snowykami/meilisearch
1
0
Fork 0
mirror of https://github.com/meilisearch/meilisearch.git synced 2025-07-31 02:40:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support ngram typos + splitwords and splitwords+synonyms in proximity

Browse Source
This commit is contained in:
Loïc Lecrenier
2023-03-13 17:21:29 +01:00
parent 14e8d0aaa2
commit 3004e281d7
9 changed files with 701 additions and 411 deletions
Download Patch File Download Diff File Expand all files Collapse all files

439
milli/src/search/new/ranking_rule_graph/proximity/build.rs
View File

@ -1,37 +1,43 @@
#![allow(clippy::too_many_arguments)]
use std::collections::BTreeMap;
use itertools::Itertools;
use super::ProximityEdge;
use crate::search::new::interner::Interner;
use crate::search::new::query_term::{LocatedQueryTerm, QueryTerm, WordDerivations};
use crate::search::new::db_cache::DatabaseCache;
use crate::search::new::interner::{Interned, Interner};
use crate::search::new::query_term::{LocatedQueryTerm, Phrase, QueryTerm, WordDerivations};
use crate::search::new::ranking_rule_graph::proximity::WordPair;
use crate::search::new::ranking_rule_graph::EdgeCondition;
use crate::search::new::{QueryNode, SearchContext};
use crate::Result;
use heed::RoTxn;
pub fn visit_from_node(
ctx: &mut SearchContext,
from_node: &QueryNode,
) -> Result<Option<(WordDerivations, i8)>> {
Ok(Some(match from_node {
) -> Result<Option<(Vec<(Option<Interned<Phrase>>, Interned<String>)>, i8)>> {
let SearchContext { derivations_interner, .. } = ctx;
let (left_phrase, left_derivations, left_end_position) = match from_node {
QueryNode::Term(LocatedQueryTerm { value: value1, positions: pos1 }) => {
match value1 {
QueryTerm::Word { derivations } => {
(ctx.derivations_interner.get(*derivations).clone(), *pos1.end())
(None, derivations_interner.get(*derivations).clone(), *pos1.end())
}
QueryTerm::Phrase { phrase: phrase1 } => {
let phrase1 = ctx.phrase_interner.get(*phrase1);
if let Some(original) = *phrase1.words.last().unwrap() {
QueryTerm::Phrase { phrase: phrase_interned } => {
let phrase = ctx.phrase_interner.get(*phrase_interned);
if let Some(original) = *phrase.words.last().unwrap() {
(
Some(*phrase_interned),
WordDerivations {
original,
zero_typo: Box::new([original]),
zero_typo: Some(original),
one_typo: Box::new([]),
two_typos: Box::new([]),
use_prefix_db: false,
use_prefix_db: None,
synonyms: Box::new([]),
split_words: None,
is_prefix: false,
prefix_of: Box::new([]),
},
*pos1.end(),
)
@ -42,190 +48,175 @@ pub fn visit_from_node(
}
}
}
QueryNode::Start => (
WordDerivations {
original: ctx.word_interner.insert(String::new()),
zero_typo: Box::new([]),
one_typo: Box::new([]),
two_typos: Box::new([]),
use_prefix_db: false,
synonyms: Box::new([]),
split_words: None,
},
-100,
),
QueryNode::Start => (None, WordDerivations::empty(&mut ctx.word_interner, ""), -1),
_ => return Ok(None),
}))
};
// left term cannot be a prefix
assert!(left_derivations.use_prefix_db.is_none() && !left_derivations.is_prefix);
let last_word_left_phrase = if let Some(left_phrase_interned) = left_phrase {
let left_phrase = ctx.phrase_interner.get(left_phrase_interned);
left_phrase.words.last().copied().unwrap()
} else {
None
};
let left_single_word_iter: Vec<(Option<Interned<Phrase>>, Interned<String>)> = left_derivations
.all_single_word_derivations_except_prefix_db()
.chain(last_word_left_phrase.iter().copied())
.map(|w| (left_phrase, w))
.collect();
let left_phrase_iter: Vec<(Option<Interned<Phrase>>, Interned<String>)> = left_derivations
.all_phrase_derivations()
.map(|left_phrase_interned: Interned<Phrase>| {
let left_phrase = ctx.phrase_interner.get(left_phrase_interned);
let last_word_left_phrase: Interned<String> =
left_phrase.words.last().unwrap().unwrap();
let r: (Option<Interned<Phrase>>, Interned<String>) =
(Some(left_phrase_interned), last_word_left_phrase);
r
})
.collect();
let mut left_word_iter = left_single_word_iter;
left_word_iter.extend(left_phrase_iter);
Ok(Some((left_word_iter, left_end_position)))
}
pub fn visit_to_node<'ctx, 'from_data>(
pub fn build_step_visit_destination_node<'ctx, 'from_data>(
ctx: &mut SearchContext<'ctx>,
conditions_interner: &mut Interner<ProximityEdge>,
from_node_data: &'from_data (Vec<(Option<Interned<Phrase>>, Interned<String>)>, i8),
to_node: &QueryNode,
from_node_data: &'from_data (WordDerivations, i8),
) -> Result<Vec<(u8, EdgeCondition<ProximityEdge>)>> {
let SearchContext { index, txn, db_cache, word_interner, derivations_interner, .. } = ctx;
// IMPORTANT! TODO: split words support
let (derivations1, pos1) = from_node_data;
let term2 = match &to_node {
let SearchContext {
index,
txn,
db_cache,
word_interner,
phrase_interner,
derivations_interner,
query_term_docids: _,
} = ctx;
let right_term = match &to_node {
QueryNode::End => return Ok(vec![(0, EdgeCondition::Unconditional)]),
QueryNode::Deleted | QueryNode::Start => return Ok(vec![]),
QueryNode::Term(term) => term,
};
let LocatedQueryTerm { value: value2, positions: pos2 } = term2;
let LocatedQueryTerm { value: right_value, positions: right_positions } = right_term;
let (derivations2, pos2, ngram_len2) = match value2 {
QueryTerm::Word { derivations } => {
(derivations_interner.get(*derivations).clone(), *pos2.start(), pos2.len())
}
QueryTerm::Phrase { phrase: phrase2 } => {
let phrase2 = ctx.phrase_interner.get(*phrase2);
if let Some(original) = *phrase2.words.first().unwrap() {
(
WordDerivations {
original,
zero_typo: Box::new([original]),
one_typo: Box::new([]),
two_typos: Box::new([]),
use_prefix_db: false,
synonyms: Box::new([]),
split_words: None,
},
*pos2.start(),
1,
)
} else {
// No word pairs if the phrase does not have a regular word as its first term
return Ok(vec![]);
let (right_phrase, right_derivations, right_start_position, right_ngram_length) =
match right_value {
QueryTerm::Word { derivations } => (
None,
derivations_interner.get(*derivations).clone(),
*right_positions.start(),
right_positions.len(),
),
QueryTerm::Phrase { phrase: right_phrase_interned } => {
let right_phrase = phrase_interner.get(*right_phrase_interned);
if let Some(original) = *right_phrase.words.first().unwrap() {
(
Some(*right_phrase_interned),
WordDerivations {
original,
zero_typo: Some(original),
one_typo: Box::new([]),
two_typos: Box::new([]),
use_prefix_db: None,
synonyms: Box::new([]),
split_words: None,
is_prefix: false,
prefix_of: Box::new([]),
},
*right_positions.start(),
1,
)
} else {
// No word pairs if the phrase does not have a regular word as its first term
return Ok(vec![]);
}
}
}
};
};
if pos1 + 1 != pos2 {
// TODO: how should this actually be handled?
// We want to effectively ignore this pair of terms
let (left_derivations, left_end_position) = from_node_data;
if left_end_position + 1 != right_start_position {
// We want to ignore this pair of terms
// Unconditionally walk through the edge without computing the docids
// But also what should the cost be?
// This can happen when, in a query like `the sun flowers are beautiful`, the term
// `flowers` is removed by the words ranking rule due to the terms matching strategy.
// The remaining query graph represents `the sun .. are beautiful`
// but `sun` and `are` have no proximity condition between them
return Ok(vec![(0, EdgeCondition::Unconditional)]);
}
let updb1 = derivations1.use_prefix_db;
let updb2 = derivations2.use_prefix_db;
// left term cannot be a prefix
assert!(!updb1);
// TODO: IMPORTANT! split words and synonyms support
let derivations1 = derivations1.all_single_word_derivations_except_prefix_db();
// TODO: eventually, we want to get rid of the uses from `orginal`
let mut cost_proximity_word_pairs = BTreeMap::<u8, BTreeMap<u8, Vec<WordPair>>>::new();
if updb2 {
for word1 in derivations1.clone() {
for proximity in 1..=(8 - ngram_len2) {
let cost = (proximity + ngram_len2 - 1) as u8;
// TODO: if we had access to the universe here, we could already check whether
// the bitmap corresponding to this word pair is disjoint with the universe or not
if db_cache
.get_word_prefix_pair_proximity_docids(
index,
txn,
word_interner,
word1,
derivations2.original,
proximity as u8,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8)
.or_default()
.push(WordPair::WordPrefix {
left: word1,
right_prefix: derivations2.original,
});
}
if db_cache
.get_prefix_word_pair_proximity_docids(
index,
txn,
word_interner,
derivations2.original,
word1,
proximity as u8 - 1,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8)
.or_default()
.push(WordPair::WordPrefixSwapped {
left_prefix: derivations2.original,
right: word1,
});
}
}
if let Some(right_prefix) = right_derivations.use_prefix_db {
for (left_phrase, left_word) in left_derivations.iter().copied() {
add_prefix_edges(
index,
txn,
db_cache,
word_interner,
right_ngram_length,
left_word,
right_prefix,
&mut cost_proximity_word_pairs,
left_phrase,
)?;
}
}
// TODO: important! support split words and synonyms as well
let derivations2 = derivations2.all_single_word_derivations_except_prefix_db();
// TODO: add safeguard in case the cartesian product is too large!
// even if we restrict the word derivations to a maximum of 100, the size of the
// caterisan product could reach a maximum of 10_000 derivations, which is way too much.
// mMaybe prioritise the product of zero typo derivations, then the product of zero-typo/one-typo
// + one-typo/zero-typo, then one-typo/one-typo, then ... until an arbitrary limit has been
// reached
let product_derivations = derivations1.cartesian_product(derivations2);
let first_word_right_phrase = if let Some(right_phrase_interned) = right_phrase {
let right_phrase = phrase_interner.get(right_phrase_interned);
right_phrase.words.first().copied().unwrap()
} else {
None
};
let right_single_word_iter: Vec<(Option<Interned<Phrase>>, Interned<String>)> =
right_derivations
.all_single_word_derivations_except_prefix_db()
.chain(first_word_right_phrase.iter().copied())
.map(|w| (right_phrase, w))
.collect();
let right_phrase_iter: Vec<(Option<Interned<Phrase>>, Interned<String>)> = right_derivations
.all_phrase_derivations()
.map(|right_phrase_interned: Interned<Phrase>| {
let right_phrase = phrase_interner.get(right_phrase_interned);
let first_word_right_phrase: Interned<String> =
right_phrase.words.first().unwrap().unwrap();
let r: (Option<Interned<Phrase>>, Interned<String>) =
(Some(right_phrase_interned), first_word_right_phrase);
r
})
.collect();
let mut right_word_iter = right_single_word_iter;
right_word_iter.extend(right_phrase_iter);
for (word1, word2) in product_derivations {
for proximity in 1..=(8 - ngram_len2) {
let cost = (proximity + ngram_len2 - 1) as u8;
if db_cache
.get_word_pair_proximity_docids(
index,
txn,
word_interner,
word1,
word2,
proximity as u8,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8)
.or_default()
.push(WordPair::Words { left: word1, right: word2 });
}
if proximity > 1
&& db_cache
.get_word_pair_proximity_docids(
index,
txn,
word_interner,
word2,
word1,
proximity as u8 - 1,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8 - 1)
.or_default()
.push(WordPair::Words { left: word2, right: word1 });
}
for (left_phrase, left_word) in left_derivations.iter().copied() {
for (right_phrase, right_word) in right_word_iter.iter().copied() {
add_non_prefix_edges(
index,
txn,
db_cache,
word_interner,
right_ngram_length,
left_word,
right_word,
&mut cost_proximity_word_pairs,
&[left_phrase, right_phrase].iter().copied().flatten().collect::<Vec<_>>(),
)?;
}
}
let mut new_edges =
cost_proximity_word_pairs
.into_iter()
@ -243,6 +234,124 @@ pub fn visit_to_node<'ctx, 'from_data>(
edges
})
.collect::<Vec<_>>();
new_edges.push((8 + (ngram_len2 - 1) as u8, EdgeCondition::Unconditional));
new_edges.push((8 + (right_ngram_length - 1) as u8, EdgeCondition::Unconditional));
Ok(new_edges)
}
fn add_prefix_edges<'ctx>(
index: &mut &crate::Index,
txn: &'ctx RoTxn,
db_cache: &mut DatabaseCache<'ctx>,
word_interner: &mut Interner<String>,
right_ngram_length: usize,
left_word: Interned<String>,
right_prefix: Interned<String>,
cost_proximity_word_pairs: &mut BTreeMap<u8, BTreeMap<u8, Vec<WordPair>>>,
left_phrase: Option<Interned<Phrase>>,
) -> Result<()> {
for proximity in 1..=(8 - right_ngram_length) {
let cost = (proximity + right_ngram_length - 1) as u8;
// TODO: if we had access to the universe here, we could already check whether
// the bitmap corresponding to this word pair is disjoint with the universe or not
if db_cache
.get_word_prefix_pair_proximity_docids(
index,
txn,
word_interner,
left_word,
right_prefix,
proximity as u8,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8)
.or_default()
.push(WordPair::WordPrefix {
phrases: left_phrase.into_iter().collect(),
left: left_word,
right_prefix,
});
}
// No swapping when computing the proximity between a phrase and a word
if left_phrase.is_none()
&& db_cache
.get_prefix_word_pair_proximity_docids(
index,
txn,
word_interner,
right_prefix,
left_word,
proximity as u8 - 1,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8)
.or_default()
.push(WordPair::WordPrefixSwapped { left_prefix: right_prefix, right: left_word });
}
}
Ok(())
}
fn add_non_prefix_edges<'ctx>(
index: &mut &crate::Index,
txn: &'ctx RoTxn,
db_cache: &mut DatabaseCache<'ctx>,
word_interner: &mut Interner<String>,
right_ngram_length: usize,
word1: Interned<String>,
word2: Interned<String>,
cost_proximity_word_pairs: &mut BTreeMap<u8, BTreeMap<u8, Vec<WordPair>>>,
phrases: &[Interned<Phrase>],
) -> Result<()> {
for proximity in 1..=(8 - right_ngram_length) {
let cost = (proximity + right_ngram_length - 1) as u8;
if db_cache
.get_word_pair_proximity_docids(
index,
txn,
word_interner,
word1,
word2,
proximity as u8,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8)
.or_default()
.push(WordPair::Words { phrases: phrases.to_vec(), left: word1, right: word2 });
}
if proximity > 1
// no swapping when either term is a phrase
&& phrases.is_empty()
&& db_cache
.get_word_pair_proximity_docids(
index,
txn,
word_interner,
word2,
word1,
proximity as u8 - 1,
)?
.is_some()
{
cost_proximity_word_pairs
.entry(cost)
.or_default()
.entry(proximity as u8 - 1)
.or_default()
.push(WordPair::Words { phrases: vec![], left: word2, right: word1 });
}
}
Ok(())
}

85
milli/src/search/new/ranking_rule_graph/proximity/compute_docids.rs
View File

@ -13,24 +13,61 @@ pub fn compute_docids<'ctx>(
let ProximityEdge { pairs, proximity } = edge;
let mut pair_docids = RoaringBitmap::new();
for pair in pairs.iter() {
let bytes = match pair {
WordPair::Words { left, right } => db_cache.get_word_pair_proximity_docids(
index,
txn,
word_interner,
*left,
*right,
*proximity,
),
WordPair::WordPrefix { left, right_prefix } => db_cache
.get_word_prefix_pair_proximity_docids(
index,
txn,
word_interner,
*left,
*right_prefix,
*proximity,
),
let pair = match pair {
WordPair::Words { phrases, left, right } => {
let mut docids = db_cache
.get_word_pair_proximity_docids(
index,
txn,
word_interner,
*left,
*right,
*proximity,
)?
.map(CboRoaringBitmapCodec::deserialize_from)
.transpose()?
.unwrap_or_default();
if !docids.is_empty() {
for phrase in phrases {
docids &= ctx.query_term_docids.get_phrase_docids(
index,
txn,
db_cache,
word_interner,
&ctx.phrase_interner,
*phrase,
)?;
}
}
docids
}
WordPair::WordPrefix { phrases, left, right_prefix } => {
let mut docids = db_cache
.get_word_prefix_pair_proximity_docids(
index,
txn,
word_interner,
*left,
*right_prefix,
*proximity,
)?
.map(CboRoaringBitmapCodec::deserialize_from)
.transpose()?
.unwrap_or_default();
if !docids.is_empty() {
for phrase in phrases {
docids &= ctx.query_term_docids.get_phrase_docids(
index,
txn,
db_cache,
word_interner,
&ctx.phrase_interner,
*phrase,
)?;
}
}
docids
}
WordPair::WordPrefixSwapped { left_prefix, right } => db_cache
.get_prefix_word_pair_proximity_docids(
index,
@ -39,11 +76,13 @@ pub fn compute_docids<'ctx>(
*left_prefix,
*right,
*proximity,
),
}?;
// TODO: deserialize bitmap within a universe, and (maybe) using a bump allocator?
let bitmap = universe
& bytes.map(CboRoaringBitmapCodec::deserialize_from).transpose()?.unwrap_or_default();
)?
.map(CboRoaringBitmapCodec::deserialize_from)
.transpose()?
.unwrap_or_default(),
};
// TODO: deserialize bitmap within a universe
let bitmap = universe & pair;
pair_docids |= bitmap;
}

32
milli/src/search/new/ranking_rule_graph/proximity/mod.rs
View File

@ -7,16 +7,27 @@ use super::empty_paths_cache::EmptyPathsCache;
use super::{EdgeCondition, RankingRuleGraphTrait};
use crate::search::new::interner::{Interned, Interner};
use crate::search::new::logger::SearchLogger;
use crate::search::new::query_term::WordDerivations;
use crate::search::new::query_term::Phrase;
use crate::search::new::small_bitmap::SmallBitmap;
use crate::search::new::{QueryGraph, QueryNode, SearchContext};
use crate::Result;
#[derive(Clone, PartialEq, Eq, Hash)]
pub enum WordPair {
Words { left: Interned<String>, right: Interned<String> },
WordPrefix { left: Interned<String>, right_prefix: Interned<String> },
WordPrefixSwapped { left_prefix: Interned<String>, right: Interned<String> },
Words {
phrases: Vec<Interned<Phrase>>,
left: Interned<String>,
right: Interned<String>,
},
WordPrefix {
phrases: Vec<Interned<Phrase>>,
left: Interned<String>,
right_prefix: Interned<String>,
},
WordPrefixSwapped {
left_prefix: Interned<String>,
right: Interned<String>,
},
}
#[derive(Clone, PartialEq, Eq, Hash)]
@ -29,7 +40,7 @@ pub enum ProximityGraph {}
impl RankingRuleGraphTrait for ProximityGraph {
type EdgeCondition = ProximityEdge;
type BuildVisitedFromNode = (WordDerivations, i8);
type BuildVisitedFromNode = (Vec<(Option<Interned<Phrase>>, Interned<String>)>, i8);
fn label_for_edge_condition(edge: &Self::EdgeCondition) -> String {
let ProximityEdge { pairs, proximity } = edge;
@ -54,10 +65,15 @@ impl RankingRuleGraphTrait for ProximityGraph {
fn build_step_visit_destination_node<'from_data, 'ctx: 'from_data>(
ctx: &mut SearchContext<'ctx>,
conditions_interner: &mut Interner<Self::EdgeCondition>,
to_node: &QueryNode,
from_node_data: &'from_data Self::BuildVisitedFromNode,
dest_node: &QueryNode,
source_node_data: &'from_data Self::BuildVisitedFromNode,
) -> Result<Vec<(u8, EdgeCondition<Self::EdgeCondition>)>> {
build::visit_to_node(ctx, conditions_interner, to_node, from_node_data)
build::build_step_visit_destination_node(
ctx,
conditions_interner,
source_node_data,
dest_node,
)
}
fn log_state(

47
milli/src/search/new/ranking_rule_graph/typo/mod.rs
View File

@ -84,7 +84,7 @@ impl RankingRuleGraphTrait for TypoGraph {
) -> Result<Vec<(u8, EdgeCondition<Self::EdgeCondition>)>> {
let SearchContext { derivations_interner, .. } = ctx;
match to_node {
QueryNode::Term(LocatedQueryTerm { value, .. }) => match *value {
QueryNode::Term(LocatedQueryTerm { value, positions }) => match *value {
QueryTerm::Phrase { phrase } => Ok(vec![(
0,
EdgeCondition::Conditional(
@ -93,57 +93,62 @@ impl RankingRuleGraphTrait for TypoGraph {
)]),
QueryTerm::Word { derivations } => {
let mut edges = vec![];
// Ngrams have a base typo cost
// 2-gram -> equivalent to 1 typo
// 3-gram -> equivalent to 2 typos
let base_cost = positions.len().max(2) as u8;
for nbr_typos in 0..=2 {
let derivations = derivations_interner.get(derivations).clone();
let new_derivations = match nbr_typos {
0 => {
// TODO: think about how split words and synonyms should be handled here
// TODO: what about ngrams?
// Maybe 2grams should have one typo by default and 3grams 2 typos by default
WordDerivations {
original: derivations.original,
synonyms: derivations.synonyms,
split_words: None,
zero_typo: derivations.zero_typo,
one_typo: Box::new([]),
two_typos: Box::new([]),
use_prefix_db: derivations.use_prefix_db,
}
}
0 => WordDerivations {
original: derivations.original,
is_prefix: derivations.is_prefix,
zero_typo: derivations.zero_typo,
prefix_of: derivations.prefix_of,
synonyms: derivations.synonyms,
split_words: None,
one_typo: Box::new([]),
two_typos: Box::new([]),
use_prefix_db: derivations.use_prefix_db,
},
1 => {
// What about split words and synonyms here?
WordDerivations {
original: derivations.original,
is_prefix: false,
zero_typo: None,
prefix_of: Box::new([]),
synonyms: Box::new([]),
split_words: derivations.split_words,
zero_typo: Box::new([]),
one_typo: derivations.one_typo,
two_typos: Box::new([]),
use_prefix_db: false, // false because all items from use_prefix_db haev 0 typos
use_prefix_db: None, // false because all items from use_prefix_db have 0 typos
}
}
2 => {
// What about split words and synonyms here?
WordDerivations {
original: derivations.original,
zero_typo: None,
is_prefix: false,
prefix_of: Box::new([]),
synonyms: Box::new([]),
split_words: None,
zero_typo: Box::new([]),
one_typo: Box::new([]),
two_typos: derivations.two_typos,
use_prefix_db: false, // false because all items from use_prefix_db haev 0 typos
use_prefix_db: None, // false because all items from use_prefix_db have 0 typos
}
}
_ => panic!(),
};
if !new_derivations.is_empty() {
edges.push((
nbr_typos,
nbr_typos as u8 + base_cost,
EdgeCondition::Conditional(conditions_interner.insert(
TypoEdge::Word {
derivations: derivations_interner.insert(new_derivations),
nbr_typos,
nbr_typos: nbr_typos as u8,
},
)),
))