Skip to main content

Retrieval

A question is searched two ways inside its workspace:
  • By meaning: the question’s embedding against every passage’s embedding.
  • By keyword: the question’s words of three letters or more against the passages’ full text. One rare word is enough to match.
The two rankings are merged with reciprocal-rank fusion: 24 candidates in, the top 6 passages out. A passage that ranks well in both rises to the top.

The answer

The model gets the question and the six passages, numbered, and is told to:
  • answer only from the passages;
  • end every sentence that states a fact with the number of the passage that supports it, like [1] or [2][3];
  • say exactly “I can’t find this in your sources.” when the passages don’t answer the question.

Citations

The numbers in the answer are renumbered 1, 2, 3 in reading order. Each citation carries:

Refusals

A refusal is an answer, not an error. It means the workspace has nothing on the question. Add a source that covers it, or check that the source finished ingesting. To check how often Context refuses correctly, and how often its answers are grounded, score the workspace with Evals.

Models

Answers are written by open-weight models. If one is busy, the next one in the list answers.