Retrieval settings
The two settings that decide which passages the model gets to see, and when to touch them.
Before the model writes, the search picks the passages it may use. Two settings under Settings, in the block Retrieval, steer that choice. The defaults are measured, not guessed, and most agents never need to change them.
How the passages are chosen
- The search ranks your passages by how close they are in meaning to the question.
- Passages below the Minimum similarity are dropped.
- Of what is left, the best Top-K results go to the model.
If nothing is left after step 2, the model is told that nothing was found, and it says it does not know (or answers from a fact in your instructions).
Top-K results
How many passages the model receives, from 1 to 10. The default is 5.
- Raise it (7 to 10) when answers are incomplete because the information is spread over several pages, such as a question that needs both the price list and the delivery terms.
- Lower it (2 to 3) when answers pull in things that were not asked, or mix two products up. Fewer passages means less to confuse.
More passages is not simply better. Each passage is about 500 tokens of reading, and a model that gets ten of them when two were relevant has eight chances to be led astray.
Minimum similarity
How close a passage must be to the question to count at all, from 0 to 1. The default is 0.25.
That looks low, and it is on purpose. Similarity is not a percentage of correctness. In our measurements a passage that is entirely about the question scores around 0.67, and a relevant paragraph inside a broader page scores around 0.29. A threshold of 0.5 would throw away most good answers.
- Raise it a little (0.30 to 0.35) when the agent answers questions it should have declined, using passages that are only vaguely related.
- Lower it (0.15 to 0.20) when the agent says "I don't know" about things your sources plainly cover. This happens most with very short questions and with questions in another language than your sources.
Change it in small steps and test with the same questions each time.
What these settings do not change
Confidence is measured the same way on every agent, always over the same number of candidates, whatever these two settings are. Moving a slider never makes the numbers in Overview look better or worse by itself. That is deliberate: a number you can change by turning a knob tells you nothing.
Before you touch them
Most weak answers are not a settings problem. If the information is not in your sources, no setting finds it. Check Fixing weak answers first. Settings are the last step there, not the first.