Proposition Extraction

Proposition Extraction is the capability that decomposes source documents into discrete, independently verifiable factual claims using language models. Rather than treating paragraphs or passages as the smallest unit of evidence, Proposition Extraction identifies individual assertions within the text and makes each one separately retrievable, enabling claim-level precision in evidence matching.

Traditional retrieval systems operate at the passage level. A query about the melting point of a specific alloy might retrieve an entire paragraph that discusses the alloy's properties, including information about tensile strength, corrosion resistance, and thermal conductivity alongside the melting point. While the passage is relevant, most of the retrieved text is not directly pertinent to the specific question. Proposition Extraction solves this by extracting the individual claim about the melting point as a standalone proposition with its own embedding.

The extraction process uses a language model to identify discrete factual statements within each semantic chunk. A paragraph might contain multiple propositions such as specific numerical values, causal relationships, temporal claims, and categorical assertions. Each extracted proposition is validated for coherence and self-containedness, ensuring it can be understood without the surrounding context. Propositions are linked back to their source chunks and documents through the provenance system, maintaining the full context chain.

Each proposition receives its own embedding vector, creating a fine-grained retrieval index that complements the chunk-level index. When Engine processes a query, it can retrieve evidence at both the proposition level for precise claim matching and the chunk level for broader contextual understanding. The combination produces evidence sets that are both precisely targeted and contextually rich.

Proposition Extraction is particularly valuable for the Trust Gradient and stress testing systems. Because propositions represent individual claims, the platform can assess confidence and fragility at the individual assertion level rather than the passage level. A paragraph might contain five propositions, three well-supported and two thinly evidenced. Without proposition extraction, the entire paragraph would receive a single aggregate confidence score that obscures these important differences.

The extracted propositions also feed into the Coverage Labelling system. By identifying every factual claim made in an answer and checking whether each one has direct propositional support in the evidence, the platform can provide much more granular coverage assessments than would be possible with passage-level matching alone.

For the CueCrux platform, Proposition Extraction represents a commitment to precision in evidence matching. It enables the platform to answer the question of whether a specific claim is supported by the evidence not just at the level of the evidence being generally relevant but at the level of the evidence directly stating the same fact.