meilisearch/milli/src/search/matching_words.rs

281 lines
9.3 KiB
Rust

use std::cmp::{min, Reverse};
use std::collections::{BTreeMap, HashSet};
use std::ops::{Index, IndexMut};
use levenshtein_automata::{Distance, DFA};
use super::build_dfa;
use crate::search::query_tree::{Operation, Query};
type IsPrefix = bool;
/// Structure created from a query tree
/// referencing words that match the given query tree.
#[derive(Default)]
pub struct MatchingWords {
dfas: Vec<(DFA, String, u8, IsPrefix)>,
}
impl MatchingWords {
pub fn from_query_tree(tree: &Operation) -> Self {
// fetch matchable words from the query tree
let mut dfas: Vec<_> = fetch_queries(tree)
.into_iter()
// create DFAs for each word
.map(|(w, t, p)| (build_dfa(w, t, p), w.to_string(), t, p))
.collect();
// Sort word by len in DESC order prioritizing the longuest word,
// in order to highlight the longuest part of the matched word.
dfas.sort_unstable_by_key(|(_dfa, query_word, _typo, _is_prefix)| {
Reverse(query_word.len())
});
Self { dfas }
}
/// Returns the number of matching bytes if the word matches one of the query words.
pub fn matching_bytes(&self, word_to_highlight: &str) -> Option<usize> {
self.dfas.iter().find_map(|(dfa, query_word, typo, is_prefix)| {
match dfa.eval(word_to_highlight) {
Distance::Exact(t) if t <= *typo => {
if *is_prefix {
let len = bytes_to_highlight(word_to_highlight, query_word);
Some(len)
} else {
Some(word_to_highlight.len())
}
}
_otherwise => None,
}
})
}
}
/// Lists all words which can be considered as a match for the query tree.
fn fetch_queries(tree: &Operation) -> HashSet<(&str, u8, IsPrefix)> {
fn resolve_ops<'a>(tree: &'a Operation, out: &mut HashSet<(&'a str, u8, IsPrefix)>) {
match tree {
Operation::Or(_, ops) | Operation::And(ops) => {
ops.as_slice().iter().for_each(|op| resolve_ops(op, out));
}
Operation::Query(Query { prefix, kind }) => {
let typo = if kind.is_exact() { 0 } else { kind.typo() };
out.insert((kind.word(), typo, *prefix));
}
Operation::Phrase(words) => {
for word in words {
out.insert((word, 0, false));
}
}
}
}
let mut queries = HashSet::new();
resolve_ops(tree, &mut queries);
queries
}
// A simple wrapper around vec so we can get contiguous but index it like it's 2D array.
struct N2Array<T> {
y_size: usize,
buf: Vec<T>,
}
impl<T: Clone> N2Array<T> {
fn new(x: usize, y: usize, value: T) -> N2Array<T> {
N2Array { y_size: y, buf: vec![value; x * y] }
}
}
impl<T> Index<(usize, usize)> for N2Array<T> {
type Output = T;
#[inline]
fn index(&self, (x, y): (usize, usize)) -> &T {
&self.buf[(x * self.y_size) + y]
}
}
impl<T> IndexMut<(usize, usize)> for N2Array<T> {
#[inline]
fn index_mut(&mut self, (x, y): (usize, usize)) -> &mut T {
&mut self.buf[(x * self.y_size) + y]
}
}
/// Returns the number of **bytes** we want to highlight in the `source` word.
/// Basically we want to highlight as much characters as possible in the source until it has too much
/// typos (= 2)
/// The algorithm is a modified
/// [Damerau-Levenshtein](https://en.wikipedia.org/wiki/Damerau%E2%80%93Levenshtein_distance)
fn bytes_to_highlight(source: &str, target: &str) -> usize {
let n = source.chars().count();
let m = target.chars().count();
if n == 0 {
return 0;
}
// since we allow two typos we can send two characters even if it's completely wrong
if m < 3 {
return source.chars().take(m).map(|c| c.len_utf8()).sum();
}
if n == m && source == target {
return source.len();
}
let inf = n + m;
let mut matrix = N2Array::new(n + 2, m + 2, 0);
matrix[(0, 0)] = inf;
for i in 0..=n {
matrix[(i + 1, 0)] = inf;
matrix[(i + 1, 1)] = i;
}
for j in 0..=m {
matrix[(0, j + 1)] = inf;
matrix[(1, j + 1)] = j;
}
let mut last_row = BTreeMap::new();
for (row, char_s) in source.chars().enumerate() {
let mut last_match_col = 0;
let row = row + 1;
for (col, char_t) in target.chars().enumerate() {
let col = col + 1;
let last_match_row = *last_row.get(&char_t).unwrap_or(&0);
let cost = if char_s == char_t { 0 } else { 1 };
let dist_add = matrix[(row, col + 1)] + 1;
let dist_del = matrix[(row + 1, col)] + 1;
let dist_sub = matrix[(row, col)] + cost;
let dist_trans = matrix[(last_match_row, last_match_col)]
+ (row - last_match_row - 1)
+ 1
+ (col - last_match_col - 1);
let dist = min(min(dist_add, dist_del), min(dist_sub, dist_trans));
matrix[(row + 1, col + 1)] = dist;
if cost == 0 {
last_match_col = col;
}
}
last_row.insert(char_s, row);
}
let mut minimum = (u32::max_value(), 0);
for x in 0..=m {
let dist = matrix[(n + 1, x + 1)] as u32;
if dist < minimum.0 {
minimum = (dist, x);
}
}
// everything was done characters wise and now we want to returns a number of bytes
source.chars().take(minimum.1).map(|c| c.len_utf8()).sum()
}
#[cfg(test)]
mod tests {
use std::str::from_utf8;
use super::*;
use crate::search::query_tree::{Operation, Query, QueryKind};
use crate::MatchingWords;
#[test]
fn test_bytes_to_highlight() {
struct TestBytesToHighlight {
query: &'static str,
text: &'static str,
length: usize,
}
let tests = [
TestBytesToHighlight { query: "bip", text: "bip", length: "bip".len() },
TestBytesToHighlight { query: "bip", text: "boup", length: "bip".len() },
TestBytesToHighlight {
query: "Levenshtein",
text: "Levenshtein",
length: "Levenshtein".len(),
},
// we get to the end of our word with only one typo
TestBytesToHighlight {
query: "Levenste",
text: "Levenshtein",
length: "Levenste".len(),
},
// we get our third and last authorized typo right on the last character
TestBytesToHighlight {
query: "Levenstein",
text: "Levenshte",
length: "Levenste".len(),
},
// we get to the end of our word with only two typos at the beginning
TestBytesToHighlight {
query: "Bavenshtein",
text: "Levenshtein",
length: "Bavenshtein".len(),
},
TestBytesToHighlight {
query: "Альфа", text: "Альфой", length: "Альф".len()
},
TestBytesToHighlight {
query: "Go💼", text: "Go💼od luck.", length: "Go💼".len()
},
TestBytesToHighlight {
query: "Go💼od", text: "Go💼od luck.", length: "Go💼od".len()
},
TestBytesToHighlight {
query: "chäräcters",
text: "chäräcters",
length: "chäräcters".len(),
},
TestBytesToHighlight { query: "ch", text: "chäräcters", length: "ch".len() },
TestBytesToHighlight { query: "chär", text: "chäräcters", length: "chär".len() },
];
for test in &tests {
let length = bytes_to_highlight(test.text, test.query);
assert_eq!(length, test.length, r#"lenght between: "{}" "{}""#, test.query, test.text);
assert!(
from_utf8(&test.query.as_bytes()[..length]).is_ok(),
r#"converting {}[..{}] to an utf8 str failed"#,
test.query,
length
);
}
}
#[test]
fn matching_words() {
let query_tree = Operation::Or(
false,
vec![Operation::And(vec![
Operation::Query(Query {
prefix: true,
kind: QueryKind::exact("split".to_string()),
}),
Operation::Query(Query {
prefix: false,
kind: QueryKind::exact("this".to_string()),
}),
Operation::Query(Query {
prefix: true,
kind: QueryKind::tolerant(1, "world".to_string()),
}),
])],
);
let matching_words = MatchingWords::from_query_tree(&query_tree);
assert_eq!(matching_words.matching_bytes("word"), Some(3));
assert_eq!(matching_words.matching_bytes("nyc"), None);
assert_eq!(matching_words.matching_bytes("world"), Some(5));
assert_eq!(matching_words.matching_bytes("splitted"), Some(5));
assert_eq!(matching_words.matching_bytes("thisnew"), None);
assert_eq!(matching_words.matching_bytes("borld"), Some(5));
assert_eq!(matching_words.matching_bytes("wordsplit"), Some(4));
}
}