---
url: >-
  https://adk.nht.io/api/@nhtio/adk/batteries/context/thrift/relevance/variables/RELEVANCE_FLOOR_MIN.md
description: >-
  The relevance floor's lower bound — the minimum fraction of the current
  query's content words a turn must share to be kept, used when the window is
  nearly empty (permissive: keep almost anything that shares ANY real overlap).
---

# Variable: RELEVANCE\_FLOOR\_MIN

```ts
const RELEVANCE_FLOOR_MIN: 0.125 = 0.125;
```

Defined in: [src/batteries/context/thrift/relevance.ts:44](https://github.com/NHTIO/ADK/blob/v1.20260713.1/src/src/batteries/context/thrift/relevance.ts#L44)

The relevance floor's lower bound — the minimum fraction of the current query's content words a
turn must share to be kept, used when the window is nearly empty (permissive: keep almost anything
that shares ANY real overlap).

## Remarks

CALIBRATED against a relevance oracle over a 94-turn stress corpus, triple-confirmed across three
model families acting as independent judges (gpt-5.5, gemini-3.5-flash, claude-haiku-4.5): the
F1-optimal floor (best balance of precision and recall when window budget is abundant) landed at
`0.125` for gpt-5.5 and claude-haiku-4.5 exactly, and `0.111` for gemini-3.5-flash — `0.125` was
chosen as the shared default (a hair stricter than gemini's number, i.e. slightly more permissive
than not, which the calibration run characterized as "slightly permissive" rather than
under-inclusive — the safer direction to err when the window has room to spare).