MeiliSearch/milli/src/search/criteria/proximity.rs

499 lines
19 KiB
Rust
Raw Normal View History

2021-03-04 16:07:07 +01:00
use std::collections::{BTreeMap, HashMap, btree_map};
2021-02-22 17:17:01 +01:00
use std::mem::take;
use roaring::RoaringBitmap;
2021-02-24 15:37:37 +01:00
use log::debug;
2021-02-22 17:17:01 +01:00
use crate::{DocumentId, Position, search::{query_tree::QueryKind, word_derivations}};
2021-02-22 17:17:01 +01:00
use crate::search::query_tree::{maximum_proximity, Operation, Query};
use super::{Candidates, Criterion, CriterionResult, Context, query_docids, query_pair_proximity_docids};
pub struct Proximity<'t> {
ctx: &'t dyn Context,
query_tree: Option<(usize, Operation)>,
proximity: u8,
candidates: Candidates,
2021-02-25 16:54:41 +01:00
bucket_candidates: RoaringBitmap,
2021-02-22 17:17:01 +01:00
parent: Option<Box<dyn Criterion + 't>>,
candidates_cache: HashMap<(Operation, u8), Vec<(Query, Query, RoaringBitmap)>>,
2021-03-04 16:07:07 +01:00
plane_sweep_cache: Option<btree_map::IntoIter<u8, RoaringBitmap>>,
2021-02-22 17:17:01 +01:00
}
impl<'t> Proximity<'t> {
pub fn initial(
ctx: &'t dyn Context,
query_tree: Option<Operation>,
candidates: Option<RoaringBitmap>,
) -> Self
2021-02-22 17:17:01 +01:00
{
Proximity {
2021-02-22 17:17:01 +01:00
ctx,
query_tree: query_tree.map(|op| (maximum_proximity(&op), op)),
proximity: 0,
candidates: candidates.map_or_else(Candidates::default, Candidates::Allowed),
2021-02-25 16:54:41 +01:00
bucket_candidates: RoaringBitmap::new(),
2021-02-22 17:17:01 +01:00
parent: None,
candidates_cache: HashMap::new(),
2021-03-04 16:07:07 +01:00
plane_sweep_cache: None,
}
2021-02-22 17:17:01 +01:00
}
pub fn new(ctx: &'t dyn Context, parent: Box<dyn Criterion + 't>) -> Self {
Proximity {
2021-02-22 17:17:01 +01:00
ctx,
query_tree: None,
proximity: 0,
candidates: Candidates::default(),
2021-02-25 16:54:41 +01:00
bucket_candidates: RoaringBitmap::new(),
2021-02-22 17:17:01 +01:00
parent: Some(parent),
candidates_cache: HashMap::new(),
2021-03-04 16:07:07 +01:00
plane_sweep_cache: None,
}
2021-02-22 17:17:01 +01:00
}
}
impl<'t> Criterion for Proximity<'t> {
fn next(&mut self) -> anyhow::Result<Option<CriterionResult>> {
use Candidates::{Allowed, Forbidden};
loop {
2021-02-24 15:37:37 +01:00
debug!("Proximity at iteration {} (max {:?}) ({:?})",
self.proximity,
self.query_tree.as_ref().map(|(mp, _)| mp),
self.candidates,
);
2021-02-22 17:17:01 +01:00
match (&mut self.query_tree, &mut self.candidates) {
(_, Allowed(candidates)) if candidates.is_empty() => {
return Ok(Some(CriterionResult {
query_tree: self.query_tree.take().map(|(_, qt)| qt),
candidates: take(&mut self.candidates).into_inner(),
bucket_candidates: take(&mut self.bucket_candidates),
}));
2021-02-22 17:17:01 +01:00
},
(Some((max_prox, query_tree)), Allowed(candidates)) => {
if self.proximity as usize > *max_prox {
2021-03-04 16:07:07 +01:00
// reset state to (None, Forbidden(_))
2021-02-22 17:17:01 +01:00
self.query_tree = None;
self.candidates = Candidates::default();
} else {
2021-03-04 16:07:07 +01:00
let mut new_candidates = if candidates.len() <= 1000 {
if let Some(cache) = self.plane_sweep_cache.as_mut() {
match cache.next() {
Some((p, candidates)) => {
self.proximity = p;
candidates
},
None => {
// reset state to (None, Forbidden(_))
self.query_tree = None;
self.candidates = Candidates::default();
continue
},
}
} else {
let cache = resolve_plane_sweep_candidates(
self.ctx,
query_tree,
candidates
)?;
self.plane_sweep_cache = Some(cache.into_iter());
continue
}
} else { // use set theory based algorithm
resolve_candidates(
self.ctx,
&query_tree,
self.proximity,
&mut self.candidates_cache,
)?
};
2021-02-22 17:17:01 +01:00
new_candidates.intersect_with(&candidates);
candidates.difference_with(&new_candidates);
self.proximity += 1;
let bucket_candidates = match self.parent {
2021-02-25 16:54:41 +01:00
Some(_) => take(&mut self.bucket_candidates),
None => new_candidates.clone(),
2021-02-22 17:17:01 +01:00
};
return Ok(Some(CriterionResult {
query_tree: Some(query_tree.clone()),
candidates: new_candidates,
bucket_candidates,
}));
}
},
(Some((max_prox, query_tree)), Forbidden(candidates)) => {
if self.proximity as usize > *max_prox {
self.query_tree = None;
self.candidates = Candidates::default();
} else {
let mut new_candidates = resolve_candidates(
self.ctx,
&query_tree,
self.proximity,
&mut self.candidates_cache,
)?;
new_candidates.difference_with(&candidates);
candidates.union_with(&new_candidates);
self.proximity += 1;
let bucket_candidates = match self.parent {
2021-02-25 16:54:41 +01:00
Some(_) => take(&mut self.bucket_candidates),
None => new_candidates.clone(),
2021-02-22 17:17:01 +01:00
};
return Ok(Some(CriterionResult {
query_tree: Some(query_tree.clone()),
candidates: new_candidates,
bucket_candidates,
}));
}
},
(None, Allowed(_)) => {
let candidates = take(&mut self.candidates).into_inner();
return Ok(Some(CriterionResult {
query_tree: None,
candidates: candidates.clone(),
2021-02-25 16:54:41 +01:00
bucket_candidates: candidates,
2021-02-22 17:17:01 +01:00
}));
},
(None, Forbidden(_)) => {
match self.parent.as_mut() {
Some(parent) => {
match parent.next()? {
Some(CriterionResult { query_tree, candidates, bucket_candidates }) => {
self.query_tree = query_tree.map(|op| (maximum_proximity(&op), op));
self.proximity = 0;
self.candidates = Candidates::Allowed(candidates);
2021-02-25 17:28:20 +01:00
self.bucket_candidates.union_with(&bucket_candidates);
2021-03-04 16:07:07 +01:00
self.plane_sweep_cache = None;
2021-02-22 17:17:01 +01:00
},
None => return Ok(None),
}
},
None => return Ok(None),
}
},
}
}
}
}
fn resolve_candidates<'t>(
ctx: &'t dyn Context,
query_tree: &Operation,
proximity: u8,
cache: &mut HashMap<(Operation, u8), Vec<(Query, Query, RoaringBitmap)>>,
) -> anyhow::Result<RoaringBitmap>
{
fn resolve_operation<'t>(
ctx: &'t dyn Context,
query_tree: &Operation,
proximity: u8,
cache: &mut HashMap<(Operation, u8), Vec<(Query, Query, RoaringBitmap)>>,
) -> anyhow::Result<Vec<(Query, Query, RoaringBitmap)>>
{
use Operation::{And, Consecutive, Or, Query};
let result = match query_tree {
And(ops) => mdfs(ctx, ops, proximity, cache)?,
Consecutive(ops) => if proximity == 0 {
mdfs(ctx, ops, 0, cache)?
} else {
Default::default()
},
Or(_, ops) => {
let mut output = Vec::new();
for op in ops {
let result = resolve_operation(ctx, op, proximity, cache)?;
output.extend(result);
}
output
},
Query(q) => if proximity == 0 {
let candidates = query_docids(ctx, q)?;
vec![(q.clone(), q.clone(), candidates)]
} else {
Default::default()
},
};
Ok(result)
}
fn mdfs_pair<'t>(
ctx: &'t dyn Context,
left: &Operation,
right: &Operation,
proximity: u8,
cache: &mut HashMap<(Operation, u8), Vec<(Query, Query, RoaringBitmap)>>,
) -> anyhow::Result<Vec<(Query, Query, RoaringBitmap)>>
{
2021-03-02 14:46:50 +01:00
fn pair_combinations(mana: u8, left_max: u8) -> impl Iterator<Item = (u8, u8)> {
(0..=mana.min(left_max)).map(move |m| (m, mana - m))
2021-02-22 17:17:01 +01:00
}
2021-03-02 14:46:50 +01:00
let pair_max_proximity = 7;
2021-02-22 17:17:01 +01:00
let mut output = Vec::new();
2021-03-02 14:46:50 +01:00
for (pair_p, left_right_p) in pair_combinations(proximity, pair_max_proximity) {
for (left_p, right_p) in pair_combinations(left_right_p, left_right_p) {
2021-02-22 17:17:01 +01:00
let left_key = (left.clone(), left_p);
if !cache.contains_key(&left_key) {
let candidates = resolve_operation(ctx, left, left_p, cache)?;
cache.insert(left_key.clone(), candidates);
}
let right_key = (right.clone(), right_p);
if !cache.contains_key(&right_key) {
let candidates = resolve_operation(ctx, right, right_p, cache)?;
cache.insert(right_key.clone(), candidates);
}
let lefts = cache.get(&left_key).unwrap();
let rights = cache.get(&right_key).unwrap();
for (ll, lr, lcandidates) in lefts {
for (rl, rr, rcandidates) in rights {
let mut candidates = query_pair_proximity_docids(ctx, lr, rl, pair_p + 1)?;
if lcandidates.len() < rcandidates.len() {
candidates.intersect_with(lcandidates);
candidates.intersect_with(rcandidates);
} else {
candidates.intersect_with(rcandidates);
candidates.intersect_with(lcandidates);
}
if !candidates.is_empty() {
output.push((ll.clone(), rr.clone(), candidates));
}
}
}
}
}
Ok(output)
}
fn mdfs<'t>(
ctx: &'t dyn Context,
branches: &[Operation],
proximity: u8,
cache: &mut HashMap<(Operation, u8), Vec<(Query, Query, RoaringBitmap)>>,
) -> anyhow::Result<Vec<(Query, Query, RoaringBitmap)>>
{
// Extract the first two elements but gives the tail
// that is just after the first element.
let next = branches.split_first().map(|(h1, t)| {
(h1, t.split_first().map(|(h2, _)| (h2, t)))
});
match next {
Some((head1, Some((head2, [_])))) => mdfs_pair(ctx, head1, head2, proximity, cache),
Some((head1, Some((head2, tail)))) => {
let mut output = Vec::new();
for p in 0..=proximity {
for (lhead, _, head_candidates) in mdfs_pair(ctx, head1, head2, p, cache)? {
if !head_candidates.is_empty() {
for (_, rtail, mut candidates) in mdfs(ctx, tail, proximity - p, cache)? {
candidates.intersect_with(&head_candidates);
if !candidates.is_empty() {
output.push((lhead.clone(), rtail, candidates));
}
}
}
}
}
Ok(output)
},
Some((head1, None)) => resolve_operation(ctx, head1, proximity, cache),
None => return Ok(Default::default()),
}
}
let mut candidates = RoaringBitmap::new();
for (_, _, cds) in resolve_operation(ctx, query_tree, proximity, cache)? {
candidates.union_with(&cds);
}
Ok(candidates)
}
fn resolve_plane_sweep_candidates<'t>(
ctx: &'t dyn Context,
query_tree: &Operation,
allowed_candidates: &RoaringBitmap,
) -> anyhow::Result<BTreeMap<u8, RoaringBitmap>>
{
/// FIXME may be buggy with query like "new new york"
fn plane_sweep<'t>(
ctx: &'t dyn Context,
operations: &[Operation],
docid: DocumentId,
consecutive: bool,
) -> anyhow::Result<Vec<(Position, u8, Position)>> {
fn compute_groups_proximity(groups: &Vec<(usize, (Position, u8, Position))>, consecutive: bool) -> Option<(Position, u8, Position)> {
// take the inner proximity of the first group as initial
let mut proximity = groups.first()?.1.1;
let left_most_pos = groups.first()?.1.0;
let right_most_pos = groups.last()?.1.2;
for pair in groups.windows(2) {
if let [(i1, (_, _, rpos1)), (i2, (lpos2, prox2, _))] = pair {
// if a pair overlap, meaning that they share at least a word, we return None
if rpos1 >= lpos2 { return None }
// if groups are in the good order (query order) we remove 1 to the proximity
// the proximity is clamped to 7
let pair_proximity = if i1 < i2 {
(*lpos2 - *rpos1 - 1).min(7)
} else {
(*lpos2 - *rpos1).min(7)
};
proximity += pair_proximity as u8 + prox2;
}
}
// if groups should be consecutives, we will only accept groups with a proximity of 0
if !consecutive || proximity == 0 {
Some((left_most_pos, proximity, right_most_pos))
} else { None }
}
let groups_len = operations.len();
let mut groups_positions = Vec::with_capacity(groups_len);
for operation in operations {
let positions = resolve_operation(ctx, operation, docid)?;
groups_positions.push(positions.into_iter());
}
// Pop top elements of each list.
let mut current = Vec::with_capacity(groups_len);
for (i, positions) in groups_positions.iter_mut().enumerate() {
match positions.next() {
Some(p) => current.push((i, p)),
// if a group return None, it means that the document does not contain all the words,
// we return an empty result.
None => return Ok(Vec::new()),
}
}
// Sort k elements by their positions.
current.sort_unstable_by_key(|(_, p)| *p);
// Find leftmost and rightmost group and their positions.
let mut leftmost = *current.first().unwrap();
let mut rightmost = *current.last().unwrap();
let mut output = Vec::new();
loop {
// Find the position p of the next elements of a list of the leftmost group.
// If the list is empty, break the loop.
let p = groups_positions[leftmost.0].next().map(|p| (leftmost.0, p));
// let q be the position q of second group of the interval.
let q = current[1];
let mut leftmost_index = 0;
// If p > r, then the interval [l, r] is minimal and
// we insert it into the heap according to its size.
if p.map_or(true, |p| p.1 > rightmost.1) {
leftmost_index = current[0].0;
if let Some(group) = compute_groups_proximity(&current, consecutive) {
output.push(group);
}
}
// TODO not sure about breaking here or when the p list is found empty.
let p = match p {
Some(p) => p,
None => break,
};
// Remove the leftmost group P in the interval,
// and pop the same group from a list.
current[leftmost_index] = p;
if p.1 > rightmost.1 {
// if [l, r] is minimal, let r = p and l = q.
rightmost = p;
leftmost = q;
} else {
// Ohterwise, let l = min{p,q}.
leftmost = if p.1 < q.1 { p } else { q };
}
// Then update the interval and order of groups_positions in the interval.
current.sort_unstable_by_key(|(_, p)| *p);
}
// Sort the list according to the size and the positions.
output.sort_unstable();
Ok(output)
}
fn resolve_operation<'t>(
ctx: &'t dyn Context,
query_tree: &Operation,
docid: DocumentId,
) -> anyhow::Result<Vec<(Position, u8, Position)>> {
use Operation::{And, Consecutive, Or};
match query_tree {
And(ops) => plane_sweep(ctx, ops, docid, false),
Consecutive(ops) => plane_sweep(ctx, ops, docid, true),
Or(_, ops) => {
let mut result = Vec::new();
for op in ops {
result.extend(resolve_operation(ctx, op, docid)?)
}
result.sort_unstable();
Ok(result)
},
Operation::Query(Query {prefix, kind}) => {
let fst = ctx.words_fst();
let words = match kind {
QueryKind::Exact { word, .. } => {
if *prefix {
word_derivations(word, true, 0, fst)?
} else {
vec![(word.to_string(), 0)]
}
},
QueryKind::Tolerant { typo, word } => {
word_derivations(word, *prefix, *typo, fst)?
}
};
let mut result = Vec::new();
for (word, _) in words {
if let Some(positions) = ctx.docid_word_positions(docid, &word)? {
let iter = positions.iter().map(|p| (p, 0, p));
result.extend(iter);
}
}
result.sort_unstable();
Ok(result)
}
}
}
let mut candidates = BTreeMap::new();
for docid in allowed_candidates {
let positions = resolve_operation(ctx, query_tree, docid)?;
let best_proximity = positions.into_iter().min_by_key(|(_, proximity, _)| *proximity);
let best_proximity = best_proximity.map(|(_, proximity, _)| proximity).unwrap_or(7);
candidates.entry(best_proximity).or_insert_with(RoaringBitmap::new).insert(docid);
}
Ok(candidates)
}