Introduce a first draft of the best_proximity algorithm

This commit is contained in:
Kerollmops 2020-06-08 18:05:14 +02:00
parent dfdaceb410
commit 2a6d6a7f69
No known key found for this signature in database
GPG Key ID: 92ADA4E935E71FA4
3 changed files with 218 additions and 22 deletions

195
src/best_proximity.rs Normal file
View File

@ -0,0 +1,195 @@
use std::cmp;
const ONE_ATTRIBUTE: u32 = 1000;
const MAX_INDEX: u32 = ONE_ATTRIBUTE - 1;
const MAX_DISTANCE: u32 = 8;
// Returns the attribute and index parts.
fn extract_position(position: u32) -> (u32, u32) {
(position / ONE_ATTRIBUTE, position % ONE_ATTRIBUTE)
}
// Returns a position from the two parts of it.
fn construct_position(attr: u32, index: u32) -> u32 {
attr * ONE_ATTRIBUTE + index
}
// TODO we should use an sdset::Set for `next_positions`.
// Returns the positions to focus that will give the best possible proximity.
fn best_proximity_for(current_position: u32, proximity: u32, next_positions: &[u32]) -> Option<(u32, Vec<u32>)> {
let (current_attr, _) = extract_position(current_position);
match proximity {
// look at i+0
0 => {
match next_positions.binary_search(&current_position) {
Ok(_) => Some((0, vec![current_position])),
Err(_) => best_proximity_for(current_position, proximity + 1, next_positions),
}
},
// look at i+1
1 => {
let position = current_position + 1;
let (attr, _) = extract_position(position);
// We must check that we do not overflowed the current attribute. If so,
// we must check for a bigger proximity that we will be able to find behind.
if current_attr == attr {
match next_positions.binary_search(&position) {
Ok(_) => Some((1, vec![position])),
Err(_) => best_proximity_for(current_position, proximity + 1, next_positions),
}
} else {
best_proximity_for(current_position, proximity + 1, next_positions)
}
},
// look at i-(p-1), i+p
2..=7 => {
let mut output = Vec::new();
// Behind the current_position
if let Some(position) = current_position.checked_sub(proximity - 1) {
let (attr, _) = extract_position(position);
// We must make sure we are not looking at a word at the end of another attribute.
if current_attr == attr && next_positions.binary_search(&position).is_ok() {
output.push(position);
}
}
// In front of the current_position
let position = current_position + proximity;
let (attr, _) = extract_position(position);
// We must make sure we are not looking at a word at the end of another attribute.
if current_attr == attr && next_positions.binary_search(&position).is_ok() {
output.push(position);
}
if output.is_empty() {
best_proximity_for(current_position, proximity + 1, next_positions)
} else {
Some((proximity, output))
}
},
// look at i+8 and all above and i-(8-1) and all below
8 => {
let mut output = Vec::new();
// Make sure we look at the latest index of the previous attr.
if let Some(previous_position) = construct_position(current_attr, 0).checked_sub(1) {
let position = current_position.saturating_sub(7).max(previous_position);
match dbg!(next_positions.binary_search(&position)) {
Ok(i) => output.extend_from_slice(&next_positions[..=i]),
Err(i) => if let Some(i) = i.checked_sub(1) {
if let Some(positions) = next_positions.get(..=i) {
output.extend_from_slice(positions)
}
},
}
}
// Make sure the position doesn't overflow to the next attribute.
let position = (current_position + 8).min(construct_position(current_attr + 1, 0));
match next_positions.binary_search(&position) {
Ok(i) => output.extend_from_slice(&next_positions[i..]),
Err(i) => if let Some(positions) = next_positions.get(i..) {
output.extend_from_slice(positions);
},
}
if output.is_empty() {
None
} else {
Some((8, output))
}
}
_ => None,
}
}
pub struct BestProximity {
positions: Vec<Vec<u32>>,
current_proximity: Option<(u32, Vec<(u32, usize)>)>, // where we are
}
impl BestProximity {
pub fn new(positions: Vec<Vec<u32>>) -> BestProximity {
BestProximity { positions, current_proximity: None }
}
}
impl Iterator for BestProximity {
type Item = (u32, Vec<u32>);
fn next(&mut self) -> Option<Self::Item> {
let output = Vec::new();
let best_proximity = 0;
for (i, positions) in self.positions.iter().enumerate() {
if let Some(next_positions) = self.positions.get(i + 1) {
for x in positions {
let p = next_positions.binary_search(&x);
let y = next_positions.get(p.unwrap_or_else(|p| p));
eprintln!("{:?} gives {:?} ({:?})", x, p, y);
}
}
}
// match &mut self.current_proximity {
// Some((_prox, _pos)) => {
// // ...
// },
// None => {
// // ...
// },
// }
Some((best_proximity, output))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn same_attribute() {
let positions = vec![
vec![0, 2, 3, 4 ],
vec![ 1, ],
vec![ 3, 6],
];
let mut iter = BestProximity::new(positions);
assert_eq!(iter.next(), Some((1+2, vec![0, 1, 3]))); // 3
assert_eq!(iter.next(), Some((2+2, vec![2, 1, 3]))); // 4
assert_eq!(iter.next(), Some((3+2, vec![3, 1, 3]))); // 5
assert_eq!(iter.next(), Some((1+5, vec![0, 1, 6]))); // 6
assert_eq!(iter.next(), Some((4+2, vec![4, 1, 3]))); // 6
assert_eq!(iter.next(), Some((2+5, vec![2, 1, 6]))); // 7
assert_eq!(iter.next(), Some((3+5, vec![3, 1, 6]))); // 8
assert_eq!(iter.next(), Some((4+5, vec![4, 1, 6]))); // 9
assert_eq!(iter.next(), None);
}
#[test]
fn easy_best_proximity_for() {
// classic
assert_eq!(best_proximity_for(0, 0, &[0]), Some((0, vec![0])));
assert_eq!(best_proximity_for(0, 1, &[0]), None);
assert_eq!(best_proximity_for(1, 1, &[0]), Some((2, vec![0])));
assert_eq!(best_proximity_for(0, 1, &[0, 1]), Some((1, vec![1])));
assert_eq!(best_proximity_for(1, 1, &[0, 2]), Some((1, vec![2])));
assert_eq!(best_proximity_for(1, 2, &[0, 2]), Some((2, vec![0])));
assert_eq!(best_proximity_for(1, 2, &[0, 3]), Some((2, vec![0, 3])));
// limits
assert_eq!(best_proximity_for(2, 7, &[0, 9]), Some((7, vec![9])));
assert_eq!(best_proximity_for(12, 7, &[6, 19]), Some((7, vec![6, 19])));
// another attribute
assert_eq!(best_proximity_for(1000, 7, &[994, 1007]), Some((7, vec![1007])));
assert_eq!(best_proximity_for(1004, 7, &[994, 1011]), Some((7, vec![1011])));
assert_eq!(best_proximity_for(1004, 8, &[900, 913, 1000, 1012, 2012]), Some((8, vec![900, 913, 1012, 2012])));
assert_eq!(best_proximity_for(1009, 8, &[900, 913, 1002, 1012, 2012]), Some((8, vec![900, 913, 1002, 2012])));
}
}

View File

@ -273,17 +273,17 @@ fn index_csv(mut rdr: csv::Reader<File>) -> anyhow::Result<MtblKvStore> {
.or_insert_with(FastMap4::default).entry(position) // positions .or_insert_with(FastMap4::default).entry(position) // positions
.or_insert_with(RoaringBitmap::new).insert(document_id); // document ids .or_insert_with(RoaringBitmap::new).insert(document_id); // document ids
// We save the documents ids under the position and prefix of the word we have seen it. // // We save the documents ids under the position and prefix of the word we have seen it.
if let Some(prefix) = word.as_bytes().get(0..word.len().min(5)) { // if let Some(prefix) = word.as_bytes().get(0..word.len().min(5)) {
for i in 1..=prefix.len() { // for i in 1..=prefix.len() {
prefix_postings_attrs.entry(SmallVec32::from(&prefix[..i])) // prefix_postings_attrs.entry(SmallVec32::from(&prefix[..i]))
.or_insert_with(RoaringBitmap::new).insert(position); // .or_insert_with(RoaringBitmap::new).insert(position);
prefix_postings_ids.entry(SmallVec32::from(&prefix[..i])) // prefix_postings_ids.entry(SmallVec32::from(&prefix[..i]))
.or_insert_with(FastMap4::default).entry(position) // positions // .or_insert_with(FastMap4::default).entry(position) // positions
.or_insert_with(RoaringBitmap::new).insert(document_id); // document ids // .or_insert_with(RoaringBitmap::new).insert(document_id); // document ids
} // }
} // }
} }
} }
} }

View File

@ -1,9 +1,9 @@
mod best_proximity;
mod query_tokens; mod query_tokens;
use std::borrow::Cow; use std::borrow::Cow;
use std::collections::HashMap; use std::collections::HashMap;
use std::hash::BuildHasherDefault; use std::hash::BuildHasherDefault;
use std::time::Instant;
use cow_utils::CowUtils; use cow_utils::CowUtils;
use fst::{IntoStreamer, Streamer}; use fst::{IntoStreamer, Streamer};
@ -15,6 +15,7 @@ use once_cell::sync::Lazy;
use roaring::RoaringBitmap; use roaring::RoaringBitmap;
use self::query_tokens::{QueryTokens, QueryToken}; use self::query_tokens::{QueryTokens, QueryToken};
use self::best_proximity::BestProximity;
// Building these factories is not free. // Building these factories is not free.
static LEVDIST0: Lazy<LevBuilder> = Lazy::new(|| LevBuilder::new(0, true)); static LEVDIST0: Lazy<LevBuilder> = Lazy::new(|| LevBuilder::new(0, true));
@ -88,10 +89,12 @@ impl Index {
}); });
let mut words_positions = Vec::new(); let mut words_positions = Vec::new();
let mut positions = Vec::new();
for (word, is_prefix, dfa) in dfas { for (word, is_prefix, dfa) in dfas {
let mut count = 0; let mut count = 0;
let mut union_positions = RoaringBitmap::default(); let mut union_positions = RoaringBitmap::default();
if word.len() <= 4 && is_prefix { if false && word.len() <= 4 && is_prefix {
if let Some(ids) = self.prefix_postings_attrs.get(rtxn, word.as_bytes())? { if let Some(ids) = self.prefix_postings_attrs.get(rtxn, word.as_bytes())? {
let right = RoaringBitmap::deserialize_from(ids)?; let right = RoaringBitmap::deserialize_from(ids)?;
union_positions.union_with(&right); union_positions.union_with(&right);
@ -110,23 +113,22 @@ impl Index {
} }
eprintln!("{} words for {:?} we have found positions {:?}", count, word, union_positions); eprintln!("{} words for {:?} we have found positions {:?}", count, word, union_positions);
words_positions.push((word, is_prefix, dfa, union_positions)); words_positions.push((word, is_prefix, dfa));
positions.push(union_positions.iter().collect());
} }
use itertools::EitherOrBoth; // let positions = BestProximity::new(positions).next().unwrap_or_default();
let (a, b) = (&words_positions[0].3, &words_positions[1].3); let _positions: Vec<Vec<u32>> = positions;
let positions: Vec<_> = itertools::merge_join_by(a, b, |a, b| (a + 1).cmp(b)).flat_map(EitherOrBoth::both).collect(); let positions = vec![0u32];
eprintln!("best proximity {:?}", positions);
if positions.is_empty() { return Ok(Vec::new()); }
let mut intersect_docids: Option<RoaringBitmap> = None; let mut intersect_docids: Option<RoaringBitmap> = None;
for (i, (word, is_prefix, dfa, _)) in words_positions.into_iter().take(2).enumerate() { for ((word, is_prefix, dfa), pos) in words_positions.into_iter().zip(positions) {
let mut count = 0; let mut count = 0;
let mut union_docids = RoaringBitmap::default(); let mut union_docids = RoaringBitmap::default();
if word.len() <= 4 && is_prefix { if false && word.len() <= 4 && is_prefix {
let mut key = word.as_bytes()[..word.len().min(5)].to_vec(); let mut key = word.as_bytes()[..word.len().min(5)].to_vec();
let pos = if i == 0 { positions[0].0 } else { positions[0].1 };
key.extend_from_slice(&pos.to_be_bytes()); key.extend_from_slice(&pos.to_be_bytes());
if let Some(ids) = self.prefix_postings_ids.get(rtxn, &key)? { if let Some(ids) = self.prefix_postings_ids.get(rtxn, &key)? {
let right = RoaringBitmap::deserialize_from(ids)?; let right = RoaringBitmap::deserialize_from(ids)?;
@ -138,7 +140,6 @@ impl Index {
while let Some(word) = stream.next() { while let Some(word) = stream.next() {
let word = std::str::from_utf8(word)?; let word = std::str::from_utf8(word)?;
let mut key = word.as_bytes().to_vec(); let mut key = word.as_bytes().to_vec();
let pos = if i == 0 { positions[0].0 } else { positions[0].1 };
key.extend_from_slice(&pos.to_be_bytes()); key.extend_from_slice(&pos.to_be_bytes());
if let Some(attrs) = self.postings_ids.get(rtxn, &key)? { if let Some(attrs) = self.postings_ids.get(rtxn, &key)? {
let right = RoaringBitmap::deserialize_from(attrs)?; let right = RoaringBitmap::deserialize_from(attrs)?;