StructuredDocument

StructuredDocument wraps raw text inputs, runs optional extractors, and then delegates to a chunker strategy. Each node keeps a docId and links to generated chunks.

Workflow

1. Create a document node (or multiple nodes).
2. Run extractors (title, summary, keywords, questions).
3. Chunk the document using a strategy.
4. Read the link graph for document → chunk relationships.

Usage

import { StructuredDocument } from "@voltagent/rag";

const doc = StructuredDocument.fromText(`# Heading
Body paragraph with details.`);

doc.extract({ title: true, summary: true, keywords: true, questions: true });

const { chunks } = doc.chunk({ strategy: "markdown", maxTokens: 200 });

// Example chunk output:
// [
//   {
//     content: "Body paragraph with details.",
//     metadata: { sourceType: "markdown", blockType: "paragraph", headingPath: ["Heading"], docId: "<doc-id>" },
//   },
// ]

const nodes = doc.getNodes();
const links = doc.getLinkGraph();

// links example:
// { "<doc-id>": ["<chunk-id>"] }

Strategies

• "markdown": uses MarkdownChunker
• "html": uses HtmlChunker
• "json": uses JsonChunker
• "latex": uses LatexChunker
• "recursive": uses RecursiveChunker
• "sentence": uses SentenceChunker
• "token": uses TokenChunker
• "table": uses TableChunker
• "code": uses CodeChunker
• "auto": heuristic: JSON parseable → json, HTML tags → html, LaTeX commands → latex, fenced code → code, table-like pipes → table, else recursive

chunk() forwards maxTokens to the underlying chunker. Chunker-specific options (tokenizer, parser, labels) use chunker defaults; configure those chunkers directly if you need custom tokenization or parsing.

Extractors

Enable any of the following booleans:

• title
• summary
• keywords
• questions

Each extractor appends metadata to the document node; chunk metadata always includes docId so downstream consumers can relate chunks back to their source nodes.