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

Merge Phrase and WordDerivations into one structure

Browse Source
This commit is contained in:
Loïc Lecrenier
2023-03-14 10:54:55 +01:00
parent 3004e281d7
commit 31628c5cd4
11 changed files with 335 additions and 645 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
milli/src/search/new/ranking_rule_graph/build.rs
View File

@ -22,28 +22,21 @@ impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
let mut edges_store = vec![];
let mut edges_of_node = vec![];
for (node_idx, node) in graph_nodes.iter().enumerate() {
for (source_idx, source_node) in graph_nodes.iter().enumerate() {
edges_of_node.push(HashSet::new());
let new_edges = edges_of_node.last_mut().unwrap();
let Some(source_node_data) = G::build_step_visit_source_node(ctx, node)? else { continue };
for successor_idx in graph_edges[node_idx].successors.iter() {
let dest_node = &graph_nodes[successor_idx as usize];
let edges = G::build_step_visit_destination_node(
ctx,
&mut conditions_interner,
dest_node,
&source_node_data,
)?;
for dest_idx in graph_edges[source_idx].successors.iter() {
let dest_node = &graph_nodes[dest_idx as usize];
let edges = G::build_edges(ctx, &mut conditions_interner, source_node, dest_node)?;
if edges.is_empty() {
continue;
}
for (cost, condition) in edges {
edges_store.push(Some(Edge {
source_node: node_idx as u16,
dest_node: successor_idx,
source_node: source_idx as u16,
dest_node: dest_idx,
cost,
condition,
}));

18
milli/src/search/new/ranking_rule_graph/mod.rs
View File

@ -80,11 +80,6 @@ pub trait RankingRuleGraphTrait: Sized {
/// in [`resolve_edge_condition`](RankingRuleGraphTrait::resolve_edge_condition).
type EdgeCondition: Sized + Clone + PartialEq + Eq + Hash;
/// A structure used in the construction of the graph, created when a
/// query graph source node is visited. It is used to determine the cost
/// and condition of a ranking rule edge when the destination node is visited.
type BuildVisitedFromNode;
/// Return the label of the given edge condition, to be used when visualising
/// the ranking rule graph.
fn label_for_edge_condition(edge: &Self::EdgeCondition) -> String;
@ -97,22 +92,13 @@ pub trait RankingRuleGraphTrait: Sized {
universe: &RoaringBitmap,
) -> Result<RoaringBitmap>;
/// Prepare to build the edges outgoing from `source_node`.
///
/// This call is followed by zero, one or more calls to [`build_step_visit_destination_node`](RankingRuleGraphTrait::build_step_visit_destination_node),
/// which builds the actual edges.
fn build_step_visit_source_node<'ctx>(
ctx: &mut SearchContext<'ctx>,
source_node: &QueryNode,
) -> Result<Option<Self::BuildVisitedFromNode>>;
/// Return the cost and condition of the edges going from the previously visited node
/// (with [`build_step_visit_source_node`](RankingRuleGraphTrait::build_step_visit_source_node)) to `dest_node`.
fn build_step_visit_destination_node<'from_data, 'ctx: 'from_data>(
fn build_edges<'ctx>(
ctx: &mut SearchContext<'ctx>,
conditions_interner: &mut Interner<Self::EdgeCondition>,
source_node: &QueryNode,
dest_node: &QueryNode,
source_node_data: &'from_data Self::BuildVisitedFromNode,
) -> Result<Vec<(u8, EdgeCondition<Self::EdgeCondition>)>>;
fn log_state(

185
milli/src/search/new/ranking_rule_graph/proximity/build.rs
View File

@ -4,89 +4,40 @@ use std::collections::BTreeMap;
use super::ProximityEdge;
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::query_term::{LocatedQueryTerm, Phrase, QueryTerm};
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<(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 } => {
(None, derivations_interner.get(*derivations).clone(), *pos1.end())
}
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: 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([]),
},
*pos1.end(),
)
} else {
// No word pairs if the phrase does not have a regular word as its last term
return Ok(None);
}
}
}
}
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)))
fn last_word_of_term_iter<'t>(
t: &'t QueryTerm,
phrase_interner: &'t Interner<Phrase>,
) -> impl Iterator<Item = (Option<Interned<Phrase>>, Interned<String>)> + 't {
t.all_single_words_except_prefix_db().map(|w| (None, w)).chain(t.all_phrases().flat_map(
move |p| {
let phrase = phrase_interner.get(p);
phrase.words.last().unwrap().map(|last| (Some(p), last))
},
))
}
fn first_word_of_term_iter<'t>(
t: &'t QueryTerm,
phrase_interner: &'t Interner<Phrase>,
) -> impl Iterator<Item = (Interned<String>, Option<Interned<Phrase>>)> + 't {
t.all_single_words_except_prefix_db().map(|w| (w, None)).chain(t.all_phrases().flat_map(
move |p| {
let phrase = phrase_interner.get(p);
phrase.words.first().unwrap().map(|first| (first, Some(p)))
},
))
}
pub fn build_step_visit_destination_node<'ctx, 'from_data>(
pub fn build_edges<'ctx>(
ctx: &mut SearchContext<'ctx>,
conditions_interner: &mut Interner<ProximityEdge>,
from_node_data: &'from_data (Vec<(Option<Interned<Phrase>>, Interned<String>)>, i8),
from_node: &QueryNode,
to_node: &QueryNode,
) -> Result<Vec<(u8, EdgeCondition<ProximityEdge>)>> {
let SearchContext {
@ -95,9 +46,19 @@ pub fn build_step_visit_destination_node<'ctx, 'from_data>(
db_cache,
word_interner,
phrase_interner,
derivations_interner,
query_term_docids: _,
term_interner,
term_docids: _,
} = ctx;
let (left_term, left_end_position) = match from_node {
QueryNode::Term(LocatedQueryTerm { value, positions }) => {
(term_interner.get(*value), *positions.end())
}
QueryNode::Deleted => return Ok(vec![]),
QueryNode::Start => return Ok(vec![(0, EdgeCondition::Unconditional)]),
QueryNode::End => return Ok(vec![]),
};
let right_term = match &to_node {
QueryNode::End => return Ok(vec![(0, EdgeCondition::Unconditional)]),
QueryNode::Deleted | QueryNode::Start => return Ok(vec![]),
@ -105,47 +66,14 @@ pub fn build_step_visit_destination_node<'ctx, 'from_data>(
};
let LocatedQueryTerm { value: right_value, positions: right_positions } = right_term;
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![]);
}
}
};
let (left_derivations, left_end_position) = from_node_data;
let (right_term, right_start_position, right_ngram_length) =
(term_interner.get(*right_value), *right_positions.start(), right_positions.len());
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
// 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.
// `flowers` is removed by the `words` ranking rule.
// 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)]);
@ -153,8 +81,8 @@ pub fn build_step_visit_destination_node<'ctx, 'from_data>(
let mut cost_proximity_word_pairs = BTreeMap::<u8, BTreeMap<u8, Vec<WordPair>>>::new();
if let Some(right_prefix) = right_derivations.use_prefix_db {
for (left_phrase, left_word) in left_derivations.iter().copied() {
if let Some(right_prefix) = right_term.use_prefix_db {
for (left_phrase, left_word) in last_word_of_term_iter(left_term, phrase_interner) {
add_prefix_edges(
index,
txn,
@ -172,37 +100,12 @@ pub fn build_step_visit_destination_node<'ctx, 'from_data>(
// 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
// Maybe 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 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 (left_phrase, left_word) in left_derivations.iter().copied() {
for (right_phrase, right_word) in right_word_iter.iter().copied() {
for (left_phrase, left_word) in last_word_of_term_iter(left_term, phrase_interner) {
for (right_word, right_phrase) in first_word_of_term_iter(right_term, phrase_interner) {
add_non_prefix_edges(
index,
txn,

4
milli/src/search/new/ranking_rule_graph/proximity/compute_docids.rs
View File

@ -29,7 +29,7 @@ pub fn compute_docids<'ctx>(
.unwrap_or_default();
if !docids.is_empty() {
for phrase in phrases {
docids &= ctx.query_term_docids.get_phrase_docids(
docids &= ctx.term_docids.get_phrase_docids(
index,
txn,
db_cache,
@ -56,7 +56,7 @@ pub fn compute_docids<'ctx>(
.unwrap_or_default();
if !docids.is_empty() {
for phrase in phrases {
docids &= ctx.query_term_docids.get_phrase_docids(
docids &= ctx.term_docids.get_phrase_docids(
index,
txn,
db_cache,

19
milli/src/search/new/ranking_rule_graph/proximity/mod.rs
View File

@ -40,7 +40,6 @@ pub enum ProximityGraph {}
impl RankingRuleGraphTrait for ProximityGraph {
type EdgeCondition = ProximityEdge;
type BuildVisitedFromNode = (Vec<(Option<Interned<Phrase>>, Interned<String>)>, i8);
fn label_for_edge_condition(edge: &Self::EdgeCondition) -> String {
let ProximityEdge { pairs, proximity } = edge;
@ -55,25 +54,13 @@ impl RankingRuleGraphTrait for ProximityGraph {
compute_docids::compute_docids(ctx, edge, universe)
}
fn build_step_visit_source_node<'ctx>(
ctx: &mut SearchContext<'ctx>,
from_node: &QueryNode,
) -> Result<Option<Self::BuildVisitedFromNode>> {
build::visit_from_node(ctx, from_node)
}
fn build_step_visit_destination_node<'from_data, 'ctx: 'from_data>(
fn build_edges<'ctx>(
ctx: &mut SearchContext<'ctx>,
conditions_interner: &mut Interner<Self::EdgeCondition>,
source_node: &QueryNode,
dest_node: &QueryNode,
source_node_data: &'from_data Self::BuildVisitedFromNode,
) -> Result<Vec<(u8, EdgeCondition<Self::EdgeCondition>)>> {
build::build_step_visit_destination_node(
ctx,
conditions_interner,
source_node_data,
dest_node,
)
build::build_edges(ctx, conditions_interner, source_node, dest_node)
}
fn log_state(

195
milli/src/search/new/ranking_rule_graph/typo/mod.rs
View File

@ -4,28 +4,24 @@ use super::empty_paths_cache::EmptyPathsCache;
use super::{EdgeCondition, RankingRuleGraph, RankingRuleGraphTrait};
use crate::search::new::interner::{Interned, Interner};
use crate::search::new::logger::SearchLogger;
use crate::search::new::query_term::{LocatedQueryTerm, Phrase, QueryTerm, WordDerivations};
use crate::search::new::query_term::{LocatedQueryTerm, QueryTerm};
use crate::search::new::small_bitmap::SmallBitmap;
use crate::search::new::{QueryGraph, QueryNode, SearchContext};
use crate::Result;
#[derive(Clone, PartialEq, Eq, Hash)]
pub enum TypoEdge {
Phrase { phrase: Interned<Phrase> },
Word { derivations: Interned<WordDerivations>, nbr_typos: u8 },
pub struct TypoEdge {
term: Interned<QueryTerm>,
nbr_typos: u8,
}
pub enum TypoGraph {}
impl RankingRuleGraphTrait for TypoGraph {
type EdgeCondition = TypoEdge;
type BuildVisitedFromNode = ();
fn label_for_edge_condition(edge: &Self::EdgeCondition) -> String {
match edge {
TypoEdge::Phrase { .. } => ", 0 typos".to_owned(),
TypoEdge::Word { nbr_typos, .. } => format!(", {nbr_typos} typos"),
}
format!(", {} typos", edge.nbr_typos)
}
fn resolve_edge_condition<'db_cache, 'ctx>(
@ -39,124 +35,101 @@ impl RankingRuleGraphTrait for TypoGraph {
db_cache,
word_interner,
phrase_interner,
derivations_interner,
query_term_docids,
term_interner,
term_docids: query_term_docids,
} = ctx;
match edge {
&TypoEdge::Phrase { phrase } => Ok(universe
& query_term_docids.get_phrase_docids(
index,
txn,
db_cache,
word_interner,
phrase_interner,
phrase,
)?),
TypoEdge::Word { derivations, .. } => {
let docids = universe
& query_term_docids.get_word_derivations_docids(
index,
txn,
db_cache,
word_interner,
derivations_interner,
phrase_interner,
*derivations,
)?;
Ok(docids)
}
}
let docids = universe
& query_term_docids.get_query_term_docids(
index,
txn,
db_cache,
word_interner,
term_interner,
phrase_interner,
edge.term,
)?;
Ok(docids)
}
fn build_step_visit_source_node<'ctx>(
_ctx: &mut SearchContext<'ctx>,
_from_node: &QueryNode,
) -> Result<Option<Self::BuildVisitedFromNode>> {
Ok(Some(()))
}
fn build_step_visit_destination_node<'from_data, 'ctx: 'from_data>(
fn build_edges<'ctx>(
ctx: &mut SearchContext<'ctx>,
conditions_interner: &mut Interner<Self::EdgeCondition>,
_from_node: &QueryNode,
to_node: &QueryNode,
_from_node_data: &'from_data Self::BuildVisitedFromNode,
) -> Result<Vec<(u8, EdgeCondition<Self::EdgeCondition>)>> {
let SearchContext { derivations_interner, .. } = ctx;
let SearchContext { term_interner, .. } = ctx;
match to_node {
QueryNode::Term(LocatedQueryTerm { value, positions }) => match *value {
QueryTerm::Phrase { phrase } => Ok(vec![(
0,
EdgeCondition::Conditional(
conditions_interner.insert(TypoEdge::Phrase { phrase }),
),
)]),
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;
QueryNode::Term(LocatedQueryTerm { value, positions }) => {
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 => WordDerivations {
original: derivations.original,
is_prefix: derivations.is_prefix,
zero_typo: derivations.zero_typo,
prefix_of: derivations.prefix_of,
synonyms: derivations.synonyms,
for nbr_typos in 0..=2 {
let term = term_interner.get(*value).clone();
let new_term = match nbr_typos {
0 => QueryTerm {
original: term.original,
is_prefix: term.is_prefix,
zero_typo: term.zero_typo,
prefix_of: term.prefix_of,
synonyms: term.synonyms,
split_words: None,
one_typo: Box::new([]),
two_typos: Box::new([]),
use_prefix_db: term.use_prefix_db,
is_ngram: term.is_ngram,
phrase: term.phrase,
},
1 => {
// What about split words and synonyms here?
QueryTerm {
original: term.original,
is_prefix: false,
zero_typo: None,
prefix_of: Box::new([]),
synonyms: Box::new([]),
split_words: term.split_words,
one_typo: term.one_typo,
two_typos: Box::new([]),
use_prefix_db: None, // false because all items from use_prefix_db have 0 typos
is_ngram: term.is_ngram,
phrase: None,
}
}
2 => {
// What about split words and synonyms here?
QueryTerm {
original: term.original,
zero_typo: None,
is_prefix: false,
prefix_of: Box::new([]),
synonyms: Box::new([]),
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,
one_typo: derivations.one_typo,
two_typos: Box::new([]),
use_prefix_db: None, // false because all items from use_prefix_db have 0 typos
}
two_typos: term.two_typos,
use_prefix_db: None, // false because all items from use_prefix_db have 0 typos
is_ngram: term.is_ngram,
phrase: None,
}
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,
one_typo: Box::new([]),
two_typos: derivations.two_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 as u8 + base_cost,
EdgeCondition::Conditional(conditions_interner.insert(
TypoEdge::Word {
derivations: derivations_interner.insert(new_derivations),
nbr_typos: nbr_typos as u8,
},
)),
))
}
_ => panic!(),
};
if !new_term.is_empty() {
edges.push((
nbr_typos as u8 + base_cost,
EdgeCondition::Conditional(conditions_interner.insert(TypoEdge {
term: term_interner.insert(new_term),
nbr_typos: nbr_typos as u8,
})),
))
}
Ok(edges)
}
},
Ok(edges)
}
QueryNode::End => Ok(vec![(0, EdgeCondition::Unconditional)]),
QueryNode::Deleted | QueryNode::Start => panic!(),
}