WrapModelCall 包装模型调用,在调用前检查是否需要总结
(ctx context.Context, req *ModelRequest, handler ModelCallHandler)
| 134 | |
| 135 | // WrapModelCall 包装模型调用,在调用前检查是否需要总结 |
| 136 | func (m *SummarizationMiddleware) WrapModelCall(ctx context.Context, req *ModelRequest, handler ModelCallHandler) (*ModelResponse, error) { |
| 137 | messages := req.Messages |
| 138 | if len(messages) == 0 { |
| 139 | return handler(ctx, req) |
| 140 | } |
| 141 | |
| 142 | // 计算当前消息的 token 数 |
| 143 | totalTokens := m.tokenCounter(messages) |
| 144 | |
| 145 | sumLog.Debug(ctx, "current tokens", map[string]any{"tokens": totalTokens, "threshold": m.maxTokensBeforeSummary}) |
| 146 | |
| 147 | // 如果未超过阈值,直接返回 |
| 148 | if totalTokens <= m.maxTokensBeforeSummary { |
| 149 | return handler(ctx, req) |
| 150 | } |
| 151 | |
| 152 | sumLog.Info(ctx, "token threshold exceeded, triggering summarization", nil) |
| 153 | |
| 154 | // 如果启用了渐进式压缩,使用新的压缩策略 |
| 155 | if m.enableProgressiveCompact { |
| 156 | compactedMessages, err := m.progressiveCompact(ctx, messages, m.maxTokensBeforeSummary) |
| 157 | if err != nil { |
| 158 | sumLog.Error(ctx, "progressive compact failed, using traditional method", map[string]any{"error": err.Error()}) |
| 159 | // 失败时回退到传统方法 |
| 160 | } else { |
| 161 | req.Messages = compactedMessages |
| 162 | newTokens := m.tokenCounter(compactedMessages) |
| 163 | sumLog.Info(ctx, "progressive compaction complete", map[string]any{ |
| 164 | "before": len(messages), |
| 165 | "after": len(compactedMessages), |
| 166 | "tokens_before": totalTokens, |
| 167 | "tokens_after": newTokens, |
| 168 | }) |
| 169 | return handler(ctx, req) |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | // 分离 system messages 和其他消息 |
| 174 | var systemMessages []types.Message |
| 175 | var regularMessages []types.Message |
| 176 | |
| 177 | for _, msg := range messages { |
| 178 | if msg.Role == types.MessageRoleSystem { |
| 179 | systemMessages = append(systemMessages, msg) |
| 180 | } else { |
| 181 | regularMessages = append(regularMessages, msg) |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | // 如果常规消息少于或等于要保留的数量,不进行总结 |
| 186 | if len(regularMessages) <= m.messagesToKeep { |
| 187 | sumLog.Debug(ctx, "not enough messages to summarize", map[string]any{"have": len(regularMessages), "keep": m.messagesToKeep}) |
| 188 | return handler(ctx, req) |
| 189 | } |
| 190 | |
| 191 | // 计算要总结的消息 |
| 192 | numToSummarize := len(regularMessages) - m.messagesToKeep |
| 193 | messagesToSummarize := regularMessages[:numToSummarize] |