Decontextualization
In fact, small models perform usually better than larger ones. Thinking models, in particular, are slow and they analyze rather than extract.
Raw text is dense, co-referential, and context-dependent. Sentences refer to things like “it”, “the company”, “this approach”… entities that only resolve if you’ve read the paragraph before. Feed that directly to an extractor and you get ambiguous predicates, dangling references, and relations that fall apart outside their original context.
Atomic proposition extraction is a pre-processing step that fixes this. Before any triple is produced, the document is decomposed into minimal, self-contained claims, each expressing exactly one fact, with all pronouns and references resolved to their full entity names. The extractor then operates on units that are already semantically isolated. Something like KGGen, a Stanford pipeline built on top of DSPy, achieves 66% extraction accuracy this way. The improvement comes not from a better extraction prompt, but from the structure of the input.
Decontextualization (creating a self-contained chunk) is the step most implementations skip or underestimate. It requires a sliding context window. Each chunk is resolved against prior material before extraction begins. If done well, it enables parallelization and cutting wall-clock time significantly.
On the model side, latent reasoning optimization (LaTRO) offers a complementary lever. Rather than relying on chain-of-thought prompting or external reward signals, LaTRO treats reasoning as sampling from a latent distribution and optimizes it variationally hence improving both reasoning quality and self-evaluation in a single training loop. Applied to the extraction step, it can sharpen triple inference from propositions without touching the pipeline structure.
For temporal knowledge graphs, Graphiti takes this further by attaching validity intervals to each extracted fact, enabling the graph to represent a world that changes over time rather than a static snapshot.