Agents recall Docs passages only when they are about the question: the meaning floor is 0.7, measured on real pages, so "hello" or an unrelated question no longer pulls in runbooks · 0637405 · flagon-io/g1t

8

9

/**

10

* The least cosine similarity a passage needs to count as being about the

11

−

* query. bge-base-en-v1.5 scores unrelated English around 0.4 to 0.55 and

12

−

* a passage on the asked-about topic from about 0.65; 0.6 keeps recall

13

−

* quiet when the docs say nothing about it.

11

+

* query. Measured on g1t's own docs folder (606 passages, chunked and

12

+

* embedded as here, bge-base-en-v1.5 with Workers AI's mean pooling): the

13

+

* best passage for 16 questions the docs answer scored 0.72 to 0.84, while

14

+

* the best for 16 they don't ("hello", a recipe, Postgres tuning, a

15

+

* vacation policy, SSO with Okta) scored 0.54 to 0.67. At 0.6, 11 of those

16

+

* 16 pulled in passages (up to 89 above it); 0.7 keeps every answer and

17

+

* keeps recall quiet when the docs say nothing about the question.

14

18

*/

15

−

export const MEANING_FLOOR = 0.6;

19

+

export const MEANING_FLOOR = 0.7;

16

20

/** The score a passage found only by its words carries: below the floor, so callers can tell. */

17

21

export const WORDS_SCORE = 0.5;

18

22

/** Nearest passages asked of the index. */

46	46		});
47	47
48	48		test("required spaces come first, then the rest", () => {
49		−	const picked = pickPassages([c("a:0", "s1", 0.9), c("b:0", "s2", 0.65), c("c:0", "s1", 0.85)], { allowed: new Set(["s1", "s2"]), required: ["s2"], limit: 2 });
	49	+	const picked = pickPassages([c("a:0", "s1", 0.9), c("b:0", "s2", 0.75), c("c:0", "s1", 0.85)], { allowed: new Set(["s1", "s2"]), required: ["s2"], limit: 2 });
50	50		assert.deepEqual(
51	51		picked.map((p) => p.id),
52	52		["b:0", "a:0"],