> impl Iterator<Item = &(Self::Level, Range<usize>)> + '_;
/// Retrieve ranges for all sections of a given level after an offset
fn ranges_after_offset(
&self,
offset: usize,
level: Self::Level,
) -> impl Iterator<Item = &(Self::Level, Range<usize>)> + '_ {
let first_item = self.ranges().find(|(l, _)| l == &level);
self.ranges()
.filter(move |(l, sep)| l >= &level && sep.start >= offset)
.skip_while(move |(l, r)| {
first_item.is_some_and(|(_, fir)| l > &level && r.contains(&fir.start))
})
}
/// Return a unique, sorted list of all line break levels present before the next max level, added
/// to all of the base semantic levels, in order from smallest to largest
fn levels_in_remaining_text(&self, offset: usize) -> impl Iterator<Item = Self::Level> + '_ {
let existing_levels = self
.ranges()
// Only start taking them from the offset
.filter(|(_, sep)| sep.start >= offset)
.map(|(l, _)| l);
Self::PERSISTENT_LEVELS
.iter()
.chain(existing_levels)
.sorted()
.dedup()
.copied()
}
/// Split a given text into iterator over each semantic chunk
fn semantic_chunks<'splitter, 'text: 'splitter>(
&'splitter self,
offset: usize,
text: &'text str,
semantic_level: Self::Level,
) -> impl Iterator<Item = (usize, &'text str)> + 'splitter;
/// Trim the str and adjust the offset if necessary.
/// This is the default behavior, but custom semantic levels may need different behavior.
fn trim_chunk<'splitter, 'text: 'splitter>(
&'splitter self,
offset: usize,
chunk: &'text str,
) -> (usize, &'text str) {
// Figure out how many bytes we lose trimming the beginning
let diff = chunk.len() - chunk.trim_start().len();
(offset + diff, chunk.trim())
}
}
/// Returns chunks of text with their byte offsets as an iterator.
#[derive(Debug)]
struct TextChunks<'text, 'sizer, C, S, Sp>
where
C: ChunkCapacity,
S: ChunkSizer,
Sp: SemanticSplit,
{
/// Size of the chunks to generate
chunk_capacity: C,
/// How to validate chunk sizes
chunk_sizer: &'sizer S,
/// Current byte offset in the `text`
cursor: usize,
/// Splitter used for determining semantic levels.
semantic_split: Sp,
/// Original text to iterate over and generate chunks from
text: &'text str,
/// Whether or not chunks should be trimmed
trim_chunks: bool,
}
impl<'sizer, 'text: 'sizer, C, S, Sp> TextChunks<'text, 'sizer, C, S, Sp>
where
C: ChunkCapacity,
S: ChunkSizer,
Sp: SemanticSplit,
{
/// Generate new [`TextChunks`] iterator for a given text.
/// Starts with an offset of 0
fn new(chunk_capacity: C, chunk_sizer: &'sizer S, text: &'text str, trim_chunks: bool) -> Self {
Self {
cursor: 0,
chunk_capacity,
chunk_sizer,
semantic_split: Sp::new(text),
text,
trim_chunks,
}
}
/// If trim chunks is on, trim the str and adjust the offset
fn trim_chunk(&self, offset: usize, chunk: &'text str) -> (usize, &'text str) {
if self.trim_chunks {
self.semantic_split.trim_chunk(offset, chunk)
} else {
(offset, chunk)
}
}
/// Is the given text within the chunk size?
fn check_capacity(&self, offset: usize, chunk: &str) -> ChunkSize {
let (offset, chunk) = self.trim_chunk(offset, chunk);
let mut chunk_size = self.chunk_sizer.chunk_size(chunk, &self.chunk_capacity);
if let Some(max_chunk_size_offset) = chunk_size.max_chunk_size_offset.as_mut() {
*max_chunk_size_offset += offset;
}
chunk_size
}
/// Generate the next chunk, applying trimming settings.
/// Returns final byte offset and str.
/// Will return `None` if given an invalid range.
fn next_chunk(&mut self) -> Option<(usize, &'text str)> {
let start = self.cursor;
let mut end = self.cursor;
let mut equals_found = false;
let sections = self.next_sections()?.collect::<Vec<_>>();
let mut sizes = sections
.iter()
.map(|_| None)
.collect::<Vec<Option<ChunkSize>>>();
let mut low = 0;
let mut high = sections.len().saturating_sub(1);
let mut successful_index = None;
while low <= high {
let mid = low + (high - low) / 2;
let (offset, str) = sections[mid];
let text_end = offset + str.len();
let chunk = self.text.get(start..text_end)?;
let chunk_size = self.check_capacity(start, chunk);
sizes[mid] = Some(chunk_size);
match chunk_size.fits {
Ordering::Less => {
// We got further than the last one, so update end
if text_end > end {
end = text_end;
successful_index = Some(mid);
}
}
Ordering::Equal => {
// If we found a smaller equals use it. Or if this is the first equals we found
if text_end < end || !equals_found {
end = text_end;
successful_index = Some(mid);
}
equals_found = true;
}
Ordering::Greater => {
// If we're too big on our smallest run, we must return at least one section
if mid == 0 && start == end {
end = text_end;
successful_index = Some(mid);
}
}
};
// Adjust search area
if chunk_size.fits.is_lt() {
low = mid + 1;
} else if mid > 0 {
high = mid - 1;
} else {
// Nothing to adjust
break;
}
}
// Sometimes with tokenization, we can get a bigger chunk for the same amount of tokens.
if let Some((successful_index, chunk_size)) =
successful_index.and_then(|successful_index| {
Some((successful_index, sizes.get(successful_index)?.as_ref()?))
})
{
for (size, (offset, str)) in sizes.iter().zip(sections).skip(successful_index) {
let text_end = offset + str.len();
match size {
Some(size) if size.size <= chunk_size.size => {
if text_end > end {
end = text_end;
}
}
// We didn't tokenize this section yet
None => {
let chunk = self.text.get(start..text_end)?;
let size = self.check_capacity(start, chunk);
if size.size <= chunk_size.size {
if text_end > end {
end = text_end;
}
} else {
break;
}
}
_ => break,
}
}
}
self.cursor = end;
let chunk = self.text.get(start..self.cursor)?;
// Trim whitespace if user requested it
Some(self.trim_chunk(start, chunk))
}
/// Find the ideal next sections, breaking it up until we find the largest chunk.
/// Increasing length of chunk until we find biggest size to minimize validation time
/// on huge chunks
fn next_sections(&'sizer self) -> Option<impl Iterator<Item = (usize, &'text str)> + 'sizer> {
// Next levels to try. Will stop at max level. We check only levels in the next max level
// chunk so we don't bypass it if not all levels are present in every chunk.
let mut levels = self.semantic_split.levels_in_remaining_text(self.cursor);
// Get starting level
let mut semantic_level = levels.next()?;
// If we aren't at the highest semantic level, stop iterating sections that go beyond the range of the next level.
let mut max_encoded_offset = None;
for level in levels {
let (_, str) = self.semantic_chunks(level).next()?;
let chunk_size = self.check_capacity(self.cursor, str);
// If this no longer fits, we use the level we are at. Or if we already
// have the rest of the string
if chunk_size.fits.is_gt() || self.text.get(self.cursor..)? == str {
max_encoded_offset = chunk_size.max_chunk_size_offset;
break;
}
// Otherwise break up the text with the next level
semantic_level = level;
}
Some(
self.semantic_chunks(semantic_level)
// We don't want to return items at this level that go beyond the next highest semantic level, as that is most
// likely a meaningful breakpoint we want to preserve. We already know that the next highest doesn't fit anyway,
// so we should be safe to break once we reach it.
.take_while_inclusive(move |(offset, _)| {
max_encoded_offset.map_or(true, |max| offset <= &max)
})
.filter(|(_, str)| !str.is_empty()),
)
}
fn semantic_chunks(
&'sizer self,
level: <Sp as SemanticSplit>::Level,
) -> impl Iterator<Item = (usize, &'text str)> + 'sizer {
self.semantic_split.semantic_chunks(
self.cursor,
self.text.get(self.cursor..).unwrap(),
level,
)
}
}
impl<'sizer, 'text: 'sizer, C, S, Sp> Iterator for TextChunks<'text, 'sizer, C, S, Sp>
where
C: ChunkCapacity,
S: ChunkSizer,
Sp: SemanticSplit,
{
type Item = (usize, &'text str);
fn next(&mut self) -> Option<Self::Item> {
loop {
// Make sure we haven't reached the end
if self.cursor >= self.text.len() {
return None;
}
match self.next_chunk()? {
// Make sure we didn't get an empty chunk. Should only happen in
// cases where we trim.
(_, "") => continue,
c => return Some(c),
}
}
}
}
/// Given a list of separator ranges, construct the sections of the text
fn split_str_by_separator<L: Level>(
text: &str,
separator_ranges: impl Iterator<Item = (L, Range<usize>)>,
) -> impl Iterator<Item = (usize, &str)> {
let mut cursor = 0;
let mut final_match = false;
separator_ranges
.batching(move |it| {
loop {
match it.next() {
// If we've hit the end, actually return None
None if final_match => return None,
// First time we hit None, return the final section of the text
None => {
final_match = true;
return text.get(cursor..).map(|t| Either::Left(once((cursor, t))));
}
// Return text preceding match + the match
Some((level, range)) => {
let offset = cursor;
match level.split_position() {
SemanticSplitPosition::Prev => {
if range.end < cursor {
continue;
}
let section = text
.get(cursor..range.end)
.expect("invalid character sequence");
cursor = range.end;
return Some(Either::Left(once((offset, section))));
}
SemanticSplitPosition::Own => {
if range.start < cursor {
continue;
}
let prev_section = text
.get(cursor..range.start)
.expect("invalid character sequence");
if prev_section.trim().is_empty()
&& level.treat_whitespace_as_previous()
{
let section = text
.get(cursor..range.end)
.expect("invalid character sequence");
cursor = range.end;
return Some(Either::Left(once((offset, section))));
}
let separator = text
.get(range.start..range.end)
.expect("invalid character sequence");
cursor = range.end;
return Some(Either::Right(
[(offset, prev_section), (range.start, separator)].into_iter(),
));
}
SemanticSplitPosition::Next => {
if range.start < cursor {
continue;
}
let prev_section = text
.get(cursor..range.start)
.expect("invalid character sequence");
// Separator will be part of the next chunk
cursor = range.start;
return Some(Either::Left(once((offset, prev_section))));
}
}
}
}
}
})
.flatten()
.filter(|(_, s)| !s.is_empty())
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn check_chunk_capacity() {
let chunk = "12345";
assert_eq!(Characters.chunk_size(chunk, &4).fits, Ordering::Greater);
assert_eq!(Characters.chunk_size(chunk, &5).fits, Ordering::Equal);
assert_eq!(Characters.chunk_size(chunk, &6).fits, Ordering::Less);
}
#[test]
fn check_chunk_capacity_for_range() {
let chunk = "12345";
assert_eq!(
Characters.chunk_size(chunk, &(0..0)).fits,
Ordering::Greater
);
assert_eq!(
Characters.chunk_size(chunk, &(0..5)).fits,
Ordering::Greater
);
assert_eq!(Characters.chunk_size(chunk, &(5..6)).fits, Ordering::Equal);
assert_eq!(Characters.chunk_size(chunk, &(6..100)).fits, Ordering::Less);
}
#[test]
fn check_chunk_capacity_for_range_from() {
let chunk = "12345";
assert_eq!(Characters.chunk_size(chunk, &(0..)).fits, Ordering::Equal);
assert_eq!(Characters.chunk_size(chunk, &(5..)).fits, Ordering::Equal);
assert_eq!(Characters.chunk_size(chunk, &(6..)).fits, Ordering::Less);
}
#[test]
fn check_chunk_capacity_for_range_full() {
let chunk = "12345";
assert_eq!(Characters.chunk_size(chunk, &..).fits, Ordering::Equal);
}
#[test]
fn check_chunk_capacity_for_range_inclusive() {
let chunk = "12345";
assert_eq!(
Characters.chunk_size(chunk, &(0..=4)).fits,
Ordering::Greater
);
assert_eq!(Characters.chunk_size(chunk, &(5..=6)).fits, Ordering::Equal);
assert_eq!(Characters.chunk_size(chunk, &(4..=5)).fits, Ordering::Equal);
assert_eq!(
Characters.chunk_size(chunk, &(6..=100)).fits,
Ordering::Less
);
}
#[test]
fn check_chunk_capacity_for_range_to() {
let chunk = "12345";
assert_eq!(Characters.chunk_size(chunk, &(..0)).fits, Ordering::Greater);
assert_eq!(Characters.chunk_size(chunk, &(..5)).fits, Ordering::Greater);
assert_eq!(Characters.chunk_size(chunk, &(..6)).fits, Ordering::Equal);
}
#[test]
fn check_chunk_capacity_for_range_to_inclusive() {
let chunk = "12345";
assert_eq!(
Characters.chunk_size(chunk, &(..=4)).fits,
Ordering::Greater
);
assert_eq!(Characters.chunk_size(chunk, &(..=5)).fits, Ordering::Equal);
assert_eq!(Characters.chunk_size(chunk, &(..=6)).fits, Ordering::Equal);
}
#[test]
fn chunk_size_from_offsets() {
let offsets = [0..1, 1..2, 2..3];
let chunk_size = ChunkSize::from_offsets(offsets.clone().into_iter(), &1);
assert_eq!(
ChunkSize {
fits: Ordering::Greater,
size: offsets.len(),
max_chunk_size_offset: Some(1)
},
chunk_size
);
}
#[test]
fn chunk_size_from_empty_offsets() {
let offsets = [];
let chunk_size = ChunkSize::from_offsets(offsets.clone().into_iter(), &1);
assert_eq!(
ChunkSize {
fits: Ordering::Less,
size: offsets.len(),
max_chunk_size_offset: None
},
chunk_size
);
}
#[test]
fn chunk_size_from_small_offsets() {
let offsets = [0..1, 1..2, 2..3];
let chunk_size = ChunkSize::from_offsets(offsets.clone().into_iter(), &4);
assert_eq!(
ChunkSize {
fits: Ordering::Less,
size: offsets.len(),
max_chunk_size_offset: Some(3)
},
chunk_size
);
}
}