fork of https://github.com/sourcegraph/zoekt
1// Copyright 2016 Google Inc. All rights reserved.
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7// http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15package index
16
17import (
18 "bytes"
19 "fmt"
20 "math"
21 "strings"
22
23 "github.com/sourcegraph/zoekt"
24 "github.com/sourcegraph/zoekt/internal/ctags"
25)
26
27const (
28 ScoreOffset = 10_000_000
29)
30
31type chunkScore struct {
32 score float64
33 debugScore string
34 bestLine int
35}
36
37// scoreChunk calculates the score for each line in the chunk based on its candidate matches, and returns the score of
38// the best-scoring line, along with its line number.
39// Invariant: there should be at least one input candidate, len(ms) > 0.
40func (p *contentProvider) scoreChunk(ms []*candidateMatch, language string, opts *zoekt.SearchOptions) (chunkScore, []*zoekt.Symbol) {
41 nl := p.newlines()
42
43 var bestScore lineScore
44 bestLine := 0
45 var symbolInfo []*zoekt.Symbol
46
47 start := 0
48 currentLine := -1
49 for i, m := range ms {
50 lineNumber := -1
51 if !m.fileName {
52 lineNumber = nl.atOffset(m.byteOffset)
53 }
54
55 // If this match represents a new line, then score the previous line and update 'start'.
56 if i != 0 && lineNumber != currentLine {
57 score, si := p.scoreLine(ms[start:i], language, currentLine, opts)
58 symbolInfo = append(symbolInfo, si...)
59 if score.score > bestScore.score {
60 bestScore = score
61 bestLine = currentLine
62 }
63 start = i
64 }
65 currentLine = lineNumber
66 }
67
68 // Make sure to score the last line
69 line, si := p.scoreLine(ms[start:], language, currentLine, opts)
70 symbolInfo = append(symbolInfo, si...)
71 if line.score > bestScore.score {
72 bestScore = line
73 bestLine = currentLine
74 }
75
76 cs := chunkScore{
77 score: bestScore.score,
78 bestLine: bestLine,
79 }
80 if opts.DebugScore {
81 cs.debugScore = fmt.Sprintf("%s, (line: %d)", bestScore.debugScore, bestLine)
82 }
83 return cs, symbolInfo
84}
85
86type lineScore struct {
87 score float64
88 debugScore string
89}
90
91// scoreLine calculates a score for the line based on its candidate matches.
92// Invariants:
93// - All candidate matches are assumed to come from the same line in the content.
94// - If this line represents a filename, then lineNumber must be -1.
95// - There should be at least one input candidate, len(ms) > 0.
96func (p *contentProvider) scoreLine(ms []*candidateMatch, language string, lineNumber int, opts *zoekt.SearchOptions) (lineScore, []*zoekt.Symbol) {
97 if opts.UseBM25Scoring {
98 score, symbolInfo := p.scoreLineBM25(ms, lineNumber)
99 ls := lineScore{score: score}
100 if opts.DebugScore {
101 ls.debugScore = fmt.Sprintf("tfScore:%.2f, ", score)
102 }
103 return ls, symbolInfo
104 }
105
106 score := 0.0
107 what := ""
108 addScore := func(w string, s float64) {
109 if s != 0 && opts.DebugScore {
110 what += fmt.Sprintf("%s:%.2f, ", w, s)
111 }
112 score += s
113 }
114
115 filename := p.data(true)
116 var symbolInfo []*zoekt.Symbol
117
118 var bestScore lineScore
119 for i, m := range ms {
120 data := p.data(m.fileName)
121
122 endOffset := m.byteOffset + m.byteMatchSz
123 startBoundary := m.byteOffset < uint32(len(data)) && (m.byteOffset == 0 || byteClass(data[m.byteOffset-1]) != byteClass(data[m.byteOffset]))
124 endBoundary := endOffset > 0 && (endOffset == uint32(len(data)) || byteClass(data[endOffset-1]) != byteClass(data[endOffset]))
125
126 score = 0
127 what = ""
128
129 if startBoundary && endBoundary {
130 addScore("WordMatch", scoreWordMatch)
131 } else if startBoundary || endBoundary {
132 addScore("PartialWordMatch", scorePartialWordMatch)
133 }
134
135 if m.fileName {
136 sep := bytes.LastIndexByte(data, '/')
137 startMatch := int(m.byteOffset) == sep+1
138 endMatch := endOffset == uint32(len(data))
139 if startMatch && endMatch {
140 addScore("Base", scoreBase)
141 } else if startMatch || endMatch {
142 addScore("EdgeBase", (scoreBase+scorePartialBase)/2)
143 } else if sep < int(m.byteOffset) {
144 addScore("InnerBase", scorePartialBase)
145 }
146 } else if sec, si, ok := p.findSymbol(m); ok {
147 startMatch := sec.Start == m.byteOffset
148 endMatch := sec.End == endOffset
149 if startMatch && endMatch {
150 addScore("Symbol", scoreSymbol)
151 } else if startMatch || endMatch {
152 addScore("EdgeSymbol", (scoreSymbol+scorePartialSymbol)/2)
153 } else {
154 addScore("OverlapSymbol", scorePartialSymbol)
155 }
156
157 // Score based on symbol data
158 if si != nil {
159 symbolKind := ctags.ParseSymbolKind(si.Kind)
160 sym := sectionSlice(data, sec)
161
162 addScore(fmt.Sprintf("kind:%s:%s", language, si.Kind), scoreSymbolKind(language, filename, sym, symbolKind))
163
164 // This is from a symbol tree, so we need to store the symbol
165 // information.
166 if m.symbol {
167 if symbolInfo == nil {
168 symbolInfo = make([]*zoekt.Symbol, len(ms))
169 }
170 // findSymbols does not hydrate in Sym. So we need to store it.
171 si.Sym = string(sym)
172 symbolInfo[i] = si
173 }
174 }
175 }
176
177 // scoreWeight != 1 means it affects score
178 if !epsilonEqualsOne(m.scoreWeight) {
179 score = score * m.scoreWeight
180 if opts.DebugScore {
181 what += fmt.Sprintf("boost:%.2f, ", m.scoreWeight)
182 }
183 }
184
185 if score > bestScore.score {
186 bestScore.score = score
187 bestScore.debugScore = what
188 }
189 }
190
191 if opts.DebugScore {
192 bestScore.debugScore = fmt.Sprintf("score:%.2f <- %s", bestScore.score, strings.TrimSuffix(bestScore.debugScore, ", "))
193 }
194
195 return bestScore, symbolInfo
196}
197
198// scoreLineBM25 computes the score of a line according to BM25, the most common scoring algorithm for text search:
199// https://en.wikipedia.org/wiki/Okapi_BM25. Compared to the standard scoreLine algorithm, this score rewards multiple
200// term matches on a line.
201// Notes:
202// - This BM25 calculation skips inverse document frequency (idf) to keep the implementation simple.
203// - It uses the same calculateTermFrequency method as BM25 file scoring, which boosts filename and symbol matches.
204func (p *contentProvider) scoreLineBM25(ms []*candidateMatch, lineNumber int) (float64, []*zoekt.Symbol) {
205 // If this is a filename, then don't compute BM25. The score would not be comparable to line scores.
206 if lineNumber < 0 {
207 return 0, nil
208 }
209
210 // Use standard parameter defaults used in Lucene (https://lucene.apache.org/core/10_1_0/core/org/apache/lucene/search/similarities/BM25Similarity.html)
211 k, b := 1.2, 0.75
212
213 // Calculate the length ratio of this line. As a heuristic, we assume an average line length of 100 characters.
214 // Usually the calculation would be based on terms, but using bytes should work fine, as we're just computing a ratio.
215 nl := p.newlines()
216 lineLength := nl.lineStart(lineNumber+1) - nl.lineStart(lineNumber)
217 L := float64(lineLength) / 100.0
218
219 score := 0.0
220 tfs := p.calculateTermFrequency(ms)
221 for _, f := range tfs {
222 score += tfScore(k, b, L, f)
223 }
224
225 // Check if any index comes from a symbol match tree, and if so hydrate in symbol information
226 var symbolInfo []*zoekt.Symbol
227 for _, m := range ms {
228 if m.symbol {
229 if sec, si, ok := p.findSymbol(m); ok && si != nil {
230 // findSymbols does not hydrate in Sym. So we need to store it.
231 sym := sectionSlice(p.data(false), sec)
232 si.Sym = string(sym)
233 symbolInfo = append(symbolInfo, si)
234 }
235 }
236 }
237 return score, symbolInfo
238}
239
240// tfScore is the term frequency score for BM25.
241func tfScore(k float64, b float64, L float64, f int) float64 {
242 return ((k + 1.0) * float64(f)) / (k*(1.0-b+b*L) + float64(f))
243}
244
245// calculateTermFrequency computes the term frequency for the file match.
246// Notes:
247// - Filename matches count more than content matches. This mimics a common text search strategy to 'boost' matches on document titles.
248// - Symbol matches also count more than content matches, to reward matches on symbol definitions.
249func (p *contentProvider) calculateTermFrequency(cands []*candidateMatch) map[string]int {
250 // Treat each candidate match as a term and compute the frequencies. For now, ignore case sensitivity and
251 // ignore whether the index is a word boundary.
252 termFreqs := map[string]int{}
253 for _, m := range cands {
254 term := string(m.substrLowered)
255 if m.fileName || p.matchesSymbol(m) {
256 termFreqs[term] += 5
257 } else {
258 termFreqs[term]++
259 }
260 }
261
262 return termFreqs
263}
264
265// scoreFile computes a score for the file match using various scoring signals, like
266// whether there's an exact match on a symbol, the number of query clauses that matched, etc.
267func (d *indexData) scoreFile(fileMatch *zoekt.FileMatch, doc uint32, mt matchTree, known map[matchTree]bool, opts *zoekt.SearchOptions) {
268 atomMatchCount := 0
269 visitMatchAtoms(mt, known, func(mt matchTree) {
270 atomMatchCount++
271 })
272
273 addScore := func(what string, computed float64) {
274 fileMatch.AddScore(what, computed, -1, opts.DebugScore)
275 }
276
277 // atom-count boosts files with matches from more than 1 atom. The
278 // maximum boost is scoreFactorAtomMatch.
279 if atomMatchCount > 0 {
280 fileMatch.AddScore("atom", (1.0-1.0/float64(atomMatchCount))*scoreFactorAtomMatch, float64(atomMatchCount), opts.DebugScore)
281 }
282
283 maxFileScore := 0.0
284 for i := range fileMatch.LineMatches {
285 if maxFileScore < fileMatch.LineMatches[i].Score {
286 maxFileScore = fileMatch.LineMatches[i].Score
287 }
288
289 // Order by ordering in file.
290 fileMatch.LineMatches[i].Score += scoreLineOrderFactor * (1.0 - (float64(i) / float64(len(fileMatch.LineMatches))))
291 }
292
293 for i := range fileMatch.ChunkMatches {
294 if maxFileScore < fileMatch.ChunkMatches[i].Score {
295 maxFileScore = fileMatch.ChunkMatches[i].Score
296 }
297
298 // Order by ordering in file.
299 fileMatch.ChunkMatches[i].Score += scoreLineOrderFactor * (1.0 - (float64(i) / float64(len(fileMatch.ChunkMatches))))
300 }
301
302 // Maintain ordering of input files. This
303 // strictly dominates the in-file ordering of
304 // the matches.
305 addScore("fragment", maxFileScore)
306
307 // Add tiebreakers
308 //
309 // ScoreOffset shifts the score 7 digits to the left.
310 fileMatch.Score = math.Trunc(fileMatch.Score) * ScoreOffset
311
312 md := d.repoMetaData[d.repos[doc]]
313
314 // md.Rank lies in the range [0, 65535]. Hence, we have to allocate 5 digits for
315 // the rank. The scoreRepoRankFactor shifts the rank score 2 digits to the left,
316 // reserving digits 3-7 for the repo rank.
317 addScore("repo-rank", scoreRepoRankFactor*float64(md.Rank))
318
319 // digits 1-2 and the decimals are reserved for the doc order. Doc order
320 // (without the scaling factor) lies in the range [0, 1]. The upper bound is
321 // achieved for matches in the first document of a shard.
322 addScore("doc-order", scoreFileOrderFactor*(1.0-float64(doc)/float64(len(d.boundaries))))
323
324 if opts.DebugScore {
325 // To make the debug output easier to read, we split the score into the query
326 // dependent score and the tiebreaker
327 score := math.Trunc(fileMatch.Score / ScoreOffset)
328 tiebreaker := fileMatch.Score - score*ScoreOffset
329 fileMatch.Debug = fmt.Sprintf("score: %d (%.2f) <- %s", int(score), tiebreaker, strings.TrimSuffix(fileMatch.Debug, ", "))
330 }
331}
332
333// scoreFilesUsingBM25 computes the score according to BM25, the most common scoring algorithm for text search:
334// https://en.wikipedia.org/wiki/Okapi_BM25. Note that we treat the inverse document frequency (idf) as constant. This
335// is supported by our evaluations which showed that for keyword style queries, idf can down-weight the score of some
336// keywords too much, leading to a worse ranking. The intuition is that each keyword is important independently of how
337// frequent it appears in the corpus.
338//
339// Unlike standard file scoring, this scoring strategy ignores all other signals including document ranks. This keeps
340// things simple for now, since BM25 is not normalized and can be tricky to combine with other scoring signals. It also
341// ignores the individual LineMatch and ChunkMatch scores, instead calculating a score over all matches in the file.
342func (d *indexData) scoreFilesUsingBM25(fileMatch *zoekt.FileMatch, doc uint32, tf map[string]int, opts *zoekt.SearchOptions) {
343 // Use standard parameter defaults used in Lucene (https://lucene.apache.org/core/10_1_0/core/org/apache/lucene/search/similarities/BM25Similarity.html)
344 k, b := 1.2, 0.75
345
346 averageFileLength := float64(d.boundaries[d.numDocs()]) / float64(d.numDocs())
347 // This is very unlikely, but explicitly guard against division by zero.
348 if averageFileLength == 0 {
349 averageFileLength++
350 }
351
352 // Compute the file length ratio. Usually the calculation would be based on terms, but using
353 // bytes should work fine, as we're just computing a ratio.
354 fileLength := float64(d.boundaries[doc+1] - d.boundaries[doc])
355
356 L := fileLength / averageFileLength
357
358 score := 0.0
359 sumTF := 0 // Just for debugging
360 for _, f := range tf {
361 sumTF += f
362 score += tfScore(k, b, L, f)
363 }
364
365 fileMatch.Score = score
366
367 if opts.DebugScore {
368 fileMatch.Debug = fmt.Sprintf("bm25-score: %.2f <- sum-termFrequencies: %d, length-ratio: %.2f", score, sumTF, L)
369 }
370}