Add exactness ranking rules

This commit is contained in:
Louis Dureuil 2023-04-04 17:12:07 +02:00 committed by Louis Dureuil
parent 1b8e4d0301
commit 8a13ed7e3f
5 changed files with 301 additions and 6 deletions

View File

@ -0,0 +1,175 @@
use heed::BytesDecode;
use roaring::MultiOps;
use super::query_graph::QueryGraph;
use super::ranking_rules::{RankingRule, RankingRuleOutput};
use crate::search::new::query_graph::QueryNodeData;
use crate::search::new::query_term::ExactTerm;
use crate::{CboRoaringBitmapCodec, Result, SearchContext, SearchLogger};
/// FIXME:
///
/// - A lot of work done in next_bucket that start_iteration could do.
/// - Consider calling the graph based rule directly from this one.
/// - currently we did exact term, don't forget about prefix
/// - some tests
pub struct ExactAttribute {
query_graph: Option<QueryGraph>,
}
impl ExactAttribute {
pub fn new() -> Self {
Self { query_graph: None }
}
}
impl<'ctx> RankingRule<'ctx, QueryGraph> for ExactAttribute {
fn id(&self) -> String {
"exact_attribute".to_owned()
}
fn start_iteration(
&mut self,
_ctx: &mut SearchContext<'ctx>,
_logger: &mut dyn SearchLogger<QueryGraph>,
_universe: &roaring::RoaringBitmap,
query: &QueryGraph,
) -> Result<()> {
self.query_graph = Some(query.clone());
Ok(())
}
fn next_bucket(
&mut self,
ctx: &mut SearchContext<'ctx>,
_logger: &mut dyn SearchLogger<QueryGraph>,
universe: &roaring::RoaringBitmap,
) -> Result<Option<RankingRuleOutput<QueryGraph>>> {
// iterate on the nodes of the graph, retain LocatedQueryTermSubset
let query_graph = self.query_graph.as_ref().unwrap();
let mut exact_term_position_ids: Vec<(ExactTerm, u16, u8)> =
Vec::with_capacity(query_graph.nodes.len() as usize);
for (_, node) in query_graph.nodes.iter() {
match &node.data {
QueryNodeData::Term(term) => {
let exact_term = if let Some(exact_term) = term.term_subset.exact_term(ctx) {
exact_term
} else {
// FIXME: Use `None` or some function indicating that we're passing down the bucket to our child rules
return Ok(Some(RankingRuleOutput {
query: query_graph.clone(),
candidates: universe.clone(),
}));
};
exact_term_position_ids.push((
exact_term,
*term.positions.start(),
*term.term_ids.start(),
))
}
QueryNodeData::Deleted | QueryNodeData::Start | QueryNodeData::End => continue,
}
}
exact_term_position_ids.sort_by_key(|(_, _, id)| *id);
// bail if there is a "hole" (missing word) in remaining query graph
let mut previous_id = 0;
for (_, _, id) in exact_term_position_ids.iter().copied() {
if id < previous_id || id - previous_id > 1 {
// FIXME: Use `None` or some function indicating that we're passing down the bucket to our child rules
return Ok(Some(RankingRuleOutput {
query: query_graph.clone(),
candidates: universe.clone(),
}));
} else {
previous_id = id;
}
}
// sample query: "sunflower are pretty"
// sunflower at pos 0 in attr A
// are at pos 1 in attr B
// pretty at pos 2 in attr C
// We want to eliminate such document
// first check that for each term, there exists some attribute that has this term at the correct position
//"word-position-docids";
let mut candidates = universe.clone();
let words_positions: Vec<(Vec<_>, _)> = exact_term_position_ids
.iter()
.copied()
.map(|(term, position, _)| (term.interned_words(ctx).collect(), position))
.collect();
for (words, position) in &words_positions {
if candidates.is_empty() {
// FIXME: Use `None` or some function indicating that we're passing down the bucket to our child rules
return Ok(Some(RankingRuleOutput {
query: query_graph.clone(),
candidates: universe.clone(),
}));
}
'words: for (offset, word) in words.iter().enumerate() {
let offset = offset as u16;
let word = if let Some(word) = word {
word
} else {
continue 'words;
};
let word_position_docids = CboRoaringBitmapCodec::bytes_decode(
ctx.get_db_word_position_docids(*word, position + offset)?.unwrap_or_default(),
)
.unwrap_or_default();
candidates &= word_position_docids;
}
}
let candidates = candidates;
if candidates.is_empty() {
// FIXME: Use `None` or some function indicating that we're passing down the bucket to our child rules
return Ok(Some(RankingRuleOutput {
query: query_graph.clone(),
candidates: universe.clone(),
}));
}
let searchable_fields_ids = ctx.index.searchable_fields_ids(ctx.txn)?.unwrap_or_default();
let mut candidates_per_attributes = Vec::with_capacity(searchable_fields_ids.len());
// then check that there exists at least one attribute that has all of the terms
for fid in searchable_fields_ids {
let mut intersection = MultiOps::intersection(
words_positions
.iter()
.flat_map(|(words, ..)| words.iter())
// ignore stop words words in phrases
.flatten()
.map(|word| -> Result<_> {
Ok(ctx
.get_db_word_fid_docids(*word, fid)?
.map(CboRoaringBitmapCodec::bytes_decode)
.unwrap_or_default()
.unwrap_or_default())
}),
)?;
intersection &= &candidates;
if !intersection.is_empty() {
candidates_per_attributes.push(intersection);
}
}
// note we could have "false positives" where there both exist different attributes that collectively
// have the terms in the correct order and a single attribute that have all the terms, but in the incorrect order.
let candidates = MultiOps::union(candidates_per_attributes.into_iter());
Ok(Some(RankingRuleOutput { query: query_graph.clone(), candidates }))
}
fn end_iteration(
&mut self,
_ctx: &mut SearchContext<'ctx>,
_logger: &mut dyn SearchLogger<QueryGraph>,
) {
}
}

View File

@ -44,8 +44,8 @@ use super::interner::{Interned, MappedInterner};
use super::logger::SearchLogger; use super::logger::SearchLogger;
use super::query_graph::QueryNode; use super::query_graph::QueryNode;
use super::ranking_rule_graph::{ use super::ranking_rule_graph::{
ConditionDocIdsCache, DeadEndsCache, ProximityGraph, RankingRuleGraph, RankingRuleGraphTrait, ConditionDocIdsCache, DeadEndsCache, ExactnessGraph, ProximityGraph, RankingRuleGraph,
TypoGraph, RankingRuleGraphTrait, TypoGraph,
}; };
use super::small_bitmap::SmallBitmap; use super::small_bitmap::SmallBitmap;
use super::{QueryGraph, RankingRule, RankingRuleOutput, SearchContext}; use super::{QueryGraph, RankingRule, RankingRuleOutput, SearchContext};
@ -65,6 +65,12 @@ impl GraphBasedRankingRule<TypoGraph> {
Self::new_with_id("typo".to_owned(), terms_matching_strategy) Self::new_with_id("typo".to_owned(), terms_matching_strategy)
} }
} }
pub type Exactness = GraphBasedRankingRule<ExactnessGraph>;
impl GraphBasedRankingRule<ExactnessGraph> {
pub fn new() -> Self {
Self::new_with_id("exactness".to_owned(), None)
}
}
/// A generic graph-based ranking rule /// A generic graph-based ranking rule
pub struct GraphBasedRankingRule<G: RankingRuleGraphTrait> { pub struct GraphBasedRankingRule<G: RankingRuleGraphTrait> {

View File

@ -9,8 +9,9 @@ mod query_term;
mod ranking_rule_graph; mod ranking_rule_graph;
mod ranking_rules; mod ranking_rules;
mod resolve_query_graph; mod resolve_query_graph;
// TODO: documentation + comments
mod small_bitmap; mod small_bitmap;
mod exact_attribute;
// TODO: documentation + comments // TODO: documentation + comments
// implementation is currently an adaptation of the previous implementation to fit with the new model // implementation is currently an adaptation of the previous implementation to fit with the new model
mod sort; mod sort;
@ -33,6 +34,8 @@ use resolve_query_graph::PhraseDocIdsCache;
use roaring::RoaringBitmap; use roaring::RoaringBitmap;
use words::Words; use words::Words;
use self::exact_attribute::ExactAttribute;
use self::graph_based_ranking_rule::Exactness;
use self::interner::Interner; use self::interner::Interner;
use self::ranking_rules::{BoxRankingRule, RankingRule}; use self::ranking_rules::{BoxRankingRule, RankingRule};
use self::resolve_query_graph::compute_query_graph_docids; use self::resolve_query_graph::compute_query_graph_docids;
@ -150,7 +153,7 @@ fn get_ranking_rules_for_query_graph_search<'ctx>(
let mut proximity = false; let mut proximity = false;
let mut sort = false; let mut sort = false;
let attribute = false; let attribute = false;
let exactness = false; let mut exactness = false;
let mut asc = HashSet::new(); let mut asc = HashSet::new();
let mut desc = HashSet::new(); let mut desc = HashSet::new();
@ -211,8 +214,9 @@ fn get_ranking_rules_for_query_graph_search<'ctx>(
if exactness { if exactness {
continue; continue;
} }
// todo!(); ranking_rules.push(Box::new(ExactAttribute::new()));
// exactness = false; ranking_rules.push(Box::new(Exactness::new()));
exactness = true;
} }
crate::Criterion::Asc(field_name) => { crate::Criterion::Asc(field_name) => {
if asc.contains(&field_name) { if asc.contains(&field_name) {

View File

@ -0,0 +1,107 @@
use roaring::RoaringBitmap;
use super::{ComputedCondition, DeadEndsCache, RankingRuleGraph, RankingRuleGraphTrait};
use crate::search::new::interner::{DedupInterner, Interned, MappedInterner};
use crate::search::new::query_graph::{QueryGraph, QueryNode};
use crate::search::new::query_term::{ExactTerm, LocatedQueryTermSubset};
use crate::{CboRoaringBitmapCodec, Result, SearchContext, SearchLogger};
/// - Exactness as first ranking rule: TermsMatchingStrategy? prefer a document that matches 1 word exactly and no other
/// word than a doc that matches 9 words non exactly but none exactly
/// - `TermsMatchingStrategy` as a word + exactness optimization: we could consider
///
/// "naive vision"
/// condition from one node to another:
/// - word exactly present: cost 0
/// - word typo/ngram/prefix/missing: cost 1, not remove from query graph, edge btwn the two nodes, return the universe without condition when resolving, destination query term is inside
///
/// Three strategies:
/// 1. ExactAttribute: word position / word_fid_docid
/// 2. AttributeStart:
/// 3. AttributeContainsExact => implementable via `RankingRuleGraphTrait`
#[derive(Clone, PartialEq, Eq, Hash)]
pub enum ExactnessCondition {
ExactInAttribute(LocatedQueryTermSubset),
Skip(LocatedQueryTermSubset),
}
pub enum ExactnessGraph {}
fn compute_docids(
ctx: &mut SearchContext,
dest_node: &LocatedQueryTermSubset,
universe: &RoaringBitmap,
) -> Result<RoaringBitmap> {
let exact_term = if let Some(exact_term) = dest_node.term_subset.exact_term(ctx) {
exact_term
} else {
return Ok(Default::default());
};
let mut candidates = match exact_term {
ExactTerm::Phrase(phrase) => ctx.get_phrase_docids(phrase)?.clone(),
ExactTerm::Word(word) => {
if let Some(word_candidates) = ctx.get_db_word_docids(word)? {
CboRoaringBitmapCodec::deserialize_from(word_candidates)?
} else {
return Ok(Default::default());
}
}
};
// TODO: synonyms?
candidates &= universe;
Ok(candidates)
}
impl RankingRuleGraphTrait for ExactnessGraph {
type Condition = ExactnessCondition;
fn resolve_condition(
ctx: &mut SearchContext,
condition: &Self::Condition,
universe: &RoaringBitmap,
) -> Result<ComputedCondition> {
let (docids, dest_node) = match condition {
ExactnessCondition::ExactInAttribute(dest_node) => {
(compute_docids(ctx, dest_node, universe)?, dest_node)
}
ExactnessCondition::Skip(dest_node) => (universe.clone(), dest_node),
};
Ok(ComputedCondition {
docids,
universe_len: universe.len(),
start_term_subset: None,
end_term_subset: dest_node.clone(),
})
}
fn build_edges(
_ctx: &mut SearchContext,
conditions_interner: &mut DedupInterner<Self::Condition>,
_source_node: Option<&LocatedQueryTermSubset>,
dest_node: &LocatedQueryTermSubset,
) -> Result<Vec<(u32, Interned<Self::Condition>)>> {
let exact_condition = ExactnessCondition::ExactInAttribute(dest_node.clone());
let exact_condition = conditions_interner.insert(exact_condition);
let skip_condition = ExactnessCondition::Skip(dest_node.clone());
let skip_condition = conditions_interner.insert(skip_condition);
Ok(vec![(0, exact_condition), (1, skip_condition)])
}
fn log_state(
graph: &RankingRuleGraph<Self>,
paths: &[Vec<Interned<Self::Condition>>],
dead_ends_cache: &DeadEndsCache<Self::Condition>,
universe: &RoaringBitmap,
costs: &MappedInterner<QueryNode, Vec<u64>>,
cost: u64,
logger: &mut dyn SearchLogger<QueryGraph>,
) {
todo!()
}
fn label_for_condition(ctx: &mut SearchContext, condition: &Self::Condition) -> Result<String> {
todo!()
}
}

View File

@ -10,6 +10,8 @@ mod cheapest_paths;
mod condition_docids_cache; mod condition_docids_cache;
mod dead_ends_cache; mod dead_ends_cache;
/// Implementation of the `exactness` ranking rule
mod exactness;
/// Implementation of the `proximity` ranking rule /// Implementation of the `proximity` ranking rule
mod proximity; mod proximity;
/// Implementation of the `typo` ranking rule /// Implementation of the `typo` ranking rule
@ -20,6 +22,7 @@ use std::hash::Hash;
pub use cheapest_paths::PathVisitor; pub use cheapest_paths::PathVisitor;
pub use condition_docids_cache::ConditionDocIdsCache; pub use condition_docids_cache::ConditionDocIdsCache;
pub use dead_ends_cache::DeadEndsCache; pub use dead_ends_cache::DeadEndsCache;
pub use exactness::{ExactnessCondition, ExactnessGraph};
pub use proximity::{ProximityCondition, ProximityGraph}; pub use proximity::{ProximityCondition, ProximityGraph};
use roaring::RoaringBitmap; use roaring::RoaringBitmap;
pub use typo::{TypoCondition, TypoGraph}; pub use typo::{TypoCondition, TypoGraph};