What changed our advice

A useful research program records when a more elaborate approach did not help. These changes come from the existing author-reported context experiments; their scope matters.

More representations were not a universal improvement

The reported experiments found settings where multiple views helped and settings where matched chunks performed better. The advice is to start with a matched baseline and test additional views on the actual questions. See real-prose comparisons.

The objective did not reliably improve the index

The tested forms of objective-conditioned representation did not support the expected improvement. The guide therefore does not recommend treating the business objective as an automatic index-design signal. See the comparisons and limits.

Retrieval is only part of the outcome

The consuming model and the composition of the retrieved set affect what happens after search. A retrieval score is not a complete measure of task success. See diversity and the consumer.

New media are new hypotheses

These text results neither prove nor disprove a particular satellite, image, or time-series representation. Evaluate those on their own data and questions, with exact source metadata preserved.

Read what we got wrong and the research method.