Section 3 · Data 360 Fundamentals · 20% of exam
3.3 Chunking
Exam objective
Explain how chunking prepares long or complex source content for search and retrieval, and identify why chunk design affects answer quality.
What a chunk represents?
Chunking divides a large source into smaller passages that can be searched and selected independently. A whole handbook, long Knowledge article, transcript, or file may contain many topics. If a user asks one narrow question, searching and returning the whole source is inefficient and may overwhelm the prompt. A chunk is intended to retain a meaningful unit of context, such as a coherent passage, rather than an arbitrary fragment. Salesforce describes chunking as a way to make search more efficient, improve relevance, and handle large datasets while respecting the amount of text a model can process in one call.
How chunks support retrieval
During data preparation, selected source fields or document content are processed into passages. Those passages can then be vectorized and indexed. At runtime, a user’s question is compared with searchable content, and the most relevant chunks can be returned to the prompt. Imagine a long warranty guide with sections for eligibility, exclusions, and claims. A query about an excluded item should ideally retrieve the exclusion passage and enough surrounding context to interpret it, not an unrelated introduction or the entire guide. Smaller passages can improve precision, but only if they preserve meaning. A fragment that loses its heading, product name, condition, or exception may be hard to interpret correctly.
Chunking strategy and structure
Data 360 supports different strategies for dividing content. Salesforce Trailhead describes semantic-based passage extraction, which can use structural cues such as headings, lists, and emphasized subheadings as logical boundaries. Window-based extraction can use block-level elements such as paragraphs and divs, or line breaks in raw text; when formatting is absent, sentence-level extraction can be used. The principle is to align passage boundaries with the structure of the source. Well-structured HTML or documents with clear headings can help preserve meaning. Poor formatting, repeated boilerplate, or very long unstructured passages can make extraction and retrieval less effective.
Balance passage size and context
Very small chunks may match a question closely but omit definitions or qualifications found in neighboring text. Large chunks preserve more context but can dilute relevance, consume more prompt space, and return unrelated material. There is no single ideal size for every dataset. A policy manual, product catalog, and conversational transcript have different structure and query patterns. In advanced Data 360 setups, admins can choose and tune chunking strategies and configuration. In a standard Agentforce Data Library, much of the preparation is automated, so first validate the source formatting, selected fields, and resulting retrieval behavior before pursuing custom configuration.
How to diagnose a chunking problem
When an agent cites a passage that is related but incomplete, inspect whether the source itself has the answer, whether the answer is split across a boundary, and whether headings or metadata are preserved. If the retriever finds no result, check that the source was processed and the relevant field was included. If results contain too much unrelated text, consider whether passage boundaries or the search strategy are appropriate. Test several phrasings, not only a question copied from the source, and include cases where a crucial condition appears before or after the target statement. Use the index configuration and processing history available in Data 360 to understand what content was actually chunked. Avoid changing chunk settings just because one answer failed; verify the cause first.
If a scenario says an answer is split across passages or retrieval returns irrelevant document sections, examine source structure and chunk boundaries before rewriting the agent’s instructions.
Chunking does not summarize a source or guarantee that each chunk is self-contained. Retrieval quality depends on preserving enough context and indexing the right fields.
Extra concept: Clear headings, descriptive section names, and concise source records improve the chance that passages remain understandable after they are separated from the full document.
Official references:
- What Are Data Libraries? https://help.salesforce.com/s/articleView?id=data_library_concept.htm&language=en_US&type=5
- Master Search Index Configurations for Effective Data Management https://trailhead.salesforce.com/content/learn/projects/unstructured-data-in-data-cloud/create-a-search-index-configuration
Key takeaways
- Chunking converts long content into smaller passages that can be searched independently.
- Passage boundaries should preserve semantic context such as headings, conditions, and exceptions.
- Diagnose source formatting and actual chunks before changing prompt instructions or search settings.
