|
| 1 | +import {MarkdownTextSplitter} from '@langchain/textsplitters' |
| 2 | + |
| 3 | +import type {Chunk} from '../types.js' |
| 4 | + |
| 5 | +/** |
| 6 | + * Hybrid two-pass markdown splitter for translation purposes. |
| 7 | + * |
| 8 | + * Pass 1: Structural split by markdown elements (frontmatter, code blocks, headers/page directives) |
| 9 | + * Pass 2: Size-based recursive split using LangChain for chunks exceeding chunkSize |
| 10 | + * |
| 11 | + * This approach: |
| 12 | + * - Respects markdown structure and translation requirements (code blocks, frontmatter = non-translatable) |
| 13 | + * - Prevents massive chunks from overwhelming translation APIs |
| 14 | + * - Uses LangChain's intelligent splitting for size management (tries headers > paragraphs > lines) |
| 15 | + */ |
| 16 | +export class MarkdownSplitter { |
| 17 | + private readonly chunkSize: number |
| 18 | + private recursiveSplitter: MarkdownTextSplitter |
| 19 | + |
| 20 | + constructor(chunkSize = 12_000) { |
| 21 | + this.chunkSize = chunkSize |
| 22 | + |
| 23 | + // Initialize LangChain's MarkdownTextSplitter for Pass 2 |
| 24 | + this.recursiveSplitter = new MarkdownTextSplitter({ |
| 25 | + chunkOverlap: 0, // No overlap needed for translation (avoids duplicates) |
| 26 | + chunkSize, |
| 27 | + keepSeparator: true, // Preserve markdown formatting (headers, etc.) |
| 28 | + }) |
| 29 | + } |
| 30 | + |
| 31 | + /** |
| 32 | + * Default split method - automatically detects and uses the appropriate splitting strategy. |
| 33 | + * If the content contains ::page directives, splits by those. |
| 34 | + * Otherwise, splits by headers (# ## ### etc.) |
| 35 | + */ |
| 36 | + async split(markdown: string): Promise<Chunk[]> { |
| 37 | + if (/^::page/m.test(markdown)) { |
| 38 | + return this.splitByPageDirective(markdown) |
| 39 | + } |
| 40 | + |
| 41 | + return this.splitByHeaders(markdown) |
| 42 | + } |
| 43 | + |
| 44 | + /** |
| 45 | + * Splits markdown by headers (# ## ### etc.) |
| 46 | + */ |
| 47 | + async splitByHeaders(markdown: string): Promise<Chunk[]> { |
| 48 | + const firstPassChunks = this.splitInternal(markdown, /^#+\s+/m) |
| 49 | + |
| 50 | + return this.secondPassSplit(firstPassChunks) |
| 51 | + } |
| 52 | + |
| 53 | + /** |
| 54 | + * Splits markdown by custom ::page directives |
| 55 | + */ |
| 56 | + async splitByPageDirective(markdown: string): Promise<Chunk[]> { |
| 57 | + const firstPassChunks = this.splitInternal(markdown, /^::page/m) |
| 58 | + |
| 59 | + return this.secondPassSplit(firstPassChunks) |
| 60 | + } |
| 61 | + |
| 62 | + /** |
| 63 | + * Finds the index of the closing --- for frontmatter |
| 64 | + */ |
| 65 | + private findFrontmatterEnd(lines: string[]): number { |
| 66 | + return lines.slice(1).findIndex((line) => line.trim() === '---') + 1 |
| 67 | + } |
| 68 | + |
| 69 | + /** |
| 70 | + * Checks if the markdown starts with YAML frontmatter (--- ... ---) |
| 71 | + */ |
| 72 | + private hasFrontmatter(lines: string[]): boolean { |
| 73 | + if (lines.length === 0) return false |
| 74 | + |
| 75 | + return lines[0].trim() === '---' && lines.slice(1).some((line) => line.trim() === '---') |
| 76 | + } |
| 77 | + |
| 78 | + /** |
| 79 | + * Pass 2: Size-based recursive split using LangChain |
| 80 | + * |
| 81 | + * For chunks exceeding chunkSize, uses MarkdownTextSplitter to break them down |
| 82 | + * intelligently while preserving the shouldTranslate flag. |
| 83 | + * |
| 84 | + * LangChain's MarkdownTextSplitter tries separators in this order: |
| 85 | + * 1. H2-H6 headers (## ### #### etc.) |
| 86 | + * 2. Code blocks with spacing |
| 87 | + * 3. Horizontal rules (---, ***, ___) |
| 88 | + * 4. Paragraph breaks (\n\n) |
| 89 | + * 5. Line breaks (\n) |
| 90 | + * 6. Spaces |
| 91 | + * 7. Characters (last resort) |
| 92 | + */ |
| 93 | + private async secondPassSplit(chunks: Chunk[]): Promise<Chunk[]> { |
| 94 | + const finalChunks: Chunk[] = [] |
| 95 | + |
| 96 | + // Process all chunks, splitting large ones |
| 97 | + const splitPromises = chunks.map(async (chunk) => { |
| 98 | + // If chunk is within size limit, keep as-is |
| 99 | + if (chunk.content.length <= this.chunkSize) { |
| 100 | + return [chunk] |
| 101 | + } |
| 102 | + |
| 103 | + // Chunk is too large - use LangChain's recursive splitter |
| 104 | + const splitTexts = await this.recursiveSplitter.splitText(chunk.content) |
| 105 | + |
| 106 | + // Preserve the shouldTranslate flag on all sub-chunks |
| 107 | + return splitTexts.map((text) => ({ |
| 108 | + content: text, |
| 109 | + shouldTranslate: chunk.shouldTranslate, |
| 110 | + })) |
| 111 | + }) |
| 112 | + |
| 113 | + const splitResults = await Promise.all(splitPromises) |
| 114 | + |
| 115 | + // Flatten the results |
| 116 | + for (const result of splitResults) { |
| 117 | + finalChunks.push(...result) |
| 118 | + } |
| 119 | + |
| 120 | + return finalChunks |
| 121 | + } |
| 122 | + |
| 123 | + /** |
| 124 | + * Pass 1: Structural split using state machine (ported from Ruby implementation) |
| 125 | + * |
| 126 | + * Handles: |
| 127 | + * - Frontmatter detection (YAML between --- markers) |
| 128 | + * - Code block detection (``` fences) |
| 129 | + * - Split pattern matching (headers or ::page directives) |
| 130 | + * - shouldTranslate flag assignment |
| 131 | + */ |
| 132 | + private splitInternal(markdown: string, splitPattern: RegExp): Chunk[] { |
| 133 | + const chunks: Chunk[] = [] |
| 134 | + const lines = markdown.split('\n').map((line) => line + '\n') |
| 135 | + let currentPosition = 0 |
| 136 | + |
| 137 | + // Handle frontmatter if present |
| 138 | + if (this.hasFrontmatter(lines)) { |
| 139 | + const endIndex = this.findFrontmatterEnd(lines) |
| 140 | + const frontmatterContent = lines.slice(0, endIndex + 1).join('') |
| 141 | + chunks.push({ |
| 142 | + content: frontmatterContent, |
| 143 | + shouldTranslate: false, |
| 144 | + }) |
| 145 | + currentPosition = endIndex + 1 |
| 146 | + } |
| 147 | + |
| 148 | + let currentChunk: Chunk = {content: '', shouldTranslate: true} |
| 149 | + let inCodeBlock = false |
| 150 | + |
| 151 | + for (let i = currentPosition; i < lines.length; i++) { |
| 152 | + const line = lines[i] |
| 153 | + |
| 154 | + // Check for code fence (supports language identifiers and trailing whitespace) |
| 155 | + // Matches: ```, ```ruby, ```sh-session, ```python , etc. |
| 156 | + if (/^```[\w-]*\s*$/.test(line)) { |
| 157 | + if (inCodeBlock) { |
| 158 | + // End of code block |
| 159 | + currentChunk.content += line |
| 160 | + chunks.push(currentChunk) |
| 161 | + currentChunk = {content: '', shouldTranslate: true} |
| 162 | + inCodeBlock = false |
| 163 | + } else { |
| 164 | + // Start of code block |
| 165 | + if (currentChunk.content.trim()) { |
| 166 | + chunks.push(currentChunk) |
| 167 | + } |
| 168 | + |
| 169 | + currentChunk = {content: line, shouldTranslate: false} |
| 170 | + inCodeBlock = true |
| 171 | + } |
| 172 | + |
| 173 | + continue |
| 174 | + } |
| 175 | + |
| 176 | + // Handle line based on current state and patterns |
| 177 | + if (splitPattern.test(line) && !inCodeBlock) { |
| 178 | + // This line matches the split pattern and we're not in a code block |
| 179 | + if (currentChunk.content.trim()) { |
| 180 | + chunks.push(currentChunk) |
| 181 | + } |
| 182 | + |
| 183 | + currentChunk = {content: line, shouldTranslate: true} |
| 184 | + } else { |
| 185 | + // Regular line - append to current chunk |
| 186 | + currentChunk.content += line |
| 187 | + } |
| 188 | + } |
| 189 | + |
| 190 | + // Add final chunk if not empty |
| 191 | + if (currentChunk.content.trim()) { |
| 192 | + chunks.push(currentChunk) |
| 193 | + } |
| 194 | + |
| 195 | + // Filter out empty chunks |
| 196 | + return chunks.filter((chunk) => chunk.content.trim() !== '') |
| 197 | + } |
| 198 | +} |
0 commit comments