| 1 | package chunk |
| 2 | |
| 3 | import ( |
| 4 | "strings" |
| 5 | "testing" |
| 6 | ) |
| 7 | |
| 8 | func TestSplitCarriesTheHeadingTrail(t *testing.T) { |
| 9 | got := Split("# FY2024\n\n## Revenue\n\n" + strings.Repeat("word ", 60) + "\n\n## Costs\n\n" + strings.Repeat("cost ", 60)) |
| 10 | if len(got) != 2 { |
| 11 | t.Fatalf("want 2 chunks, got %d", len(got)) |
| 12 | } |
| 13 | if got[0].Heading() != "FY2024 › Revenue" { |
| 14 | t.Errorf("first trail = %q", got[0].Heading()) |
| 15 | } |
| 16 | if got[1].Heading() != "FY2024 › Costs" { |
| 17 | t.Errorf("second trail = %q", got[1].Heading()) |
| 18 | } |
| 19 | // The trail must be in the indexable text, or a query naming the |
| 20 | // section cannot find a passage that never repeats it. |
| 21 | if !strings.Contains(got[1].Indexable(), "Costs") { |
| 22 | t.Error("heading absent from indexable text") |
| 23 | } |
| 24 | } |
| 25 | |
| 26 | // The rule that pays for itself: a chunk that straddles a heading cites |
| 27 | // wrongly whichever section you label it with. |
| 28 | func TestNoChunkStraddlesAHeading(t *testing.T) { |
| 29 | // Two short sections that would happily fit in one chunk by size. |
| 30 | got := Split("## Alpha\n\n" + strings.Repeat("a ", 50) + "\n\n## Beta\n\n" + strings.Repeat("b ", 50)) |
| 31 | for _, c := range got { |
| 32 | if strings.Contains(c.Text, "a ") && strings.Contains(c.Text, "b ") { |
| 33 | t.Fatalf("chunk %d mixes two sections: %q", c.Ordinal, c.Text[:40]) |
| 34 | } |
| 35 | } |
| 36 | if len(got) != 2 { |
| 37 | t.Fatalf("want 2 chunks, got %d", len(got)) |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | func TestLongTextSplitsNearTarget(t *testing.T) { |
| 42 | // Twelve paragraphs of 100 words each: 1200 words, target 400. |
| 43 | var b strings.Builder |
| 44 | for i := 0; i < 12; i++ { |
| 45 | b.WriteString(strings.Repeat("filler ", 100)) |
| 46 | b.WriteString("\n\n") |
| 47 | } |
| 48 | got := Split(b.String()) |
| 49 | if len(got) < 2 || len(got) > 4 { |
| 50 | t.Fatalf("1200 words at target %d gave %d chunks", TargetWords, len(got)) |
| 51 | } |
| 52 | for _, c := range got { |
| 53 | if n := len(strings.Fields(c.Text)); n > TargetWords*2 { |
| 54 | t.Errorf("chunk %d has %d words, far over target", c.Ordinal, n) |
| 55 | } |
| 56 | } |
| 57 | } |
| 58 | |
| 59 | func TestOrdinalsAreDense(t *testing.T) { |
| 60 | got := Split("## A\n\n" + strings.Repeat("x ", 500) + "\n\n## B\n\nshort tail") |
| 61 | for i, c := range got { |
| 62 | if c.Ordinal != i { |
| 63 | t.Fatalf("chunk at index %d has ordinal %d; a gap reads as a missing citation", i, c.Ordinal) |
| 64 | } |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | func TestEmptyInputYieldsNoChunks(t *testing.T) { |
| 69 | for _, in := range []string{"", " ", "\n\n\n", "# Heading with no body"} { |
| 70 | if got := Split(in); len(got) != 0 { |
| 71 | t.Errorf("Split(%q) gave %d chunks, want 0", in, len(got)) |
| 72 | } |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | // A document that starts at "##" with no "#" above it pads the trail, and |
| 77 | // an unfiltered join renders "› Revenue" — a separator with nothing before |
| 78 | // it. Every earlier fixture started at "#", so nothing caught this until |
| 79 | // the page was looked at. |
| 80 | func TestHeadingDropsEmptyTrailSegments(t *testing.T) { |
| 81 | got := Split("## Revenue\n\n" + strings.Repeat("word ", 60)) |
| 82 | if len(got) != 1 { |
| 83 | t.Fatalf("want 1 chunk, got %d", len(got)) |
| 84 | } |
| 85 | if h := got[0].Heading(); h != "Revenue" { |
| 86 | t.Errorf("Heading() = %q, want %q", h, "Revenue") |
| 87 | } |
| 88 | if strings.HasPrefix(got[0].Heading(), "›") || strings.Contains(got[0].Heading(), " › ") { |
| 89 | t.Errorf("heading has a dangling separator: %q", got[0].Heading()) |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | // Deeper skips too: "###" with nothing above it. |
| 94 | func TestDeepSkipDoesNotDangle(t *testing.T) { |
| 95 | got := Split("### Deep\n\n" + strings.Repeat("word ", 60)) |
| 96 | if len(got) != 1 || got[0].Heading() != "Deep" { |
| 97 | t.Fatalf("Heading() = %q, want %q", got[0].Heading(), "Deep") |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | // The commonest Markdown style: a heading with no blank line after it. |
| 102 | // The paragraph-based chunker filed the whole thing as one chunk of body |
| 103 | // text with an empty trail, so those documents indexed with no heading at |
| 104 | // all — and the seeded corpus is full of them. |
| 105 | func TestHeadingWithNoBlankLineAfterIt(t *testing.T) { |
| 106 | got := Split("## Revenue\nJune revenue was 1,350,000.\n\n## Costs\nHosting fell after the migration.") |
| 107 | if len(got) != 2 { |
| 108 | t.Fatalf("want 2 chunks, got %d: %+v", len(got), got) |
| 109 | } |
| 110 | if got[0].Heading() != "Revenue" || got[1].Heading() != "Costs" { |
| 111 | t.Errorf("headings are %q and %q", got[0].Heading(), got[1].Heading()) |
| 112 | } |
| 113 | if strings.Contains(got[0].Text, "##") { |
| 114 | t.Errorf("the heading line leaked into the body: %q", got[0].Text) |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | // A "#" inside a fenced code block is a shell comment. Treating it as a |
| 119 | // heading filed everything after it under a trail like "now run it". |
| 120 | func TestHashInAFencedBlockIsNotAHeading(t *testing.T) { |
| 121 | got := Split("# Real heading\n\n```sh\nmake build\n\n# now run it\n\n./app\n```\n\nProse after the block.") |
| 122 | for _, c := range got { |
| 123 | if strings.Contains(c.Heading(), "now run it") { |
| 124 | t.Fatalf("a shell comment became a heading: trail %q", c.Heading()) |
| 125 | } |
| 126 | } |
| 127 | if len(got) == 0 || got[0].Heading() != "Real heading" { |
| 128 | t.Fatalf("the real heading was lost: %+v", got) |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | func TestTildeFencesToo(t *testing.T) { |
| 133 | got := Split("# Top\n\n~~~\n# not a heading\n~~~\n\nBody text here.") |
| 134 | for _, c := range got { |
| 135 | if strings.Contains(c.Heading(), "not a heading") { |
| 136 | t.Fatalf("a tilde-fenced comment became a heading: %q", c.Heading()) |
| 137 | } |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | // A document with no blank lines — a CSV, a log, hard-wrapped prose — |
| 142 | // became one chunk of the whole file, whose single citation read |
| 143 | // "passage 0" for everything. The form invites .csv, and a CSV never has |
| 144 | // a blank line. |
| 145 | func TestNoBlankLinesStillSplits(t *testing.T) { |
| 146 | var b strings.Builder |
| 147 | for i := 0; i < 4000; i++ { |
| 148 | b.WriteString("alpha,beta,gamma,delta,epsilon\n") |
| 149 | } |
| 150 | got := Split(b.String()) |
| 151 | if len(got) < 2 { |
| 152 | t.Fatalf("a 4000-line CSV produced %d chunk(s); MaxWords is not binding", len(got)) |
| 153 | } |
| 154 | for _, c := range got { |
| 155 | if n := len(strings.Fields(c.Text)); n > MaxWords*2 { |
| 156 | t.Errorf("chunk %d has %d words, over twice the ceiling", c.Ordinal, n) |
| 157 | } |
| 158 | } |
| 159 | } |
| 160 | |
| 161 | // No text may be lost. Every word of the input must appear in some chunk. |
| 162 | func TestSplitLosesNoText(t *testing.T) { |
| 163 | in := "# Top\n\nFirst para here.\n## Second\nNo blank line above.\n\n```\ncode # here\n```\n\nLast para." |
| 164 | var joined string |
| 165 | for _, c := range Split(in) { |
| 166 | joined += " " + c.Text |
| 167 | } |
| 168 | for _, want := range []string{"First para here.", "No blank line above.", "code # here", "Last para."} { |
| 169 | if !strings.Contains(joined, want) { |
| 170 | t.Errorf("lost %q from the output", want) |
| 171 | } |
| 172 | } |
| 173 | } |
| 174 | |
amadan