Propositional Chunking

Each chunk represents exactly one proposition—a single, self-contained factual claim that cannot be further split without losing meaning. This contrasts with fixed-length chunking, where a 512-token block might contain 5-7 unrelated facts. By isolating propositions, retrieval systems achieve granularity control at the finest possible level, ensuring that only the most relevant atomic chunk is returned for a given query rather than a bloated paragraph containing extraneous information.

Propositional chunks exhibit exceptionally high chunk information density—the ratio of unique factual content to total token length. Because filler words, transitions, and rhetorical structures are stripped away, every token in the chunk contributes directly to the encoded fact. This makes propositional chunks ideal for vector space positioning, as the embedding model focuses entirely on the semantic content of the fact rather than being diluted by surrounding prose.

By eliminating chunk contamination—where a single chunk contains multiple unrelated topics—propositional chunking prevents irrelevant data from leaking into the LLM's generation context. When a user queries a specific fact, the retriever returns only that precise proposition rather than a larger block that might include tangential or contradictory information. This directly improves re-ranking accuracy and reduces hallucination risk in the final synthesis.

Each propositional chunk must be fully self-contained and understandable without external context. This requires rewriting pronouns into explicit entities and resolving anaphoric references. For example, 'It increased by 15%' becomes 'Company X's Q3 2024 revenue increased by 15% year-over-year.' This decontextualization process is essential for achieving high chunk coherence and ensuring accurate retrieval regardless of which document the proposition originated from.

Propositional chunks are typically paired with rich metadata enrichment to preserve provenance and enable filtered retrieval. Each atomic fact carries structured attributes including:

• Source document identifier and URL
• Section heading hierarchy
• Publication or last-updated timestamp
• Entity tags for mentioned organizations, people, or products This metadata enables scoped queries and supports chunk attribution for citation generation in the final LLM response.

The primary cost of propositional chunking is a significantly larger vector database infrastructure footprint. A single 1,000-word document might produce 40-60 propositional chunks versus 4-6 semantic chunks. This increases embedding compute costs, index storage requirements, and retrieval latency. Engineering teams must balance this against the precision gains, often implementing hybrid retrieval strategies where propositional chunks serve as the fine-grained index while larger parent chunks provide expanded context during synthesis.

The indivisible unit of retrievable content representing a single, self-contained fact. Propositional chunking explicitly targets this granularity.

• Key characteristic: Cannot be split further without semantic loss
• Example: "Paris is the capital of France" is atomic; "Paris, the capital of France, has a population of 2.1 million" contains two atoms
• Trade-off: Maximizes retrieval precision but increases total chunk count and index size

Splits text based on meaning shifts detected via embedding similarity rather than fixed token counts. Propositional chunking is a specialized, higher-precision subset of this approach.

• Mechanism: Computes cosine similarity between adjacent sentences; splits where similarity drops below a threshold
• Contrast with propositional: Semantic chunking may group multiple related facts; propositional isolates each one
• Use case: General-purpose document segmentation where some contextual grouping is acceptable

A quality metric measuring whether a text segment contains a logically complete and self-contained idea. Propositional chunking maximizes this score by design.

• High coherence: "The sky is blue due to Rayleigh scattering" — no external context needed
• Low coherence: "Due to this phenomenon, it appears blue" — requires prior chunk for resolution
• Impact: High coherence chunks reduce hallucination risk and improve retrieval relevance

The ratio of unique factual content to total token length within a chunk. Propositional chunks exhibit maximum density by eliminating all non-essential tokens.

• Calculation: (Number of unique facts) / (Total tokens)
• Propositional advantage: Every token serves a factual purpose; no filler or transitional language
• Retrieval benefit: High-density chunks improve signal-to-noise ratio in the LLM's context window

A retrieval failure mode where a chunk contains information from multiple unrelated topics, causing irrelevant data to leak into the LLM's generation context. Propositional chunking eliminates this risk.

• Example of contamination: A paragraph discussing both "machine learning algorithms" and "data privacy regulations" in one chunk
• Propositional solution: Each atomic statement is isolated, so a query about algorithms never retrieves privacy content
• Result: Cleaner retrieval, fewer hallucinations, more precise citations

A post-retrieval stage where a more computationally intensive model re-scores initial search results. Propositional chunks benefit significantly from re-ranking due to their uniform granularity.

• Why it matters: Initial vector similarity may return near-miss chunks; re-ranking applies deeper semantic comparison
• Propositional synergy: Uniform atomic chunks give re-rankers cleaner, more comparable units to score
• Common models: Cohere Rerank, bge-reranker, cross-encoder architectures