104 lines
4.5 KiB
C#
104 lines
4.5 KiB
C#
using System;
|
||
|
||
namespace AnotherReplayReader.Utils
|
||
{
|
||
/// <summary>
|
||
/// 上下文预算策略：每模型软上限、输出余量、估算安全系数与请求护栏。
|
||
/// </summary>
|
||
internal static class AiContextBudget
|
||
{
|
||
public const int Tier1MBudget = 160_000;
|
||
public const int Tier256KBudget = 100_000;
|
||
public const double EstimatorSafetyFactor = 1.2;
|
||
public const double HardUsageRatio = 0.9;
|
||
|
||
/// <summary>
|
||
/// 获取模型的一次请求总 token 软上限。0 表示该模型不支持长录像（只能短录像单 slice）。
|
||
/// </summary>
|
||
public static int GetContextBudget(AiModel model)
|
||
{
|
||
if (model.ContextBudget is { } explicitBudget && explicitBudget > 0)
|
||
{
|
||
return explicitBudget;
|
||
}
|
||
if (model.ContextLength >= 1_000_000)
|
||
{
|
||
return Tier1MBudget;
|
||
}
|
||
if (model.ContextLength >= 200_000)
|
||
{
|
||
return Tier256KBudget;
|
||
}
|
||
return 0;
|
||
}
|
||
|
||
/// <summary>为输出/推理 tokens 预留的余量。</summary>
|
||
public static int GetOutputHeadroom(AiProvider provider, AiModel model)
|
||
{
|
||
var maxTokens = provider.DefaultMaxTokens;
|
||
return Math.Max(2 * maxTokens, 32_000);
|
||
}
|
||
|
||
/// <summary>带安全系数的 token 估算（对中文偏乐观的 bytes/2.2 估算 × 1.2）。</summary>
|
||
public static int EstimateTokens(string text) =>
|
||
(int)Math.Ceiling(AIAnalyze.EstimateTokenCount(text).EstimatedTokenCount * EstimatorSafetyFactor);
|
||
|
||
/// <summary>请求护栏：只按 prompt 估算检查（保留给旧调用/测试使用）。</summary>
|
||
public static ContextCheckResult CheckPromptUsage(
|
||
int estimatedPromptTokens,
|
||
AiProvider provider,
|
||
AiModel model)
|
||
{
|
||
return CheckUsage(estimatedPromptTokens, provider, model, includeOutputHeadroom: false);
|
||
}
|
||
|
||
/// <summary>请求护栏：把输出/推理余量也算入总用量。</summary>
|
||
public static ContextCheckResult CheckRequestUsage(
|
||
int estimatedPromptTokens,
|
||
AiProvider provider,
|
||
AiModel model)
|
||
{
|
||
return CheckUsage(estimatedPromptTokens, provider, model, includeOutputHeadroom: true);
|
||
}
|
||
|
||
private static ContextCheckResult CheckUsage(
|
||
int estimatedPromptTokens,
|
||
AiProvider provider,
|
||
AiModel model,
|
||
bool includeOutputHeadroom)
|
||
{
|
||
var headroom = includeOutputHeadroom ? GetOutputHeadroom(provider, model) : 0;
|
||
var estimatedTotal = estimatedPromptTokens + headroom;
|
||
if (model.ContextLength > 0)
|
||
{
|
||
var hardLimit = (int)(model.ContextLength * HardUsageRatio);
|
||
if (estimatedTotal > hardLimit)
|
||
{
|
||
return new ContextCheckResult(
|
||
true,
|
||
includeOutputHeadroom
|
||
? $"估算用量（输入 {estimatedPromptTokens:N0} + 输出余量 {headroom:N0} = {estimatedTotal:N0}）超过模型上下文 {model.ContextLength:N0} 的 90%，已拒绝发起请求。请改用更长上下文的模型，或缩短操作记录。"
|
||
: $"估算输入 {estimatedPromptTokens:N0} token 超过模型上下文 {model.ContextLength:N0} 的 90%，已拒绝发起请求。请改用更长上下文的模型，或缩短操作记录。");
|
||
}
|
||
}
|
||
|
||
var budget = GetContextBudget(model);
|
||
if (budget > 0 && estimatedTotal > budget)
|
||
{
|
||
return new ContextCheckResult(
|
||
false,
|
||
includeOutputHeadroom
|
||
? $"估算用量（输入 {estimatedPromptTokens:N0} + 输出余量 {headroom:N0} = {estimatedTotal:N0}）超过上下文预算 {budget:N0}（可在模型设置中调整）。"
|
||
: $"估算输入 {estimatedPromptTokens:N0} token 超过上下文预算 {budget:N0}（可在模型设置中调整），长录像将自动分段，超出部分会被压缩。");
|
||
}
|
||
return ContextCheckResult.Ok;
|
||
}
|
||
}
|
||
|
||
internal sealed record ContextCheckResult(bool Block, string Message)
|
||
{
|
||
public static ContextCheckResult Ok { get; } = new(false, string.Empty);
|
||
public bool IsOk => !Block && string.IsNullOrEmpty(Message);
|
||
}
|
||
}
|