Skip to main content

​
  Partition

Partitioning converts an unstructured file or semi-structured data record into structured output. During partitioning, Unstructured produces structured document elements and metadata in a predefined, consistent, expressive, and contextualized JSON format. Unstructured supports several partitioning strategies. They range from fast, rule-based text processing to vision language model (VLM) processing for complex layouts, handwriting, and multilanguage content. Choosing the right strategy lets you balance speed, cost, and output quality for your use case. Learn more about partitioning.

​
  Extract

Extraction produces structured output from an unstructured file or semi-structured data record. However, unlike partitioning, the structured data extractor lets you define your own target JSON schema. Unstructured then extracts values from your source file or record directly into that schema. Learn more about extraction.

​
  Enrich

Enrichments add AI-generated enhancements to partitioned output. Enrichments include image descriptions, table descriptions, table-to-HTML conversion, named entity recognition (NER), and generative OCR. Generative OCR improves text accuracy in complex documents and data records. Enriching gives your downstream applications richer, more useful data from source content that would otherwise be hard to work with, such as images, handwritten text, or dense tables. Learn more about enriching.

​
  Chunk

Chunking reorganizes partitioned output into manageable pieces sized for embedding models and optimized for retrieval precision. Instead of embedding entire documents, chunking ensures that each piece of retrieved content is focused and relevant to a user’s query. Unstructured offers several chunking strategies (by character count, section, page, or semantic similarity) so you can tune chunk boundaries to match your content and retrieval goals. Learn more about chunking.

​
  Embed

Embedding converts text output into numeric vectors using an embedding model. These vectors capture semantic meaning. Unstructured stores them alongside the text so you can load them into a vector store. Vector embeddings power similarity search in RAG applications. When a user submits a query, the application finds the chunks whose embeddings are closest to that query and returns the most relevant results. Learn more about embedding.

​
See also