2020-05-31 22:09:34 +08:00
|
|
|
use std::collections::HashMap;
|
|
|
|
use std::hash::BuildHasherDefault;
|
2020-05-31 23:48:13 +08:00
|
|
|
use std::time::Instant;
|
2020-05-31 22:09:34 +08:00
|
|
|
|
2020-05-31 23:48:13 +08:00
|
|
|
use cow_utils::CowUtils;
|
|
|
|
use fst::{IntoStreamer, Streamer};
|
2020-05-31 22:09:34 +08:00
|
|
|
use fxhash::FxHasher32;
|
2020-05-31 23:48:13 +08:00
|
|
|
use heed::types::*;
|
|
|
|
use heed::{PolyDatabase, Database};
|
2020-06-01 00:21:24 +08:00
|
|
|
use levenshtein_automata::LevenshteinAutomatonBuilder as LevBuilder;
|
|
|
|
use once_cell::sync::OnceCell;
|
2020-05-31 23:48:13 +08:00
|
|
|
use roaring::RoaringBitmap;
|
2020-05-31 22:09:34 +08:00
|
|
|
use slice_group_by::StrGroupBy;
|
|
|
|
|
2020-06-01 00:21:24 +08:00
|
|
|
static LEVDIST0: OnceCell<LevBuilder> = OnceCell::new();
|
|
|
|
static LEVDIST1: OnceCell<LevBuilder> = OnceCell::new();
|
|
|
|
static LEVDIST2: OnceCell<LevBuilder> = OnceCell::new();
|
|
|
|
|
2020-05-31 22:09:34 +08:00
|
|
|
pub type FastMap4<K, V> = HashMap<K, V, BuildHasherDefault<FxHasher32>>;
|
|
|
|
pub type SmallString32 = smallstr::SmallString<[u8; 32]>;
|
|
|
|
pub type SmallVec32 = smallvec::SmallVec<[u8; 32]>;
|
|
|
|
pub type BEU32 = heed::zerocopy::U32<heed::byteorder::BE>;
|
|
|
|
pub type DocumentId = u32;
|
|
|
|
|
|
|
|
pub fn alphanumeric_tokens(string: &str) -> impl Iterator<Item = &str> {
|
|
|
|
let is_alphanumeric = |s: &&str| s.chars().next().map_or(false, char::is_alphanumeric);
|
|
|
|
string.linear_group_by_key(|c| c.is_alphanumeric()).filter(is_alphanumeric)
|
|
|
|
}
|
2020-05-31 23:48:13 +08:00
|
|
|
|
|
|
|
#[derive(Clone)]
|
|
|
|
pub struct Index {
|
|
|
|
pub main: PolyDatabase,
|
|
|
|
pub postings_ids: Database<Str, ByteSlice>,
|
2020-06-01 00:20:49 +08:00
|
|
|
pub prefix_postings_ids: Database<Str, ByteSlice>,
|
2020-05-31 23:48:13 +08:00
|
|
|
pub documents: Database<OwnedType<BEU32>, ByteSlice>,
|
|
|
|
}
|
|
|
|
|
|
|
|
impl Index {
|
|
|
|
pub fn new(env: &heed::Env) -> heed::Result<Index> {
|
|
|
|
let main = env.create_poly_database(None)?;
|
|
|
|
let postings_ids = env.create_database(Some("postings-ids"))?;
|
2020-06-01 00:20:49 +08:00
|
|
|
let prefix_postings_ids = env.create_database(Some("prefix-postings-ids"))?;
|
2020-05-31 23:48:13 +08:00
|
|
|
let documents = env.create_database(Some("documents"))?;
|
|
|
|
|
2020-06-01 00:21:24 +08:00
|
|
|
Ok(Index { main, postings_ids, prefix_postings_ids, documents })
|
2020-05-31 23:48:13 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
pub fn headers<'t>(&self, rtxn: &'t heed::RoTxn) -> heed::Result<Option<&'t [u8]>> {
|
|
|
|
self.main.get::<_, Str, ByteSlice>(rtxn, "headers")
|
|
|
|
}
|
|
|
|
|
|
|
|
pub fn search(&self, rtxn: &heed::RoTxn, query: &str) -> anyhow::Result<Vec<DocumentId>> {
|
|
|
|
let fst = match self.main.get::<_, Str, ByteSlice>(rtxn, "words-fst")? {
|
|
|
|
Some(bytes) => fst::Set::new(bytes)?,
|
|
|
|
None => return Ok(Vec::new()),
|
|
|
|
};
|
|
|
|
|
|
|
|
// Building these factories is not free.
|
2020-06-01 00:21:24 +08:00
|
|
|
let lev0 = LEVDIST0.get_or_init(|| LevBuilder::new(0, true));
|
|
|
|
let lev1 = LEVDIST1.get_or_init(|| LevBuilder::new(1, true));
|
|
|
|
let lev2 = LEVDIST2.get_or_init(|| LevBuilder::new(2, true));
|
2020-05-31 23:48:13 +08:00
|
|
|
|
|
|
|
let words: Vec<_> = alphanumeric_tokens(query).collect();
|
2020-06-05 00:36:30 +08:00
|
|
|
let ends_with_whitespace = query.chars().last().map_or(false, char::is_whitespace);
|
2020-05-31 23:48:13 +08:00
|
|
|
let number_of_words = words.len();
|
|
|
|
let dfas = words.into_iter().enumerate().map(|(i, word)| {
|
|
|
|
let word = word.cow_to_lowercase();
|
|
|
|
let is_last = i + 1 == number_of_words;
|
2020-06-05 00:36:30 +08:00
|
|
|
let is_prefix = is_last && !ends_with_whitespace;
|
2020-05-31 23:48:13 +08:00
|
|
|
let dfa = match word.len() {
|
2020-06-05 00:36:30 +08:00
|
|
|
0..=4 => if is_prefix { lev0.build_prefix_dfa(&word) } else { lev0.build_dfa(&word) },
|
|
|
|
5..=8 => if is_prefix { lev1.build_prefix_dfa(&word) } else { lev1.build_dfa(&word) },
|
|
|
|
_ => if is_prefix { lev2.build_prefix_dfa(&word) } else { lev2.build_dfa(&word) },
|
2020-05-31 23:48:13 +08:00
|
|
|
};
|
|
|
|
(word, dfa)
|
|
|
|
});
|
|
|
|
|
|
|
|
let mut intersect_result: Option<RoaringBitmap> = None;
|
|
|
|
for (word, dfa) in dfas {
|
|
|
|
let before = Instant::now();
|
2020-06-01 00:20:49 +08:00
|
|
|
|
2020-05-31 23:48:13 +08:00
|
|
|
let mut union_result = RoaringBitmap::default();
|
2020-06-01 00:20:49 +08:00
|
|
|
if word.len() <= 4 {
|
2020-06-01 23:52:43 +08:00
|
|
|
if let Some(ids) = self.prefix_postings_ids.get(rtxn, &word[..word.len().min(5)])? {
|
2020-06-01 00:20:49 +08:00
|
|
|
union_result = RoaringBitmap::deserialize_from(ids)?;
|
|
|
|
}
|
|
|
|
} else {
|
2020-06-01 23:52:43 +08:00
|
|
|
let mut count = 0;
|
2020-06-01 00:20:49 +08:00
|
|
|
let mut stream = fst.search(dfa).into_stream();
|
|
|
|
while let Some(word) = stream.next() {
|
2020-06-01 23:52:43 +08:00
|
|
|
count += 1;
|
2020-06-01 00:20:49 +08:00
|
|
|
let word = std::str::from_utf8(word)?;
|
|
|
|
if let Some(ids) = self.postings_ids.get(rtxn, word)? {
|
|
|
|
let right = RoaringBitmap::deserialize_from(ids)?;
|
|
|
|
union_result.union_with(&right);
|
|
|
|
}
|
2020-05-31 23:48:13 +08:00
|
|
|
}
|
2020-06-01 23:52:43 +08:00
|
|
|
eprint!("with {:?} words ", count);
|
2020-05-31 23:48:13 +08:00
|
|
|
}
|
2020-06-01 00:21:24 +08:00
|
|
|
eprintln!("union for {:?} took {:.02?}", word, before.elapsed());
|
2020-05-31 23:48:13 +08:00
|
|
|
|
|
|
|
intersect_result = match intersect_result.take() {
|
|
|
|
Some(mut left) => {
|
|
|
|
let before = Instant::now();
|
|
|
|
let left_len = left.len();
|
|
|
|
left.intersect_with(&union_result);
|
2020-06-01 23:52:43 +08:00
|
|
|
eprintln!("intersect between {:?} and {:?} gives {:?} took {:.02?}",
|
|
|
|
left_len, union_result.len(), left.len(), before.elapsed());
|
2020-05-31 23:48:13 +08:00
|
|
|
Some(left)
|
|
|
|
},
|
|
|
|
None => Some(union_result),
|
|
|
|
};
|
|
|
|
}
|
|
|
|
|
2020-06-01 23:52:43 +08:00
|
|
|
eprintln!("{} candidates", intersect_result.as_ref().map_or(0, |r| r.len()));
|
|
|
|
|
2020-05-31 23:48:13 +08:00
|
|
|
Ok(intersect_result.unwrap_or_default().iter().take(20).collect())
|
|
|
|
}
|
|
|
|
}
|