|
| 1 | +import { describe, expect, it, vi } from 'vitest'; |
| 2 | + |
| 3 | +import { resolveJoiner, normalizeToArray, truncateByPunctuation, hardTruncateFromTail, trimBasedOnBatchProbe } from './trimBatchProbe'; |
| 4 | + |
| 5 | +vi.mock('tokenx', () => ({ |
| 6 | + estimateTokenCount: (str: string) => str.split(/\s+/).filter(Boolean).length, |
| 7 | +})); |
| 8 | + |
| 9 | +describe('trimBasedOnBatchProbe', () => { |
| 10 | + it('prefers compact builds to keep more segments', async () => { |
| 11 | + class BuildableChunk { |
| 12 | + constructor(private readonly detail: string, private readonly summary: string) {} |
| 13 | + build(tryCompactIfPossible?: boolean) { |
| 14 | + return tryCompactIfPossible ? this.summary : this.detail; |
| 15 | + } |
| 16 | + } |
| 17 | + |
| 18 | + const chunks = [ |
| 19 | + new BuildableChunk('very old detail segment', 'old summary'), |
| 20 | + new BuildableChunk('middle detail segment', 'mid summary'), |
| 21 | + new BuildableChunk('latest detail segment', 'latest summary'), |
| 22 | + ]; |
| 23 | + |
| 24 | + const result = await trimBasedOnBatchProbe(chunks, 6); |
| 25 | + |
| 26 | + expect(result).toBe('old summary\nmid summary\nlatest summary'); |
| 27 | + }); |
| 28 | + |
| 29 | + it('keeps newest plain string chunks when under limit', async () => { |
| 30 | + const chunks = ['older message', 'newer message']; |
| 31 | + const result = await trimBasedOnBatchProbe(chunks, 10); |
| 32 | + |
| 33 | + expect(result).toBe('older message\nnewer message'); |
| 34 | + }); |
| 35 | + |
| 36 | + it('prefers compact probe when it allows keeping more segments', async () => { |
| 37 | + class BuildableChunk { |
| 38 | + constructor(private readonly detail: string, private readonly summary: string) {} |
| 39 | + build(tryCompactIfPossible?: boolean) { |
| 40 | + return tryCompactIfPossible ? this.summary : this.detail; |
| 41 | + } |
| 42 | + } |
| 43 | + |
| 44 | + // Each detail is 3 tokens, summary is 1 token. |
| 45 | + const chunks = [ |
| 46 | + new BuildableChunk('old detail chunk', 'old'), |
| 47 | + new BuildableChunk('mid detail chunk', 'mid'), |
| 48 | + new BuildableChunk('latest detail chunk', 'latest'), |
| 49 | + ]; |
| 50 | + |
| 51 | + // With detail: 9 tokens total => only last 2 fit; with summary: 3 tokens => all 3 fit. |
| 52 | + const result = await trimBasedOnBatchProbe(chunks, 5); |
| 53 | + |
| 54 | + expect(result).toBe('old\nmid\nlatest'); |
| 55 | + }); |
| 56 | + |
| 57 | + it('truncates only the newest structured segment when all probes fail', async () => { |
| 58 | + const structured = [ |
| 59 | + '<root><a>keep me</a></root>', |
| 60 | + '<root><b>keep me too</b></root>', |
| 61 | + '<root><c>truncate me last</c></root>', |
| 62 | + ]; |
| 63 | + |
| 64 | + const result = await trimBasedOnBatchProbe(structured, 2, { tryChunkingByPunctuation: true }); |
| 65 | + |
| 66 | + // When the limit is extremely small, we still return a truncated newest segment rather than empty. |
| 67 | + expect(result.length).toBeGreaterThan(0); |
| 68 | + }); |
| 69 | + |
| 70 | + it('returns empty when truncation is disabled and nothing fits', async () => { |
| 71 | + const result = await trimBasedOnBatchProbe('too long without truncation', { |
| 72 | + tokenLimit: 1, |
| 73 | + tryChunkingByPunctuation: false, |
| 74 | + tryHardTruncation: false, |
| 75 | + }); |
| 76 | + |
| 77 | + expect(result).toBe(''); |
| 78 | + }); |
| 79 | + |
| 80 | + it('uses compact build for single buildable before truncation', async () => { |
| 81 | + class BuildableChunk { |
| 82 | + constructor(private readonly detail: string, private readonly summary: string) {} |
| 83 | + build(tryCompactIfPossible?: boolean) { |
| 84 | + return tryCompactIfPossible ? this.summary : this.detail; |
| 85 | + } |
| 86 | + } |
| 87 | + |
| 88 | + const result = await trimBasedOnBatchProbe(new BuildableChunk('too long detail text', 'short'), 2); |
| 89 | + |
| 90 | + expect(result).toBe('short'); |
| 91 | + }); |
| 92 | + |
| 93 | + it('selects the largest newest batch within limit via probing', async () => { |
| 94 | + // 5 segments of two tokens each => total 10; limit 5 should keep last 3 segments (6 tokens -> too many), so keep last 2 segments (4 tokens) under limit. |
| 95 | + const segments = ['a b', 'c d', 'e f', 'g h', 'i j']; |
| 96 | + const result = await trimBasedOnBatchProbe(segments, 5); |
| 97 | + |
| 98 | + expect(result).toBe('g h\ni j'); |
| 99 | + }); |
| 100 | + |
| 101 | + it('falls back to punctuation then hard truncation for single strings', async () => { |
| 102 | + const text = |
| 103 | + 'Older sentence should be dropped. Newest sentence should stay intact. trailing tail'; |
| 104 | + |
| 105 | + const result = await trimBasedOnBatchProbe(text, 4); |
| 106 | + |
| 107 | + expect(result).toBe('trailing tail'); |
| 108 | + |
| 109 | + const longToken = 'thisisaverylongtokenwithoutspacesorpunctuationthatkeepsgoing'; |
| 110 | + const hardResult = await trimBasedOnBatchProbe(longToken, 3); |
| 111 | + |
| 112 | + expect(hardResult.length).toBeGreaterThan(0); |
| 113 | + expect(longToken.endsWith(hardResult)).toBe(true); |
| 114 | + }); |
| 115 | + |
| 116 | + describe('__private helpers', () => { |
| 117 | + it('resolves joiners correctly', async () => { |
| 118 | + const fnJoiner = (batch: string[]) => batch.join('|'); |
| 119 | + const objJoiner = { join: (batch: string[]) => batch.join('*') }; |
| 120 | + |
| 121 | + expect(resolveJoiner()).toBe('\n'); |
| 122 | + expect(resolveJoiner(',') as string).toBe(','); |
| 123 | + expect(await (resolveJoiner(fnJoiner) as any)(['a', 'b'])).toBe('a|b'); |
| 124 | + expect(await (resolveJoiner(objJoiner) as any)(['a', 'b'])).toBe('a*b'); |
| 125 | + }); |
| 126 | + |
| 127 | + it('normalizes inputs to array', () => { |
| 128 | + expect(normalizeToArray(null)).toEqual([]); |
| 129 | + expect(normalizeToArray('one')).toEqual(['one']); |
| 130 | + expect(normalizeToArray(['a', 'b'])).toEqual(['a', 'b']); |
| 131 | + }); |
| 132 | + |
| 133 | + it('truncates by punctuation when possible', async () => { |
| 134 | + const text = 'keep this sentence. drop these words'; |
| 135 | + const result = await truncateByPunctuation(text, 3); |
| 136 | + |
| 137 | + expect(result).toBe('drop these words'); |
| 138 | + }); |
| 139 | + |
| 140 | + it('hard truncates from tail with shrinking window', async () => { |
| 141 | + const text = 'averylongtokenwithoutspacesorpunctuation'; |
| 142 | + const result = await hardTruncateFromTail(text, 2); |
| 143 | + |
| 144 | + expect(result.length).toBeGreaterThan(0); |
| 145 | + expect(text.endsWith(result)).toBe(true); |
| 146 | + }); |
| 147 | + }); |
| 148 | +}); |
0 commit comments