Entity explicitness
A chunk that relies on earlier text for its subject retrieves worse. Anthropic's Contextual Retrieval write-up gives the example of a chunk saying the company's revenue grew by 3% over the previous quarter without naming the company or period, and reports that prepending a short chunk-specific context before embedding cut top-20 retrieval failures by 35% (5.7% to 3.7%), and by 49% when also applied to BM25. That naming the product in each section has the same effect is an inference from this result, not a separate measurement.
Sources
- 1
CGrade C: reputable industry data or observation. A grade describes the provenance of the sources above; it is not a probability that a change will work. The effect basis says how the size of the effect is known: measured by a study, taken from an industry report, only documented without a size, or not quantified.
Put this evidence to work
Related entries
More in Retrieval, or search the whole base.
All Retrieval entries