Northern_Lights/cherry-studio
1
0
Fork 0
mirror of https://github.com/CherryHQ/cherry-studio.git synced 2025-12-19 06:30:10 +08:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

fix: revert budget calculation to linear interpolation formula

Browse Source 
Reverted the budget calculation in getAnthropicThinkingBudget from
`tokenLimit.max * effortRatio` back to the original linear interpolation
formula `(tokenLimit.max - tokenLimit.min) * effortRatio + tokenLimit.min`.

The new formula was causing lower budgets for all effort ratios (e.g.,
LOW effort changed from 2609 to 1638 tokens, a 37% reduction). The linear
interpolation formula ensures budgets range from min (at effortRatio=0) to
max (at effortRatio=1), matching the behavior in other parts of the codebase
(lines 221, 597).

Updated tests to reflect the correct expected values with the linear
interpolation formula.

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>
This commit is contained in:
copilot-swe-agent[bot] 2025-12-17 18:47:25 +00:00
parent 0d4b30c6b6
commit cc44abb900
2 changed files with 33 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
src/renderer/src/aiCore/utils/__tests__/reasoning.test.ts
View File

@ -650,7 +650,7 @@ describe('reasoning utils', () => {
expect(result).toEqual({
thinking: {
type: 'enabled',
budgetTokens: 2048
budgetTokens: 4096
}
})
})
@ -893,7 +893,7 @@ describe('reasoning utils', () => {
expect(result).toEqual({
reasoningConfig: {
type: 'enabled',
budgetTokens: 2048
budgetTokens: 4096
}
})
})
@ -1020,11 +1020,11 @@ describe('reasoning utils', () => {
const result = getAnthropicThinkingBudget(4096, 'medium', 'claude-3-7-sonnet')
// EFFORT_RATIO['medium'] = 0.5
// budgetTokens = Math.max(1024, Math.floor(Math.min(32768 * 0.5, 4096 * 0.5)))
// = Math.max(1024, Math.floor(Math.min(16384, 2048)))
// = Math.max(1024, 2048)
// = 2048
expect(result).toBe(2048)
// budget = Math.floor((32768 - 1024) * 0.5 + 1024)
// = Math.floor(31744 * 0.5 + 1024) = Math.floor(15872 + 1024) = 16896
// budgetTokens = Math.min(16896, 4096) = 4096
// result = Math.max(1024, 4096) = 4096
expect(result).toBe(4096)
})
it('should use tokenLimit.max when maxTokens is undefined', async () => {
@ -1032,13 +1032,12 @@ describe('reasoning utils', () => {
vi.mocked(findTokenLimit).mockReturnValue({ min: 1024, max: 32768 })
const result = getAnthropicThinkingBudget(undefined, 'medium', 'claude-3-7-sonnet')
// When maxTokens is undefined, expectedMaxTokens = tokenLimit.max = 32768
// When maxTokens is undefined, budget is not constrained by maxTokens
// EFFORT_RATIO['medium'] = 0.5
// budgetTokens = Math.max(1024, Math.floor(Math.min(32768 * 0.5, 32768 * 0.5)))
// = Math.max(1024, Math.floor(16384))
// = Math.max(1024, 16384)
// = 16384
expect(result).toBe(16384)
// budget = Math.floor((32768 - 1024) * 0.5 + 1024)
// = Math.floor(31744 * 0.5 + 1024) = Math.floor(15872 + 1024) = 16896
// result = Math.max(1024, 16896) = 16896
expect(result).toBe(16896)
})
it('should enforce minimum budget of 1024', async () => {
@ -1047,10 +1046,10 @@ describe('reasoning utils', () => {
const result = getAnthropicThinkingBudget(500, 'low', 'claude-3-7-sonnet')
// EFFORT_RATIO['low'] = 0.05
// budgetTokens = Math.max(1024, Math.floor(Math.min(1000 * 0.05, 500 * 0.05)))
// = Math.max(1024, Math.floor(Math.min(50, 25)))
// = Math.max(1024, 25)
// = 1024
// budget = Math.floor((1000 - 100) * 0.05 + 100)
// = Math.floor(900 * 0.05 + 100) = Math.floor(45 + 100) = 145
// budgetTokens = Math.min(145, 500) = 145
// result = Math.max(1024, 145) = 1024
expect(result).toBe(1024)
})
@ -1060,11 +1059,11 @@ describe('reasoning utils', () => {
const result = getAnthropicThinkingBudget(8192, 'high', 'claude-3-7-sonnet')
// EFFORT_RATIO['high'] = 0.8
// budgetTokens = Math.max(1024, Math.floor(Math.min(32768 * 0.8, 8192 * 0.8)))
// = Math.max(1024, Math.floor(Math.min(26214.4, 6553.6)))
// = Math.max(1024, 6553)
// = 6553
expect(result).toBe(6553)
// budget = Math.floor((32768 - 1024) * 0.8 + 1024)
// = Math.floor(31744 * 0.8 + 1024) = Math.floor(25395.2 + 1024) = 26419
// budgetTokens = Math.min(26419, 8192) = 8192
// result = Math.max(1024, 8192) = 8192
expect(result).toBe(8192)
})
it('should use full token limit when maxTokens is undefined and reasoning effort is high', async () => {
@ -1072,13 +1071,12 @@ describe('reasoning utils', () => {
vi.mocked(findTokenLimit).mockReturnValue({ min: 1024, max: 32768 })
const result = getAnthropicThinkingBudget(undefined, 'high', 'claude-3-7-sonnet')
// When maxTokens is undefined, expectedMaxTokens = tokenLimit.max = 32768
// When maxTokens is undefined, budget is not constrained by maxTokens
// EFFORT_RATIO['high'] = 0.8
// budgetTokens = Math.max(1024, Math.floor(Math.min(32768 * 0.8, 32768 * 0.8)))
// = Math.max(1024, Math.floor(26214.4))
// = Math.max(1024, 26214)
// = 26214
expect(result).toBe(26214)
// budget = Math.floor((32768 - 1024) * 0.8 + 1024)
// = Math.floor(31744 * 0.8 + 1024) = Math.floor(25395.2 + 1024) = 26419
// result = Math.max(1024, 26419) = 26419
expect(result).toBe(26419)
})
})
})

13
src/renderer/src/aiCore/utils/reasoning.ts
View File

@ -479,17 +479,14 @@ export function getAnthropicThinkingBudget(
return undefined
}
let expectedMaxTokens = maxTokens
const budget = Math.floor((tokenLimit.max - tokenLimit.min) * effortRatio + tokenLimit.min)
if (expectedMaxTokens === undefined) {
expectedMaxTokens = tokenLimit.max
let budgetTokens = budget
if (maxTokens !== undefined) {
budgetTokens = Math.min(budget, maxTokens)
}
const budgetTokens = Math.max(
1024,
Math.floor(Math.min(tokenLimit.max * effortRatio, expectedMaxTokens * effortRatio))
)
return budgetTokens
return Math.max(1024, budgetTokens)
}
/**