cupel 1.2.0

Context window management pipeline for LLM applications
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56

[test]
name = "KnapsackSlice selects optimally where greedy would not"
stage = "slicing"
slicer = "knapsack"

# Budget: 300 tokens, bucket_size: 100
# This is a case where knapsack differs from greedy.
#
# Items:
#   "big":     tokens=250, score=0.7 → value=7000, dw=ceil(250/100)=3
#   "small-a": tokens=150, score=0.5 → value=5000, dw=ceil(150/100)=2
#   "small-b": tokens=150, score=0.45 → value=4500, dw=ceil(150/100)=2
#
# Capacity = floor(300/100) = 3
#
# Greedy (by density):
#   big:     0.7/250 = 0.0028
#   small-a: 0.5/150 ≈ 0.00333
#   small-b: 0.45/150 = 0.003
#   → small-a first (dw=2, remaining=1), small-b next (dw=2 > 1, skip),
#     big (dw=3 > 1, skip). Only "small-a" selected. Total value = 5000.
#
# Knapsack DP:
#   At capacity 3, can fit big (dw=3, value=7000) or small-a (dw=2, value=5000).
#   big at w=3: dp[3-3]+7000 = 7000 > 0 → dp[3]=7000, keep[0][3]=true
#   small-a at w=3: dp[3-2]+5000 = 5000 < 7000 → no change
#   small-a at w=2: dp[2-2]+5000 = 5000 > 0 → dp[2]=5000, keep[1][2]=true
#   small-b at w=3: dp[3-2]+4500 = dp[1]+4500 = 4500 < 7000 → no change
#   small-b at w=2: dp[2-2]+4500 = 4500 < 5000 → no change
#
#   Reconstruct from capacity=3: keep[2][3]=false, keep[1][3]=false, keep[0][3]=true → "big"
#   → Knapsack selects "big" (value 7000 > 5000)

[config]
bucket_size = 100

[budget]
target_tokens = 300

[[scored_items]]
content = "big"
tokens = 250
score = 0.7

[[scored_items]]
content = "small-a"
tokens = 150
score = 0.5

[[scored_items]]
content = "small-b"
tokens = 150
score = 0.45

[expected]
selected_contents = ["big"]