integration with new tokenizer wip

This commit is contained in:
mpostma 2020-11-19 18:23:08 +01:00 committed by many
parent 8a4d05b7bb
commit 5e00842087
No known key found for this signature in database
GPG Key ID: 2CEF23B75189EACA
5 changed files with 94 additions and 60 deletions

View File

@ -2,13 +2,3 @@ mod dfa;
pub use self::dfa::{build_dfa, build_prefix_dfa, build_exact_dfa};
pub fn normalize_str(string: &str) -> String {
let mut string = string.to_lowercase();
if !string.contains(is_cjk) {
string = deunicode::deunicode_with_tofu(&string, "");
}
string
}

View File

@ -7,13 +7,14 @@ use std::{cmp, fmt, iter::once};
use fst::{IntoStreamer, Streamer};
use itertools::{EitherOrBoth, merge_join_by};
use meilisearch_tokenizer::split_query_string;
use sdset::{Set, SetBuf, SetOperation};
use log::debug;
use meilisearch_tokenizer::Token;
use meilisearch_tokenizer::tokenizer::{Analyzer, AnalyzerConfig};
use sdset::{Set, SetBuf, SetOperation};
use crate::database::MainT;
use crate::{store, DocumentId, DocIndex, MResult, FstSetCow};
use crate::automaton::{normalize_str, build_dfa, build_prefix_dfa, build_exact_dfa};
use crate::automaton::{build_dfa, build_prefix_dfa, build_exact_dfa};
use crate::QueryWordsMapper;
#[derive(Clone, PartialEq, Eq, Hash)]
@ -146,7 +147,7 @@ fn split_best_frequency<'a>(reader: &heed::RoTxn<MainT>, ctx: &Context, word: &'
}
fn fetch_synonyms(reader: &heed::RoTxn<MainT>, ctx: &Context, words: &[&str]) -> MResult<Vec<Vec<String>>> {
let words = normalize_str(&words.join(" "));
let words = &words.join(" ");
let set = ctx.synonyms.synonyms_fst(reader, words.as_bytes())?;
let mut strings = Vec::new();
@ -174,15 +175,25 @@ where I: IntoIterator<Item=Operation>,
const MAX_NGRAM: usize = 3;
fn split_query_string(s: &str) -> Vec<(usize, String)> {
// TODO: Use global instance instead
let analyzer = Analyzer::new(AnalyzerConfig::default());
analyzer
.analyze(s)
.tokens()
.filter(|t| !t.is_stopword())
.enumerate()
.map(|(i, Token { word, .. })| (i, word.to_string()))
.collect()
}
pub fn create_query_tree(
reader: &heed::RoTxn<MainT>,
ctx: &Context,
query: &str,
) -> MResult<(Operation, HashMap<QueryId, Range<usize>>)>
{
let words = split_query_string(query).map(str::to_lowercase);
let words = words.filter(|w| !ctx.stop_words.contains(w));
let words: Vec<_> = words.enumerate().collect();
let words = split_query_string(query);
let mut mapper = QueryWordsMapper::new(words.iter().map(|(_, w)| w));

View File

@ -2,8 +2,9 @@ use std::borrow::Cow;
use std::collections::{BTreeMap, HashMap};
use std::convert::TryFrom;
use deunicode::deunicode_with_tofu;
use meilisearch_schema::IndexedPos;
use meilisearch_tokenizer::tokenizer::{Analyzer, AnalyzerConfig};
use meilisearch_tokenizer::Token;
use sdset::SetBuf;
use crate::{DocIndex, DocumentId};
@ -18,6 +19,7 @@ pub struct RawIndexer<A> {
stop_words: fst::Set<A>,
words_doc_indexes: BTreeMap<Word, Vec<DocIndex>>,
docs_words: HashMap<DocumentId, Vec<Word>>,
analyzer: Analyzer,
}
pub struct Indexed<'a> {
@ -36,6 +38,7 @@ impl<A> RawIndexer<A> {
stop_words,
words_doc_indexes: BTreeMap::new(),
docs_words: HashMap::new(),
analyzer: Analyzer::new(AnalyzerConfig::default()),
}
}
}
@ -44,9 +47,12 @@ impl<A: AsRef<[u8]>> RawIndexer<A> {
pub fn index_text(&mut self, id: DocumentId, indexed_pos: IndexedPos, text: &str) -> usize {
let mut number_of_words = 0;
for token in Tokenizer::new(text) {
let analyzed_text = self.analyzer.analyze(text);
for (word_pos, (token_index, token)) in analyzed_text.tokens().enumerate().filter(|(_, t)| !t.is_separator()).enumerate() {
let must_continue = index_token(
token,
token_index,
word_pos,
id,
indexed_pos,
self.word_limit,
@ -69,20 +75,47 @@ impl<A: AsRef<[u8]>> RawIndexer<A> {
where
I: IntoIterator<Item = &'s str>,
{
let iter = iter.into_iter();
for token in SeqTokenizer::new(iter) {
let must_continue = index_token(
token,
id,
indexed_pos,
self.word_limit,
&self.stop_words,
&mut self.words_doc_indexes,
&mut self.docs_words,
);
let mut token_index_offset = 0;
let mut byte_offset = 0;
let mut word_offset = 0;
if !must_continue {
break;
for s in iter.into_iter() {
let current_token_index_offset = token_index_offset;
let current_byte_offset = byte_offset;
let current_word_offset = word_offset;
let analyzed_text = self.analyzer.analyze(s);
let tokens = analyzed_text
.tokens()
.enumerate()
.map(|(i, mut t)| {
t.byte_start = t.byte_start + current_byte_offset;
t.byte_end = t.byte_end + current_byte_offset;
(i + current_token_index_offset, t)
})
.enumerate()
.map(|(i, t)| (i + current_word_offset, t));
for (word_pos, (token_index, token)) in tokens {
token_index_offset = token_index + 1;
word_offset = word_pos + 1;
byte_offset = token.byte_end + 1;
let must_continue = index_token(
token,
token_index,
word_pos,
id,
indexed_pos,
self.word_limit,
&self.stop_words,
&mut self.words_doc_indexes,
&mut self.docs_words,
);
if !must_continue {
break;
}
}
}
}
@ -114,6 +147,8 @@ impl<A: AsRef<[u8]>> RawIndexer<A> {
fn index_token<A>(
token: Token,
position: usize,
word_pos: usize,
id: DocumentId,
indexed_pos: IndexedPos,
word_limit: usize,
@ -123,20 +158,14 @@ fn index_token<A>(
) -> bool
where A: AsRef<[u8]>,
{
if token.index >= word_limit {
if position >= word_limit {
return false;
}
let lower = token.word.to_lowercase();
let token = Token {
word: &lower,
..token
};
if !stop_words.contains(&token.word) {
match token_to_docindex(id, indexed_pos, token) {
if !stop_words.contains(&token.word.as_ref()) {
match token_to_docindex(id, indexed_pos, &token, word_pos) {
Some(docindex) => {
let word = Vec::from(token.word);
let word = Vec::from(token.word.as_ref());
if word.len() <= WORD_LENGTH_LIMIT {
words_doc_indexes
@ -144,20 +173,6 @@ where A: AsRef<[u8]>,
.or_insert_with(Vec::new)
.push(docindex);
docs_words.entry(id).or_insert_with(Vec::new).push(word);
if !lower.contains(is_cjk) {
let unidecoded = deunicode_with_tofu(&lower, "");
if unidecoded != lower && !unidecoded.is_empty() {
let word = Vec::from(unidecoded);
if word.len() <= WORD_LENGTH_LIMIT {
words_doc_indexes
.entry(word.clone())
.or_insert_with(Vec::new)
.push(docindex);
docs_words.entry(id).or_insert_with(Vec::new).push(word);
}
}
}
}
}
None => return false,
@ -167,8 +182,8 @@ where A: AsRef<[u8]>,
true
}
fn token_to_docindex(id: DocumentId, indexed_pos: IndexedPos, token: Token) -> Option<DocIndex> {
let word_index = u16::try_from(token.word_index).ok()?;
fn token_to_docindex(id: DocumentId, indexed_pos: IndexedPos, token: &Token, word_index: usize) -> Option<DocIndex> {
let word_index = u16::try_from(word_index).ok()?;
let char_index = u16::try_from(token.char_index).ok()?;
let char_length = u16::try_from(token.word.chars().count()).ok()?;

View File

@ -11,7 +11,6 @@ use meilisearch_core::criterion::*;
use meilisearch_core::settings::RankingRule;
use meilisearch_core::{Highlight, Index, RankedMap};
use meilisearch_schema::{FieldId, Schema};
use meilisearch_tokenizer::is_cjk;
use serde::{Deserialize, Serialize};
use serde_json::Value;
use siphasher::sip::SipHasher;
@ -344,7 +343,7 @@ pub struct SearchResult {
/// returns the start index and the length on the crop.
fn aligned_crop(text: &str, match_index: usize, context: usize) -> (usize, usize) {
let is_word_component = |c: &char| c.is_alphanumeric() && !is_cjk(*c);
let is_word_component = |c: &char| c.is_alphanumeric() && !super::is_cjk(*c);
let word_end_index = |mut index| {
if text.chars().nth(index - 1).map_or(false, |c| is_word_component(&c)) {

View File

@ -5,3 +5,22 @@ pub mod compression;
pub use authentication::Authentication;
pub use normalize_path::NormalizePath;
pub fn is_cjk(c: char) -> bool {
(c >= '\u{1100}' && c <= '\u{11ff}') // Hangul Jamo
|| (c >= '\u{2e80}' && c <= '\u{2eff}') // CJK Radicals Supplement
|| (c >= '\u{2f00}' && c <= '\u{2fdf}') // Kangxi radical
|| (c >= '\u{3000}' && c <= '\u{303f}') // Japanese-style punctuation
|| (c >= '\u{3040}' && c <= '\u{309f}') // Japanese Hiragana
|| (c >= '\u{30a0}' && c <= '\u{30ff}') // Japanese Katakana
|| (c >= '\u{3100}' && c <= '\u{312f}')
|| (c >= '\u{3130}' && c <= '\u{318F}') // Hangul Compatibility Jamo
|| (c >= '\u{3200}' && c <= '\u{32ff}') // Enclosed CJK Letters and Months
|| (c >= '\u{3400}' && c <= '\u{4dbf}') // CJK Unified Ideographs Extension A
|| (c >= '\u{4e00}' && c <= '\u{9fff}') // CJK Unified Ideographs
|| (c >= '\u{a960}' && c <= '\u{a97f}') // Hangul Jamo Extended-A
|| (c >= '\u{ac00}' && c <= '\u{d7a3}') // Hangul Syllables
|| (c >= '\u{d7b0}' && c <= '\u{d7ff}') // Hangul Jamo Extended-B
|| (c >= '\u{f900}' && c <= '\u{faff}') // CJK Compatibility Ideographs
|| (c >= '\u{ff00}' && c <= '\u{ffef}') // Full-width roman characters and half-width katakana
}