feat: preserve pages and compress context

src/shared/context-window.ts

@@ -0,0 +1,63 @@
+export const minimumModelContextWindowTokens = 32_000
+export const maximumModelContextWindowTokens = 10_000_000
+export const contextOutputAndSafetyTokens = 12_000
+export const estimatedContextRequestOverheadTokens = 4_000
+
+export type ContextWindowMessage = {
+  role: 'user' | 'assistant'
+  content: string
+}
+
+export function estimateTextTokens(value: string): number {
+  let asciiCharacters = 0
+  let nonAsciiCharacters = 0
+  for (const character of value) {
+    if (character.codePointAt(0)! <= 0x7f) {
+      asciiCharacters += 1
+    } else {
+      nonAsciiCharacters += 1
+    }
+  }
+  return Math.max(
+    1,
+    Math.ceil(asciiCharacters / 4 + nonAsciiCharacters)
+  )
+}
+
+export function estimateMessagesTokens(
+  messages: readonly ContextWindowMessage[]
+): number {
+  return messages.reduce(
+    (total, message) => total + estimateTextTokens(message.content) + 4,
+    0
+  )
+}
+
+export function estimateContextInputTokens(input: {
+  history: readonly ContextWindowMessage[]
+  prompt: string
+  summaryTokens?: number
+}): number {
+  return (
+    estimateMessagesTokens(input.history) +
+    estimateTextTokens(input.prompt) +
+    (input.summaryTokens ?? 0) +
+    estimatedContextRequestOverheadTokens
+  )
+}
+
+export function getEffectiveContextTriggerTokens(input: {
+  triggerTokens: number
+  contextWindowTokens?: number
+}): number {
+  if (input.contextWindowTokens === undefined) {
+    return input.triggerTokens
+  }
+  return Math.min(
+    input.triggerTokens,
+    Math.max(
+      input.contextWindowTokens,
+      minimumModelContextWindowTokens
+    ) - contextOutputAndSafetyTokens
+  )
+}