fork of https://github.com/sourcegraph/zoekt
Ranking: increase contribution of repo rank (#546)
The file score includes a "repo rank" signal, which is based on the
repository's number of stars. Previously, we were aggressively normalizing the
number of stars, which made the repo ranks small and close together. This PR
changes the normalization to spread it out better over the full range. This
increases its contribution to the score.
author
Julie Tibshirani
committer
GitHub
date
(Feb 26, 2023, 4:13 PM -0800)
commit
5f25b307
5f25b3073480520aae1cd145d9f3f57226ff7fbc
parent
25607737
2560773778893ff9f43ebb403ca97f85ffdae8d3
+104
-13
+3
-3
api.go
+3
-3
api.go
···
19
19
"encoding/json"
20
20
"errors"
21
21
"fmt"
22
-
"math"
23
22
"reflect"
24
23
"strconv"
25
24
"time"
···
618
617
// based on priority. Setting it on read instead of during indexing
619
618
// allows us to avoid a complete reindex.
620
619
if r.Rank == 0 && r.priority > 0 {
621
-
l := math.Log(float64(r.priority))
622
-
repo.Rank = uint16((1.0 - 1.0/math.Pow(1+l, 0.6)) * 10000)
620
+
// Normalize the repo score within [0, 1), with the midpoint at 5,000. This means popular
621
+
// repos (roughly ones with over 5,000 stars) see diminishing returns from more stars.
622
+
r.Rank = uint16(r.priority / (5000.0 + r.priority) * maxUInt16)
623
623
}
624
624
}
625
625
return nil
+91
build/e2e_test.go
+91
build/e2e_test.go
···
1241
1241
})
1242
1242
}
1243
1243
}
1244
+
1245
+
func TestRepoRanks(t *testing.T) {
1246
+
if os.Getenv("CI") == "" && checkCTags() == "" {
1247
+
t.Skip("ctags not available")
1248
+
}
1249
+
dir := t.TempDir()
1250
+
1251
+
opts := Options{
1252
+
IndexDir: dir,
1253
+
RepositoryDescription: zoekt.Repository{
1254
+
Name: "repo",
1255
+
},
1256
+
DocumentRanksVersion: "ranking",
1257
+
}
1258
+
1259
+
searchQuery := &query.Substring{Content: true, Pattern: "Inner"}
1260
+
exampleJava, err := os.ReadFile("./testdata/example.java")
1261
+
if err != nil {
1262
+
t.Fatal(err)
1263
+
}
1264
+
1265
+
cases := []struct {
1266
+
name string
1267
+
repoRank uint16
1268
+
wantScore float64
1269
+
}{
1270
+
{
1271
+
name: "no shard rank",
1272
+
// 5500 (partial symbol at boundary) + 1000 (Java class) + 500 (word match) + 10 (file order)
1273
+
wantScore: 7012.00,
1274
+
},
1275
+
{
1276
+
name: "medium shard rank",
1277
+
repoRank: 30000,
1278
+
// 5500 (partial symbol at boundary) + 1000 (Java class) + 500 (word match) + 10 (file order) + 9.16 (repo rank)
1279
+
wantScore: 7021.16,
1280
+
},
1281
+
{
1282
+
name: "high shard rank",
1283
+
repoRank: 60000,
1284
+
// 5500 (partial symbol at boundary) + 1000 (Java class) + 500 (word match) + 10 (file order) + 18.31 (repo rank)
1285
+
wantScore: 7030.31,
1286
+
},
1287
+
}
1288
+
1289
+
for _, c := range cases {
1290
+
t.Run(c.name, func(t *testing.T) {
1291
+
opts.RepositoryDescription = zoekt.Repository{
1292
+
Name: "repo",
1293
+
Rank: c.repoRank,
1294
+
}
1295
+
1296
+
b, err := NewBuilder(opts)
1297
+
if err != nil {
1298
+
t.Fatalf("NewBuilder: %v", err)
1299
+
}
1300
+
1301
+
err = b.Add(zoekt.Document{Name: "example.java", Content: exampleJava})
1302
+
if err != nil {
1303
+
t.Fatal(err)
1304
+
}
1305
+
1306
+
if err := b.Finish(); err != nil {
1307
+
t.Fatalf("Finish: %v", err)
1308
+
}
1309
+
1310
+
ss, err := shards.NewDirectorySearcher(dir)
1311
+
if err != nil {
1312
+
t.Fatalf("NewDirectorySearcher(%s): %v", dir, err)
1313
+
}
1314
+
defer ss.Close()
1315
+
1316
+
srs, err := ss.Search(context.Background(), searchQuery, &zoekt.SearchOptions{
1317
+
UseDocumentRanks: true,
1318
+
DebugScore: true,
1319
+
})
1320
+
1321
+
if err != nil {
1322
+
t.Fatal(err)
1323
+
}
1324
+
1325
+
if got, want := len(srs.Files), 1; got != want {
1326
+
t.Fatalf("file matches: want %d, got %d", want, got)
1327
+
}
1328
+
1329
+
if got := srs.Files[0].Score; math.Abs(got-c.wantScore) >= 0.01 {
1330
+
t.Fatalf("score: want %f, got %f\ndebug: %s\ndebugscore: %s", c.wantScore, got, srs.Files[0].Debug, srs.Files[0].LineMatches[0].DebugScore)
1331
+
}
1332
+
})
1333
+
}
1334
+
}
+1
-1
contentprovider.go
+1
-1
contentprovider.go
···
468
468
// equal weight with the query-dependent signals.
469
469
scoreFileRankFactor = 9000.0
470
470
scoreFileOrderFactor = 10.0
471
-
scoreShardRankFactor = 20.0
471
+
scoreRepoRankFactor = 20.0
472
472
473
473
// Used for ordering line and chunk matches within a file.
474
474
scoreLineOrderFactor = 1.0
+1
-1
eval.go
+1
-1
eval.go
···
385
385
}
386
386
387
387
fileMatch.addScore("doc-order", scoreFileOrderFactor*(1.0-float64(nextDoc)/float64(len(d.boundaries))), opts.DebugScore)
388
-
fileMatch.addScore("shard-order", scoreShardRankFactor*float64(md.Rank)/maxUInt16, opts.DebugScore)
388
+
fileMatch.addScore("repo-rank", scoreRepoRankFactor*float64(md.Rank)/maxUInt16, opts.DebugScore)
389
389
390
390
fileMatch.Branches = d.gatherBranches(nextDoc, mt, known)
391
391
sortMatchesByScore(fileMatch.LineMatches)
+8
-8
shards/shards_test.go
+8
-8
shards/shards_test.go
···
202
202
nextShardNum++
203
203
}
204
204
205
-
addShard("weekend-project", 0.25, zoekt.Document{Name: "f2", Content: []byte("foo bas")})
206
-
addShard("moderately-popular", 0.5, zoekt.Document{Name: "f3", Content: []byte("foo bar")})
207
-
addShard("weekend-project-2", 0.25, zoekt.Document{Name: "f2", Content: []byte("foo bas")})
208
-
addShard("super-star", 0.9, zoekt.Document{Name: "f1", Content: []byte("foo bar bas")})
205
+
addShard("weekend-project", 20, zoekt.Document{Name: "f2", Content: []byte("foo bas")})
206
+
addShard("moderately-popular", 500, zoekt.Document{Name: "f3", Content: []byte("foo bar")})
207
+
addShard("weekend-project-2", 20, zoekt.Document{Name: "f2", Content: []byte("foo bas")})
208
+
addShard("super-star", 5000, zoekt.Document{Name: "f1", Content: []byte("foo bar bas")})
209
209
210
210
want := []string{
211
211
"super-star",
···
244
244
nextShardNum++
245
245
}
246
246
247
-
addShard("weekend-project", 0.25, zoekt.Document{Name: "f1", Content: []byte("foobar")})
248
-
addShard("moderately-popular", 0.4, zoekt.Document{Name: "f2", Content: []byte("foobaz")})
249
-
addShard("weekend-project-2", 0.25, zoekt.Document{Name: "f3", Content: []byte("foo bar")})
250
-
addShard("super-star", 0.9, zoekt.Document{Name: "f4", Content: []byte("foo baz")},
247
+
addShard("weekend-project", 20, zoekt.Document{Name: "f1", Content: []byte("foobar")})
248
+
addShard("moderately-popular", 500, zoekt.Document{Name: "f2", Content: []byte("foobaz")})
249
+
addShard("weekend-project-2", 20, zoekt.Document{Name: "f3", Content: []byte("foo bar")})
250
+
addShard("super-star", 5000, zoekt.Document{Name: "f4", Content: []byte("foo baz")},
251
251
zoekt.Document{Name: "f5", Content: []byte("fooooo")})
252
252
253
253
// Run a stream search and gather the results