[Cherry picked from v1.23.0 - Tentative] Updated tokenizer to better matching when search for code snippets (go-gitea#32261)

This PR improves the accuracy of Gitea's code search.

Currently, Gitea does not consider statements such as
`onsole.log("hello")` as hits when the user searches for `log`. The
culprit is how both ES and Bleve are tokenizing the file contents (in
both cases, `console.log` is a whole token).

In ES' case, we changed the tokenizer to
[simple_pattern_split](https://www.elastic.co/guide/en/elasticsearch/reference/current/analysis-simplepatternsplit-tokenizer.html#:~:text=The%20simple_pattern_split%20tokenizer%20uses%20a,the%20tokenization%20is%20generally%20faster.).
In such a case, tokens are words formed by digits and letters. In
Bleve's case, it employs a
[letter](https://blevesearch.com/docs/Tokenizers/) tokenizer.

Resolves go-gitea#32220

---------

Signed-off-by: Bruno Sofiato <[email protected]>

Author: bsofiato

modules/indexer/code/bleve/bleve.go

@@ -31,6 +31,7 @@ import (
 	"github.com/blevesearch/bleve/v2/analysis/token/camelcase"
 	"github.com/blevesearch/bleve/v2/analysis/token/lowercase"
 	"github.com/blevesearch/bleve/v2/analysis/token/unicodenorm"
+	"github.com/blevesearch/bleve/v2/analysis/tokenizer/letter"
 	"github.com/blevesearch/bleve/v2/analysis/tokenizer/unicode"
 	"github.com/blevesearch/bleve/v2/mapping"
 	"github.com/blevesearch/bleve/v2/search/query"
@@ -69,7 +70,7 @@ const (
 	filenameIndexerAnalyzer  = "filenameIndexerAnalyzer"
 	filenameIndexerTokenizer = "filenameIndexerTokenizer"
 	repoIndexerDocType       = "repoIndexerDocType"
-	repoIndexerLatestVersion = 7
+	repoIndexerLatestVersion = 8
 )
 
 // generateBleveIndexMapping generates a bleve index mapping for the repo indexer
@@ -105,7 +106,7 @@ func generateBleveIndexMapping() (mapping.IndexMapping, error) {
 	} else if err := mapping.AddCustomAnalyzer(repoIndexerAnalyzer, map[string]any{
 		"type":          analyzer_custom.Name,
 		"char_filters":  []string{},
-		"tokenizer":     unicode.Name,
+		"tokenizer":     letter.Name,
 		"token_filters": []string{unicodeNormalizeName, camelcase.Name, lowercase.Name},
 	}); err != nil {
 		return nil, err

modules/indexer/code/elasticsearch/elasticsearch.go

@@ -30,7 +30,7 @@ import (
 )
 
 const (
-	esRepoIndexerLatestVersion = 2
+	esRepoIndexerLatestVersion = 3
 	// multi-match-types, currently only 2 types are used
 	// Reference: https://www.elastic.co/guide/en/elasticsearch/reference/7.0/query-dsl-multi-match-query.html#multi-match-types
 	esMultiMatchTypeBestFields   = "best_fields"
@@ -60,6 +60,10 @@ const (
 		"settings": {
     		"analysis": {
       			"analyzer": {
+					"content_analyzer": {
+						"tokenizer": "content_tokenizer",
+						"filter" : ["lowercase"]
+					},
         			"filename_path_analyzer": {
           				"tokenizer": "path_tokenizer"
         			},
@@ -68,6 +72,10 @@ const (
         			}
       			},
 				"tokenizer": {
+					"content_tokenizer": {
+						"type": "simple_pattern_split",
+						"pattern": "[^a-zA-Z0-9]"
+					},
 					"path_tokenizer": {
 						"type": "path_hierarchy",
 						"delimiter": "/"
@@ -104,7 +112,8 @@ const (
 				"content": {
 					"type": "text",
 					"term_vector": "with_positions_offsets",
-					"index": true
+					"index": true,
+					"analyzer": "content_analyzer"
 				},
 				"commit_id": {
 					"type": "keyword",

modules/indexer/code/indexer_test.go

@@ -181,6 +181,55 @@ func testIndexer(name string, t *testing.T, indexer internal.Indexer) {
 					},
 				},
 			},
+			// Search for matches on the contents of files regardless of case.
+			{
+				RepoIDs: nil,
+				Keyword: "dESCRIPTION",
+				Langs:   1,
+				Results: []codeSearchResult{
+					{
+						Filename: "README.md",
+						Content:  "# repo1\n\nDescription for repo1",
+					},
+				},
+			},
+			// Search for an exact match on the filename within the repo '62' (case insenstive).
+			// This scenario yields a single result (the file avocado.md on the repo '62')
+			{
+				RepoIDs: []int64{62},
+				Keyword: "AVOCADO.MD",
+				Langs:   1,
+				Results: []codeSearchResult{
+					{
+						Filename: "avocado.md",
+						Content:  "# repo1\n\npineaple pie of cucumber juice",
+					},
+				},
+			},
+			// Search for matches on the contents of files when the criteria is a expression.
+			{
+				RepoIDs: []int64{62},
+				Keyword: "console.log",
+				Langs:   1,
+				Results: []codeSearchResult{
+					{
+						Filename: "example-file.js",
+						Content:  "console.log(\"Hello, World!\")",
+					},
+				},
+			},
+			// Search for matches on the contents of files when the criteria is part of a expression.
+			{
+				RepoIDs: []int64{62},
+				Keyword: "log",
+				Langs:   1,
+				Results: []codeSearchResult{
+					{
+						Filename: "example-file.js",
+						Content:  "console.log(\"Hello, World!\")",
+					},
+				},
+			},
 		}
 
 		for _, kw := range keywords {

modules/indexer/internal/bleve/util.go

@@ -6,12 +6,13 @@ package bleve
 import (
 	"errors"
 	"os"
+	"unicode"
 
 	"code.gitea.io/gitea/modules/log"
 	"code.gitea.io/gitea/modules/util"
 
 	"github.com/blevesearch/bleve/v2"
-	"github.com/blevesearch/bleve/v2/analysis/tokenizer/unicode"
+	unicode_tokenizer "github.com/blevesearch/bleve/v2/analysis/tokenizer/unicode"
 	"github.com/blevesearch/bleve/v2/index/upsidedown"
 	"github.com/ethantkoenig/rupture"
 )
@@ -57,7 +58,7 @@ func openIndexer(path string, latestVersion int) (bleve.Index, int, error) {
 // may be different on two string and they still be considered equivalent.
 // Given a phrasse, its shortest word determines its fuzziness. If a phrase uses CJK (eg: `갃갃갃` `啊啊啊`), the fuzziness is zero.
 func GuessFuzzinessByKeyword(s string) int {
-	tokenizer := unicode.NewUnicodeTokenizer()
+	tokenizer := unicode_tokenizer.NewUnicodeTokenizer()
 	tokens := tokenizer.Tokenize([]byte(s))
 
 	if len(tokens) > 0 {
@@ -77,8 +78,10 @@ func guessFuzzinessByKeyword(s string) int {
 	// according to https://github.com/blevesearch/bleve/issues/1563, the supported max fuzziness is 2
 	// magic number 4 was chosen to determine the levenshtein distance per each character of a keyword
 	// BUT, when using CJK (eg: `갃갃갃` `啊啊啊`), it mismatches a lot.
+	// Likewise, queries whose terms contains characters that are *not* letters should not use fuzziness
+
 	for _, r := range s {
-		if r >= 128 {
+		if r >= 128 || !unicode.IsLetter(r) {
 			return 0
 		}
 	}

modules/indexer/internal/bleve/util_test.go

@@ -35,6 +35,14 @@ func TestBleveGuessFuzzinessByKeyword(t *testing.T) {
 			Input:     "갃갃갃",
 			Fuzziness: 0,
 		},
+		{
+			Input:     "repo1",
+			Fuzziness: 0,
+		},
+		{
+			Input:     "avocado.md",
+			Fuzziness: 0,
+		},
 	}
 
 	for _, scenario := range scenarios {

tests/gitea-repositories-meta/org42/search-by-path.git/description

@@ -4,5 +4,6 @@ This repository will be used to test code search. The snippet below shows its di
 ├── avocado.md
 ├── cucumber.md
 ├── ham.md
-└── potato
-    └── ham.md
+├── potato
+|   └── ham.md
+└── example-file.js

tests/gitea-repositories-meta/org42/search-by-path.git/info/refs

@@ -3,7 +3,7 @@
 65f1bf27bc3bf70f64657658635e66094edbcb4d	refs/heads/develop
 65f1bf27bc3bf70f64657658635e66094edbcb4d	refs/heads/feature/1
 78fb907e3a3309eae4fe8fef030874cebbf1cd5e	refs/heads/home-md-img-check
-3731fe53b763859aaf83e703ee731f6b9447ff1e	refs/heads/master
+9f894b61946fd2f7b8b9d8e370e4d62f915522f5	refs/heads/master
 62fb502a7172d4453f0322a2cc85bddffa57f07a	refs/heads/pr-to-update
 4649299398e4d39a5c09eb4f534df6f1e1eb87cc	refs/heads/sub-home-md-img-check
 3fa2f829675543ecfc16b2891aebe8bf0608a8f4	refs/notes/commits

tests/gitea-repositories-meta/org42/search-by-path.git/objects/info/commit-graph

@@ -1,2 +1,2 @@
-P pack-393dc29256bc27cb2ec73898507df710be7a3cf5.pack
+P pack-a7bef76cf6e2b46bc816936ab69306fb10aea571.pack