| 1 | //! Legacy DeepSeek-scoped prompt complexity classifier. |
| 2 | //! |
| 3 | //! This pure scorer is retained for API compatibility, but the consolidated |
| 4 | //! CLI dispatcher must not use it to resolve provider-neutral `model = "auto"`: |
| 5 | //! doing so fabricates DeepSeek model ids for every active provider. The TUI's |
| 6 | //! provider-aware router owns runtime auto selection. Callers may use this |
| 7 | //! helper only when their candidate pair is explicitly the DeepSeek pair: |
| 8 | //! |
| 9 | //! - **`deepseek-v4-pro`** — complex tasks (debugging, refactoring, design, |
| 10 | //! security review, multi-file changes, code generation, …). |
| 11 | //! - **`deepseek-v4-flash`** — simple tasks (lookups, formatting, small edits, |
| 12 | //! translation, Q&A, …). |
| 13 | //! |
| 14 | //! This is a pure rule-based classifier. It lives in the config crate because |
| 15 | //! the resolved model name is a config-level concern; the route resolver never |
| 16 | //! sees the `"auto"` sentinel or the prompt text. |
| 17 | |
| 18 | /// The resolved model name for the pro tier. |
| 19 | pub const PRO_MODEL: &str = "deepseek-v4-pro"; |
| 20 | |
| 21 | /// The resolved model name for the flash tier. |
| 22 | pub const FLASH_MODEL: &str = "deepseek-v4-flash"; |
| 23 | |
| 24 | /// The threshold score above which a task is classified as complex (pro). |
| 25 | /// Score ≥ 2 → pro, else → flash. |
| 26 | const PRO_THRESHOLD: i32 = 2; |
| 27 | |
| 28 | /// Strong indicators of a complex task. Each match adds +3. |
| 29 | const COMPLEX_STRONG: &[&str] = &[ |
| 30 | // Debugging & fixing |
| 31 | "debug", |
| 32 | "bug", |
| 33 | "fix", |
| 34 | "error", |
| 35 | "crash", |
| 36 | "异常", |
| 37 | "错误", |
| 38 | "调试", |
| 39 | "故障", |
| 40 | "排查", |
| 41 | "root cause", |
| 42 | // Architecture & design |
| 43 | "refactor", |
| 44 | "重构", |
| 45 | "architecture", |
| 46 | "架构", |
| 47 | "design pattern", |
| 48 | "系统设计", |
| 49 | "高并发", |
| 50 | "分布式", |
| 51 | "microservice", |
| 52 | // Security |
| 53 | "security", |
| 54 | "安全", |
| 55 | "vulnerability", |
| 56 | "漏洞", |
| 57 | "渗透", |
| 58 | "exploit", |
| 59 | // Code generation |
| 60 | "implement", |
| 61 | "实现", |
| 62 | "generate", |
| 63 | "生成", |
| 64 | "create", |
| 65 | "创建", |
| 66 | "build", |
| 67 | "构建", |
| 68 | "开发", |
| 69 | "prototype", |
| 70 | // Complex analysis |
| 71 | "analyze", |
| 72 | "分析", |
| 73 | "review", |
| 74 | "审查", |
| 75 | "audit", |
| 76 | "审计", |
| 77 | "optimize", |
| 78 | "优化", |
| 79 | "migrate", |
| 80 | "迁移", |
| 81 | // Multi-file / large scale |
| 82 | "multi-file", |
| 83 | "multiple files", |
| 84 | "多个文件", |
| 85 | "整个项目", |
| 86 | "full project", |
| 87 | "重构整个", |
| 88 | "large scale", |
| 89 | // Testing |
| 90 | "unit test", |
| 91 | "integration test", |
| 92 | "e2e test", |
| 93 | "测试用例", |
| 94 | "test suite", |
| 95 | "coverage", |
| 96 | // Complex logic |
| 97 | "algorithm", |
| 98 | "算法", |
| 99 | "状态机", |
| 100 | "state machine", |
| 101 | "concurrent", |
| 102 | "并行", |
| 103 | "异步", |
| 104 | "async", |
| 105 | // Documentation / PRD |
| 106 | "architecture document", |
| 107 | "设计文档", |
| 108 | "技术方案", |
| 109 | "prd", |
| 110 | ]; |
| 111 | |
| 112 | /// Medium-strength indicators. Each match adds +1. |
| 113 | const COMPLEX_MEDIUM: &[&str] = &[ |
| 114 | "change", |
| 115 | "修改", |
| 116 | "update", |
| 117 | "更新", |
| 118 | "add", |
| 119 | "添加", |
| 120 | "新增", |
| 121 | "feature", |
| 122 | "功能", |
| 123 | "improve", |
| 124 | "改进", |
| 125 | "enhance", |
| 126 | "config", |
| 127 | "配置", |
| 128 | "setup", |
| 129 | "设置", |
| 130 | "deploy", |
| 131 | "部署", |
| 132 | "ci/cd", |
| 133 | "pipeline", |
| 134 | "script", |
| 135 | "脚本", |
| 136 | "tool", |
| 137 | "工具", |
| 138 | "api", |
| 139 | "interface", |
| 140 | "接口", |
| 141 | "endpoint", |
| 142 | "database", |
| 143 | "数据库", |
| 144 | "schema", |
| 145 | "query", |
| 146 | "document", |
| 147 | "文档", |
| 148 | "readme", |
| 149 | ]; |
| 150 | |
| 151 | /// Simple-task indicators. Each match subtracts -1. |
| 152 | const SIMPLE: &[&str] = &[ |
| 153 | "find", |
| 154 | "查找", |
| 155 | "search", |
| 156 | "搜索", |
| 157 | "look up", |
| 158 | "查询", |
| 159 | "what is", |
| 160 | "什么是", |
| 161 | "explain", |
| 162 | "解释", |
| 163 | "tell me", |
| 164 | "告诉我", |
| 165 | "how to", |
| 166 | "如何", |
| 167 | "format", |
| 168 | "格式化", |
| 169 | "pretty", |
| 170 | "list", |
| 171 | "列出", |
| 172 | "show", |
| 173 | "显示", |
| 174 | "print", |
| 175 | "rename", |
| 176 | "重命名", |
| 177 | "move", |
| 178 | "移动", |
| 179 | "copy", |
| 180 | "复制", |
| 181 | "delete", |
| 182 | "删除", |
| 183 | "remove", |
| 184 | "typo", |
| 185 | "拼写", |
| 186 | "spelling", |
| 187 | "grammar", |
| 188 | "quick", |
| 189 | "快速", |
| 190 | "simple", |
| 191 | "简单", |
| 192 | "hello world", |
| 193 | "demo", |
| 194 | "example", |
| 195 | "示例", |
| 196 | "translate", |
| 197 | "翻译", |
| 198 | "convert", |
| 199 | "转换", |
| 200 | "short", |
| 201 | "简短", |
| 202 | "brief", |
| 203 | "简要", |
| 204 | ]; |
| 205 | |
| 206 | /// Classify a prompt for the legacy DeepSeek candidate pair. |
| 207 | /// |
| 208 | /// Uses a simple scoring system: |
| 209 | /// - Strong complex keyword: +3 |
| 210 | /// - Medium complex keyword: +1 |
| 211 | /// - Simple keyword: -1 |
| 212 | /// - Prompt length > 500 chars: +2, > 200 chars: +1 |
| 213 | /// - Contains code fence or backtick: +1 |
| 214 | /// - Contains a file path: +1 |
| 215 | /// - Multi-line (> 5 newlines): +1 |
| 216 | /// |
| 217 | /// Total ≥ 2 → `PRO_MODEL`, else → `FLASH_MODEL`. |
| 218 | #[must_use] |
| 219 | pub fn classify(prompt: &str) -> &'static str { |
| 220 | if score(prompt) >= PRO_THRESHOLD { |
| 221 | PRO_MODEL |
| 222 | } else { |
| 223 | FLASH_MODEL |
| 224 | } |
| 225 | } |
| 226 | |
| 227 | /// Compute the raw complexity score for a prompt. |
| 228 | #[must_use] |
| 229 | pub fn score(prompt: &str) -> i32 { |
| 230 | let lower = prompt.to_ascii_lowercase(); |
| 231 | let mut score = 0i32; |
| 232 | |
| 233 | // Strong complex keywords: +3 (first match only to avoid overcounting) |
| 234 | if COMPLEX_STRONG.iter().any(|kw| lower.contains(kw)) { |
| 235 | score += 3; |
| 236 | } |
| 237 | |
| 238 | // Medium complex keywords: +1 each |
| 239 | for kw in COMPLEX_MEDIUM { |
| 240 | if lower.contains(kw) { |
| 241 | score += 1; |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | // Simple keywords: -1 each |
| 246 | for kw in SIMPLE { |
| 247 | if lower.contains(kw) { |
| 248 | score -= 1; |
| 249 | } |
| 250 | } |
| 251 | |
| 252 | // Length factor: long prompts tend to be more complex. |
| 253 | // |
| 254 | // Counted in characters, not bytes, as the doc comment above states. Half |
| 255 | // the keyword lists here are Chinese, so CJK input is a first-class case — |
| 256 | // and every CJK character is three UTF-8 bytes, which made `prompt.len()` |
| 257 | // award the long-prompt bonus at a third of the documented length. A |
| 258 | // 200-character Chinese prompt scored +2 (600 bytes) and classified as |
| 259 | // complex, where the same-length English prompt scored 0. |
| 260 | let len = prompt.chars().count(); |
| 261 | if len > 500 { |
| 262 | score += 2; |
| 263 | } else if len > 200 { |
| 264 | score += 1; |
| 265 | } |
| 266 | |
| 267 | // Code fence or backtick: actual coding task |
| 268 | if prompt.contains("```") || prompt.contains('`') { |
| 269 | score += 1; |
| 270 | } |
| 271 | |
| 272 | // File path pattern: e.g. /path/to/file.rs or C:\path |
| 273 | // Simple heuristic: path-like sequences contain / or \ and . |
| 274 | if (prompt.contains('/') || prompt.contains('\\')) && prompt.contains('.') { |
| 275 | score += 1; |
| 276 | } |
| 277 | |
| 278 | // Multi-line: more lines = more context |
| 279 | if prompt.chars().filter(|&c| c == '\n').count() > 5 { |
| 280 | score += 1; |
| 281 | } |
| 282 | |
| 283 | score |
| 284 | } |
| 285 | |
| 286 | #[cfg(test)] |
| 287 | mod tests { |
| 288 | use super::*; |
| 289 | |
| 290 | #[test] |
| 291 | fn test_debug_task_uses_pro() { |
| 292 | assert_eq!(classify("帮我调试这个bug，程序崩溃了"), PRO_MODEL); |
| 293 | } |
| 294 | |
| 295 | #[test] |
| 296 | fn test_refactor_task_uses_pro() { |
| 297 | assert_eq!( |
| 298 | classify("refactor the user module with a new architecture"), |
| 299 | PRO_MODEL |
| 300 | ); |
| 301 | } |
| 302 | |
| 303 | #[test] |
| 304 | fn test_security_review_uses_pro() { |
| 305 | assert_eq!( |
| 306 | classify("review this code for security vulnerabilities"), |
| 307 | PRO_MODEL |
| 308 | ); |
| 309 | } |
| 310 | |
| 311 | #[test] |
| 312 | fn test_simple_lookup_uses_flash() { |
| 313 | assert_eq!(classify("查找昨天的日志文件"), FLASH_MODEL); |
| 314 | } |
| 315 | |
| 316 | #[test] |
| 317 | fn test_translation_uses_flash() { |
| 318 | assert_eq!(classify("translate this to Chinese"), FLASH_MODEL); |
| 319 | } |
| 320 | |
| 321 | #[test] |
| 322 | fn test_formatting_uses_flash() { |
| 323 | assert_eq!(classify("format this code"), FLASH_MODEL); |
| 324 | } |
| 325 | |
| 326 | #[test] |
| 327 | fn test_long_prompt_gets_bonus() { |
| 328 | let long = "a".repeat(300); |
| 329 | // No keywords, long prompt gives +1, total = 1 < 2 → flash |
| 330 | assert_eq!(classify(&long), FLASH_MODEL); |
| 331 | } |
| 332 | |
| 333 | #[test] |
| 334 | fn test_very_long_prompt_gets_more_bonus() { |
| 335 | let long = "a".repeat(600); |
| 336 | // No keywords, very long prompt gives +2, total = 2 → pro |
| 337 | assert_eq!(classify(&long), PRO_MODEL); |
| 338 | } |
| 339 | |
| 340 | #[test] |
| 341 | fn test_code_block_gets_bonus() { |
| 342 | // Code block without keywords, +1, total = 1 < 2 → flash |
| 343 | assert_eq!(classify("```\nhello\n```"), FLASH_MODEL); |
| 344 | } |
| 345 | |
| 346 | #[test] |
| 347 | fn test_mixed_keywords_pro_wins() { |
| 348 | // "refactor" is strong (+3), "explain" is simple (-1), total = 2 → pro |
| 349 | assert_eq!(classify("refactor and explain the code"), PRO_MODEL); |
| 350 | } |
| 351 | |
| 352 | #[test] |
| 353 | fn test_implement_task_uses_pro() { |
| 354 | assert_eq!( |
| 355 | classify("implement a new feature for the user module"), |
| 356 | PRO_MODEL |
| 357 | ); |
| 358 | } |
| 359 | |
| 360 | #[test] |
| 361 | fn test_quick_question_uses_flash() { |
| 362 | assert_eq!(classify("what is the capital of France?"), FLASH_MODEL); |
| 363 | } |
| 364 | |
| 365 | #[test] |
| 366 | fn length_bonus_counts_characters_not_utf8_bytes() { |
| 367 | // Keyword-free prompts of identical *length* must score identically |
| 368 | // regardless of script. "啊" is three UTF-8 bytes, so a byte-counted |
| 369 | // length factor gave the Chinese prompt a bonus the English one did |
| 370 | // not earn — and at 200 characters it flipped the classification. |
| 371 | let english = "a".repeat(200); |
| 372 | let chinese = "啊".repeat(200); |
| 373 | assert_eq!(score(&chinese), score(&english)); |
| 374 | assert_eq!(classify(&chinese), FLASH_MODEL); |
| 375 | |
| 376 | let english_long = "a".repeat(600); |
| 377 | let chinese_long = "啊".repeat(600); |
| 378 | assert_eq!(score(&chinese_long), score(&english_long)); |
| 379 | assert_eq!(classify(&chinese_long), PRO_MODEL); |
| 380 | } |
| 381 | |
| 382 | #[test] |
| 383 | fn test_score_never_negative() { |
| 384 | // Even for very simple queries, score should be predictable |
| 385 | let s = score("hello world"); |
| 386 | assert!(s >= -10); // sanity check |
| 387 | } |
| 388 | } |
| 389 |