Propositional Indexing

Each chunk represents a single, self-contained factual statement extracted from the source document. A complex paragraph containing three distinct legal holdings is split into three separate, independently retrievable propositions. This prevents the semantic dilution that occurs when a dense embedding represents multiple disparate concepts simultaneously.

Every proposition is engineered to be fully interpretable in isolation. The indexing process rewrites or annotates chunks to resolve co-references and implicit dependencies. For example, 'The court held that it was unconstitutional' becomes 'The Marbury v. Madison court held that Section 13 of the Judiciary Act of 1789 was unconstitutional.' This eliminates the need to retrieve surrounding context to understand a chunk's meaning.

Because each vector represents a single, unambiguous fact, the cosine similarity between a query embedding and a proposition embedding yields a highly reliable relevance signal. This contrasts sharply with naive chunking, where a 512-token passage might contain one relevant sentence buried within eleven irrelevant ones, producing a misleadingly high similarity score.

Each proposition chunk maintains a strict one-to-one or many-to-one mapping to its source citation. When a generative model grounds a claim on a retrieved proposition, the provenance is unambiguous. This enables high-precision citation verification, as there is no ambiguity about which sentence within a larger paragraph actually supports the generated output.

Propositional indexing enables hard negative mining during embedding model fine-tuning. The system can identify pairs of propositions that are lexically similar but legally distinct—such as a majority holding and a dissent on the same issue—and explicitly train the model to push their embeddings apart in vector space, improving retrieval discrimination.

Legal queries are often themselves specific propositions. A user asking 'Is a warrant required for cell site location data?' is seeking a direct propositional match. Propositional indexing aligns the granularity of the index with the granularity of the typical legal research query, enabling direct fact-to-fact matching rather than broad topic alignment.

A retrieval architecture that searches using small, focused sentence chunks but returns the larger parent paragraph or section to the generator. This preserves surrounding legal context while maintaining the precision of propositional indexing. The technique ensures that an atomic factual proposition is not interpreted in isolation, providing the full statutory or precedential framework.

A preprocessing technique that prepends chunk-specific explanatory context to each text chunk before embedding. This prevents isolated propositions from losing their legal meaning when retrieved. For example, a proposition stating 'the defendant must file within 30 days' is prepended with context like 'This excerpt is from FRCP Rule 59(e) regarding motions to alter or amend a judgment.'

A post-retrieval step where a computationally intensive cross-encoder model re-orders a candidate list of propositions to prioritize those most semantically relevant to a complex legal query. This is critical after propositional indexing, as the atomic chunks may have high lexical overlap but different legal implications that require deeper semantic comparison.

The process of forcing a generative model to anchor every factual claim or legal proposition in its output to a specific, verifiable source document chunk. Propositional indexing directly enables this by providing the granular, atomic evidence units that can be cited individually, rather than citing an entire page or section.

An iterative search process where the answer to an initial query is used to formulate a secondary query to find connecting authority. Propositional indexing excels here because each hop can retrieve a single, precise factual statement rather than a large block of text containing irrelevant information, enabling the construction of a clean logical evidence chain.

A natural language inference task that determines whether a specific legal hypothesis can be logically concluded from a given set of premises. Propositional indexing provides the ideal input format for entailment models by isolating the exact factual premises from case law, allowing for binary true/false assessment of whether a conclusion is supported.