From bb1ab428dbcf2021dbf1bfd4b889e91ced53608a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Cl=C3=A9ment=20Renault?= <clement@meilisearch.com>
Date: Sun, 6 Sep 2020 17:33:26 +0200
Subject: [PATCH] Use another function to define the proximity

---
 Cargo.lock         |  2 +-
 Cargo.toml         |  2 +-
 src/bin/indexer.rs |  5 +++--
 src/search.rs      | 34 +++++++++++++++++++++++++++++++---
 4 files changed, 36 insertions(+), 7 deletions(-)

diff --git a/Cargo.lock b/Cargo.lock
index b68215ac9..390983638 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1145,7 +1145,7 @@ dependencies = [
 [[package]]
 name = "near-proximity"
 version = "0.1.0"
-source = "git+https://github.com/Kerollmops/plane-sweep-proximity?rev=2f5ad5c#2f5ad5cdafde54731cd75d17ec6228ea3ca1f9b4"
+source = "git+https://github.com/Kerollmops/plane-sweep-proximity?rev=6608205#66082058537f6fe7709adc4690048d62f3c0e9b7"
 dependencies = [
  "tinyvec",
 ]
diff --git a/Cargo.toml b/Cargo.toml
index 4e130c583..39e155d87 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -22,7 +22,7 @@ levenshtein_automata = { version = "0.2.0", features = ["fst_automaton"] }
 memmap = "0.7.0"
 once_cell = "1.4.0"
 oxidized-mtbl = { git = "https://github.com/Kerollmops/oxidized-mtbl.git", rev = "5426182" }
-near-proximity = { git = "https://github.com/Kerollmops/plane-sweep-proximity", rev = "2f5ad5c" }
+near-proximity = { git = "https://github.com/Kerollmops/plane-sweep-proximity", rev = "6608205" }
 rayon = "1.3.1"
 ringtail = "0.3.0"
 roaring = { git = "https://github.com/Kerollmops/roaring-rs.git", branch = "mem-usage" }
diff --git a/src/bin/indexer.rs b/src/bin/indexer.rs
index 1a37874c0..e66d63d7d 100644
--- a/src/bin/indexer.rs
+++ b/src/bin/indexer.rs
@@ -199,13 +199,14 @@ impl Store {
         let mut key = vec![WORD_DOCID_POSITIONS_BYTE];
         let mut buffer = Vec::new();
 
-        // We serialize the document ids into a buffer
         // We prefix the words by the document id.
         key.extend_from_slice(&id.to_be_bytes());
+        let base_size = key.len();
 
         for (word, positions) in iter {
-            key.truncate(1 + 4);
+            key.truncate(base_size);
             key.extend_from_slice(word.as_bytes());
+            // We serialize the positions into a buffer.
             buffer.clear();
             buffer.reserve(positions.serialized_size());
             positions.serialize_into(&mut buffer)?;
diff --git a/src/search.rs b/src/search.rs
index 9758862e7..a545ab77d 100644
--- a/src/search.rs
+++ b/src/search.rs
@@ -1,4 +1,5 @@
 use std::collections::{HashMap, HashSet};
+use std::cmp;
 
 use fst::{IntoStreamer, Streamer};
 use levenshtein_automata::DFA;
@@ -10,7 +11,7 @@ use roaring::bitmap::{IntoIter, RoaringBitmap};
 use near_proximity::near_proximity;
 
 use crate::query_tokens::{QueryTokens, QueryToken};
-use crate::{Index, DocumentId};
+use crate::{Index, DocumentId, Position};
 
 // Building these factories is not free.
 static LEVDIST0: Lazy<LevBuilder> = Lazy::new(|| LevBuilder::new(0, true));
@@ -153,7 +154,7 @@ impl<'a> Search<'a> {
                 if docids.contains(candidate) {
                     match index.docid_word_positions.get(rtxn, &(candidate, word))? {
                         Some(positions) => union_positions.union_with(&positions),
-                        None => error!("position missing for candidate {} and word {}", candidate, word),
+                        None => error!("position missing for candidate {} and word {:?}", candidate, word),
                     }
                 }
             }
@@ -194,10 +195,37 @@ impl<'a> Search<'a> {
         let min_proximity = derived_words.len() as u32 - 1;
         let mut number_min_proximity = 0;
 
+        // TODO move this function elsewhere
+        fn compute_proximity(path: &[Position]) -> u32 {
+            const ONE_ATTRIBUTE: u32 = 1000;
+            const MAX_DISTANCE: u32 = 8;
+
+            fn index_proximity(lhs: u32, rhs: u32) -> u32 {
+                if lhs <= rhs {
+                    cmp::min(rhs - lhs, MAX_DISTANCE)
+                } else {
+                    cmp::min((lhs - rhs) + 1, MAX_DISTANCE)
+                }
+            }
+
+            fn positions_proximity(lhs: u32, rhs: u32) -> u32 {
+                let (lhs_attr, lhs_index) = extract_position(lhs);
+                let (rhs_attr, rhs_index) = extract_position(rhs);
+                if lhs_attr != rhs_attr { MAX_DISTANCE }
+                else { index_proximity(lhs_index, rhs_index) }
+            }
+
+            fn extract_position(position: u32) -> (u32, u32) {
+                (position / ONE_ATTRIBUTE, position % ONE_ATTRIBUTE)
+            }
+
+             path.windows(2).map(|w| positions_proximity(w[0], w[1])).sum::<u32>()
+        }
+
         let mut paths = Vec::new();
         for candidate in candidates {
             let keywords = Self::fecth_keywords(rtxn, index, &derived_words, candidate)?;
-            near_proximity(keywords, &mut paths);
+            near_proximity(keywords, &mut paths, compute_proximity);
             if let Some((prox, _path)) = paths.first() {
                 documents.push((*prox, candidate));
                 if *prox == min_proximity {