Chunking is the process of splitting a long document into smaller passages, called chunks, so an AI system can embed, index, and retrieve each piece independently. In commercial real estate document work, chunking breaks a 90-page lease or offering memorandum into passages small enough to search by meaning and feed to a language model.
How Does Chunking Work?
Chunking works by cutting a document into passages of a target size, usually measured in tokens, then embedding each passage as a separate vector for retrieval. A boundary rule decides where each cut falls, and an overlap setting repeats a slice of text across neighboring chunks so a clause split across a boundary still appears intact in at least one passage.
A common starting point is a recursive splitter set to 400 to 512 tokens with 10 to 20 percent overlap, per a 2025 practitioner guide from Firecrawl. The overlap keeps context from being lost at cut points. Strategies differ in where they place boundaries.
Strategy | How it cuts | Trade-off |
Fixed-size | Every N tokens | Fast, but splits mid-clause |
Recursive | On paragraph, then sentence, then word | Preserves structure, common default |
Semantic | Where meaning shifts | Highest accuracy, highest cost |
Page or section | On document layout | Keeps a full clause together |
Why Chunking Matters
Chunking matters because retrieval quality caps answer quality, and chunk boundaries decide what a model sees. A chunk that splits a rent escalation clause in half hands the model an incomplete provision, and the answer inherits the gap. Chunk too small and context is lost; chunk too large and irrelevant text dilutes the match.
The size choice is measurable, not arbitrary. An NVIDIA technical analysis found semantic chunking can lift retrieval accuracy meaningfully over naive fixed-size baselines on some tasks, while a January 2026 systematic study cited by Firecrawl identified a "context cliff" near 2,500 tokens where response quality drops. For dense legal text like a lease, keeping a full clause inside one chunk is the practical priority, which favors recursive or section-aware splitting over blind fixed cuts.
Example
Chunking a lease is clearest with a worked count. A 90-page lease at roughly 500 words per page holds about 45,000 words, which at 1.33 tokens per word, the standard English ratio reported by llmtest.io, is about 59,850 tokens.
Input | Value |
Pages | 90 |
Words per page | 500 |
Total words | 45,000 |
Tokens per word | 1.33 |
Total tokens | 59,850 |
Chunk size | 500 tokens |
Overlap | 15 percent (75 tokens) |
With 500-token chunks and a 15 percent overlap, each chunk advances 425 net tokens. Dividing 59,850 by 425 gives about 141 chunks. Instead of feeding all 59,850 tokens to a model for one rent question, the system retrieves the few chunks nearest the query, cutting the text the model reads by more than 99 percent while keeping the relevant clause intact.
Variations and Edge Cases
Chunking behaves differently by document type, and CRE documents mix prose, tables, and forms in one file. The variants below matter when a rent roll table and a dense legal clause live in the same document.
Variant | Behavior |
Table-aware chunking | Keeps a rent roll row set together instead of splitting a table |
Overlapping windows | Repeats text across chunks so boundary clauses stay whole |
Parent-child chunking | Retrieves a small chunk, then expands to its larger parent for context |
Layout-based chunking | Cuts on headings and sections rather than token count |
Metadata tagging | Attaches page and section labels so a citation can point back to the source |
Chunking vs Tokenization
Chunking is often confused with tokenization, but they operate at different scales. Chunking splits a document into passages of many words for retrieval. Tokenization splits text into sub-word units a model reads one at a time.
Tokenization is the finer step: it breaks "escalation" into pieces a model processes and is measured in tokens per word. Chunking sits above it, grouping hundreds of tokens into a retrievable passage. Chunk size is defined in tokens, so tokenization is the unit chunking counts in, but the two answer different questions: tokenization asks how a model reads text, chunking asks how a document is divided for search.
Frequently Asked Questions
What is chunking in document AI?Chunking is the process of splitting a long document into smaller passages, called chunks, so an AI system can embed, index, and retrieve each piece independently. In commercial real estate it breaks a lease or offering memorandum into passages small enough to search by meaning and feed to a language model.
What chunk size and overlap should I use?A common starting point is 400 to 512 tokens with 10 to 20 percent overlap, per a 2025 Firecrawl practitioner guide, and 15 percent overlap performed best on the FinanceBench task at 1,024-token chunks. For dense lease clauses, keeping a full provision inside one chunk matters more than any fixed number.
Why does chunking affect extraction accuracy?Chunking affects accuracy because retrieval quality caps answer quality, and a chunk that splits a clause in half hands the model an incomplete provision. If the right passage is never retrieved intact, the model cannot answer from it, no matter how capable the model is.
Related Terms
Retrieval-Augmented Generation
Vector Embedding
Tokenization
Semantic Search
Extraction Accuracy