Document chunking
Splitting a document into smaller coherent sections for search and model context.
Which context must remain within the section?
Document chunking splits documents into smaller coherent sections for retrieval or presentation to a model. Boundaries can follow sentences, headings and tables. This can accommodate input limits and make distinct topics easier to retrieve accurately. Size and possible overlap depend on content and task; there is no universally ideal length. Each part still needs links to its source, version and applicable access rules.
In a fictional knowledge collection, a table identifies the equipment identifier as a required field for the portal form. If only the table row is retrieved without column headings or the section title, it can become unclear which reporting route the requirement applies to. Preparation therefore carries the relevant heading and table context with the section. Planned search questions assess whether retrieved sections preserve this distinction and lead back to the original source. When a version is replaced, the affected search sections must also be updated traceably.
Sources
Related terms
- Context windowThe context window limits how many tokens a language model can consider in one processing step. It is not a permanent knowledge archive.
- Language model tokenTokens are processing units for language models. They do not reliably correspond to whole words and differ from design tokens.
- EmbeddingAn embedding represents content as a numeric vector. The model and comparison method determine which similarities become visible.
- Retrieval-augmented generationRAG combines information retrieval with generated answers. Access to sources does not remove the need for quality and permission design.
- Semantic searchSearch that finds and ranks results by their semantic relevance to the query.
- Hybrid searchCombining lexical search and vector search into a shared result list.
Your project with ASES
Describe your project to us. We will get back to you with an initial assessment.