1use anyhow::anyhow;
2use serde::{Deserialize, Serialize};
3use strum::EnumIter;
4
5#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
6#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
7pub enum BedrockModelMode {
8 #[default]
9 Default,
10 Thinking {
11 budget_tokens: Option<u64>,
12 },
13}
14
15#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
16#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, EnumIter)]
17pub enum Model {
18 // Anthropic models (already included)
19 #[default]
20 #[serde(rename = "claude-3-5-sonnet-v2", alias = "claude-3-5-sonnet-latest")]
21 Claude3_5SonnetV2,
22 #[serde(rename = "claude-3-7-sonnet", alias = "claude-3-7-sonnet-latest")]
23 Claude3_7Sonnet,
24 #[serde(
25 rename = "claude-3-7-sonnet-thinking",
26 alias = "claude-3-7-sonnet-thinking-latest"
27 )]
28 Claude3_7SonnetThinking,
29 #[serde(rename = "claude-3-opus", alias = "claude-3-opus-latest")]
30 Claude3Opus,
31 #[serde(rename = "claude-3-sonnet", alias = "claude-3-sonnet-latest")]
32 Claude3Sonnet,
33 #[serde(rename = "claude-3-5-haiku", alias = "claude-3-5-haiku-latest")]
34 Claude3_5Haiku,
35 Claude3_5Sonnet,
36 Claude3Haiku,
37 // Amazon Nova Models
38 AmazonNovaLite,
39 AmazonNovaMicro,
40 AmazonNovaPro,
41 AmazonNovaPremier,
42 // AI21 models
43 AI21J2GrandeInstruct,
44 AI21J2JumboInstruct,
45 AI21J2Mid,
46 AI21J2MidV1,
47 AI21J2Ultra,
48 AI21J2UltraV1_8k,
49 AI21J2UltraV1,
50 AI21JambaInstructV1,
51 AI21Jamba15LargeV1,
52 AI21Jamba15MiniV1,
53 // Cohere models
54 CohereCommandTextV14_4k,
55 CohereCommandRV1,
56 CohereCommandRPlusV1,
57 CohereCommandLightTextV14_4k,
58 // DeepSeek
59 DeepSeekR1,
60 // Meta models
61 MetaLlama38BInstructV1,
62 MetaLlama370BInstructV1,
63 MetaLlama318BInstructV1_128k,
64 MetaLlama318BInstructV1,
65 MetaLlama3170BInstructV1_128k,
66 MetaLlama3170BInstructV1,
67 MetaLlama3211BInstructV1,
68 MetaLlama3290BInstructV1,
69 MetaLlama321BInstructV1,
70 MetaLlama323BInstructV1,
71 // Mistral models
72 MistralMistral7BInstructV0,
73 MistralMixtral8x7BInstructV0,
74 MistralMistralLarge2402V1,
75 MistralMistralSmall2402V1,
76 // Writer models
77 PalmyraWriterX5,
78 PalmyraWriterX4,
79 #[serde(rename = "custom")]
80 Custom {
81 name: String,
82 max_tokens: usize,
83 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
84 display_name: Option<String>,
85 max_output_tokens: Option<u32>,
86 default_temperature: Option<f32>,
87 },
88}
89
90impl Model {
91 pub fn default_fast() -> Self {
92 Self::Claude3_5Haiku
93 }
94
95 pub fn from_id(id: &str) -> anyhow::Result<Self> {
96 if id.starts_with("claude-3-5-sonnet-v2") {
97 Ok(Self::Claude3_5SonnetV2)
98 } else if id.starts_with("claude-3-opus") {
99 Ok(Self::Claude3Opus)
100 } else if id.starts_with("claude-3-sonnet") {
101 Ok(Self::Claude3Sonnet)
102 } else if id.starts_with("claude-3-5-haiku") {
103 Ok(Self::Claude3_5Haiku)
104 } else if id.starts_with("claude-3-7-sonnet") {
105 Ok(Self::Claude3_7Sonnet)
106 } else if id.starts_with("claude-3-7-sonnet-thinking") {
107 Ok(Self::Claude3_7SonnetThinking)
108 } else {
109 Err(anyhow!("invalid model id"))
110 }
111 }
112
113 pub fn id(&self) -> &str {
114 match self {
115 Model::Claude3_5SonnetV2 => "anthropic.claude-3-5-sonnet-20241022-v2:0",
116 Model::Claude3_5Sonnet => "anthropic.claude-3-5-sonnet-20240620-v1:0",
117 Model::Claude3Opus => "anthropic.claude-3-opus-20240229-v1:0",
118 Model::Claude3Sonnet => "anthropic.claude-3-sonnet-20240229-v1:0",
119 Model::Claude3Haiku => "anthropic.claude-3-haiku-20240307-v1:0",
120 Model::Claude3_5Haiku => "anthropic.claude-3-5-haiku-20241022-v1:0",
121 Model::Claude3_7Sonnet | Model::Claude3_7SonnetThinking => {
122 "anthropic.claude-3-7-sonnet-20250219-v1:0"
123 }
124 Model::AmazonNovaLite => "amazon.nova-lite-v1:0",
125 Model::AmazonNovaMicro => "amazon.nova-micro-v1:0",
126 Model::AmazonNovaPro => "amazon.nova-pro-v1:0",
127 Model::AmazonNovaPremier => "amazon.nova-premier-v1:0",
128 Model::DeepSeekR1 => "us.deepseek.r1-v1:0",
129 Model::AI21J2GrandeInstruct => "ai21.j2-grande-instruct",
130 Model::AI21J2JumboInstruct => "ai21.j2-jumbo-instruct",
131 Model::AI21J2Mid => "ai21.j2-mid",
132 Model::AI21J2MidV1 => "ai21.j2-mid-v1",
133 Model::AI21J2Ultra => "ai21.j2-ultra",
134 Model::AI21J2UltraV1_8k => "ai21.j2-ultra-v1:0:8k",
135 Model::AI21J2UltraV1 => "ai21.j2-ultra-v1",
136 Model::AI21JambaInstructV1 => "ai21.jamba-instruct-v1:0",
137 Model::AI21Jamba15LargeV1 => "ai21.jamba-1-5-large-v1:0",
138 Model::AI21Jamba15MiniV1 => "ai21.jamba-1-5-mini-v1:0",
139 Model::CohereCommandTextV14_4k => "cohere.command-text-v14:7:4k",
140 Model::CohereCommandRV1 => "cohere.command-r-v1:0",
141 Model::CohereCommandRPlusV1 => "cohere.command-r-plus-v1:0",
142 Model::CohereCommandLightTextV14_4k => "cohere.command-light-text-v14:7:4k",
143 Model::MetaLlama38BInstructV1 => "meta.llama3-8b-instruct-v1:0",
144 Model::MetaLlama370BInstructV1 => "meta.llama3-70b-instruct-v1:0",
145 Model::MetaLlama318BInstructV1_128k => "meta.llama3-1-8b-instruct-v1:0:128k",
146 Model::MetaLlama318BInstructV1 => "meta.llama3-1-8b-instruct-v1:0",
147 Model::MetaLlama3170BInstructV1_128k => "meta.llama3-1-70b-instruct-v1:0:128k",
148 Model::MetaLlama3170BInstructV1 => "meta.llama3-1-70b-instruct-v1:0",
149 Model::MetaLlama3211BInstructV1 => "meta.llama3-2-11b-instruct-v1:0",
150 Model::MetaLlama3290BInstructV1 => "meta.llama3-2-90b-instruct-v1:0",
151 Model::MetaLlama321BInstructV1 => "meta.llama3-2-1b-instruct-v1:0",
152 Model::MetaLlama323BInstructV1 => "meta.llama3-2-3b-instruct-v1:0",
153 Model::MistralMistral7BInstructV0 => "mistral.mistral-7b-instruct-v0:2",
154 Model::MistralMixtral8x7BInstructV0 => "mistral.mixtral-8x7b-instruct-v0:1",
155 Model::MistralMistralLarge2402V1 => "mistral.mistral-large-2402-v1:0",
156 Model::MistralMistralSmall2402V1 => "mistral.mistral-small-2402-v1:0",
157 Model::PalmyraWriterX4 => "writer.palmyra-x4-v1:0",
158 Model::PalmyraWriterX5 => "writer.palmyra-x5-v1:0",
159 Self::Custom { name, .. } => name,
160 }
161 }
162
163 pub fn display_name(&self) -> &str {
164 match self {
165 Self::Claude3_5SonnetV2 => "Claude 3.5 Sonnet v2",
166 Self::Claude3_5Sonnet => "Claude 3.5 Sonnet",
167 Self::Claude3Opus => "Claude 3 Opus",
168 Self::Claude3Sonnet => "Claude 3 Sonnet",
169 Self::Claude3Haiku => "Claude 3 Haiku",
170 Self::Claude3_5Haiku => "Claude 3.5 Haiku",
171 Self::Claude3_7Sonnet => "Claude 3.7 Sonnet",
172 Self::Claude3_7SonnetThinking => "Claude 3.7 Sonnet Thinking",
173 Self::AmazonNovaLite => "Amazon Nova Lite",
174 Self::AmazonNovaMicro => "Amazon Nova Micro",
175 Self::AmazonNovaPro => "Amazon Nova Pro",
176 Self::AmazonNovaPremier => "Amazon Nova Premier",
177 Self::DeepSeekR1 => "DeepSeek R1",
178 Self::AI21J2GrandeInstruct => "AI21 Jurassic2 Grande Instruct",
179 Self::AI21J2JumboInstruct => "AI21 Jurassic2 Jumbo Instruct",
180 Self::AI21J2Mid => "AI21 Jurassic2 Mid",
181 Self::AI21J2MidV1 => "AI21 Jurassic2 Mid V1",
182 Self::AI21J2Ultra => "AI21 Jurassic2 Ultra",
183 Self::AI21J2UltraV1_8k => "AI21 Jurassic2 Ultra V1 8K",
184 Self::AI21J2UltraV1 => "AI21 Jurassic2 Ultra V1",
185 Self::AI21JambaInstructV1 => "AI21 Jamba Instruct",
186 Self::AI21Jamba15LargeV1 => "AI21 Jamba 1.5 Large",
187 Self::AI21Jamba15MiniV1 => "AI21 Jamba 1.5 Mini",
188 Self::CohereCommandTextV14_4k => "Cohere Command Text V14 4K",
189 Self::CohereCommandRV1 => "Cohere Command R V1",
190 Self::CohereCommandRPlusV1 => "Cohere Command R Plus V1",
191 Self::CohereCommandLightTextV14_4k => "Cohere Command Light Text V14 4K",
192 Self::MetaLlama38BInstructV1 => "Meta Llama 3 8B Instruct V1",
193 Self::MetaLlama370BInstructV1 => "Meta Llama 3 70B Instruct V1",
194 Self::MetaLlama318BInstructV1_128k => "Meta Llama 3 1.8B Instruct V1 128K",
195 Self::MetaLlama318BInstructV1 => "Meta Llama 3 1.8B Instruct V1",
196 Self::MetaLlama3170BInstructV1_128k => "Meta Llama 3 1 70B Instruct V1 128K",
197 Self::MetaLlama3170BInstructV1 => "Meta Llama 3 1 70B Instruct V1",
198 Self::MetaLlama3211BInstructV1 => "Meta Llama 3 2 11B Instruct V1",
199 Self::MetaLlama3290BInstructV1 => "Meta Llama 3 2 90B Instruct V1",
200 Self::MetaLlama321BInstructV1 => "Meta Llama 3 2 1B Instruct V1",
201 Self::MetaLlama323BInstructV1 => "Meta Llama 3 2 3B Instruct V1",
202 Self::MistralMistral7BInstructV0 => "Mistral 7B Instruct V0",
203 Self::MistralMixtral8x7BInstructV0 => "Mistral Mixtral 8x7B Instruct V0",
204 Self::MistralMistralLarge2402V1 => "Mistral Large 2402 V1",
205 Self::MistralMistralSmall2402V1 => "Mistral Small 2402 V1",
206 Self::PalmyraWriterX5 => "Writer Palmyra X5",
207 Self::PalmyraWriterX4 => "Writer Palmyra X4",
208 Self::Custom {
209 display_name, name, ..
210 } => display_name.as_deref().unwrap_or(name),
211 }
212 }
213
214 pub fn max_token_count(&self) -> usize {
215 match self {
216 Self::Claude3_5SonnetV2
217 | Self::Claude3Opus
218 | Self::Claude3Sonnet
219 | Self::Claude3_5Haiku
220 | Self::Claude3_7Sonnet => 200_000,
221 Self::AmazonNovaPremier => 1_000_000,
222 Self::PalmyraWriterX5 => 1_000_000,
223 Self::PalmyraWriterX4 => 128_000,
224 Self::Custom { max_tokens, .. } => *max_tokens,
225 _ => 200_000,
226 }
227 }
228
229 pub fn max_output_tokens(&self) -> u32 {
230 match self {
231 Self::Claude3Opus | Self::Claude3Sonnet | Self::Claude3_5Haiku => 4_096,
232 Self::Claude3_7Sonnet | Self::Claude3_7SonnetThinking => 128_000,
233 Self::Claude3_5SonnetV2 | Self::PalmyraWriterX4 | Self::PalmyraWriterX5 => 8_192,
234 Self::Custom {
235 max_output_tokens, ..
236 } => max_output_tokens.unwrap_or(4_096),
237 _ => 4_096,
238 }
239 }
240
241 pub fn default_temperature(&self) -> f32 {
242 match self {
243 Self::Claude3_5SonnetV2
244 | Self::Claude3Opus
245 | Self::Claude3Sonnet
246 | Self::Claude3_5Haiku
247 | Self::Claude3_7Sonnet => 1.0,
248 Self::Custom {
249 default_temperature,
250 ..
251 } => default_temperature.unwrap_or(1.0),
252 _ => 1.0,
253 }
254 }
255
256 pub fn supports_tool_use(&self) -> bool {
257 match self {
258 // Anthropic Claude 3 models (all support tool use)
259 Self::Claude3Opus
260 | Self::Claude3Sonnet
261 | Self::Claude3_5Sonnet
262 | Self::Claude3_5SonnetV2
263 | Self::Claude3_7Sonnet
264 | Self::Claude3_7SonnetThinking
265 | Self::Claude3_5Haiku => true,
266
267 // Amazon Nova models (all support tool use)
268 Self::AmazonNovaPremier
269 | Self::AmazonNovaPro
270 | Self::AmazonNovaLite
271 | Self::AmazonNovaMicro => true,
272
273 // AI21 Jamba 1.5 models support tool use
274 Self::AI21Jamba15LargeV1 | Self::AI21Jamba15MiniV1 => true,
275
276 // Cohere Command R models support tool use
277 Self::CohereCommandRV1 | Self::CohereCommandRPlusV1 => true,
278
279 // All other models don't support tool use
280 // Including Meta Llama 3.2, AI21 Jurassic, and others
281 _ => false,
282 }
283 }
284
285 pub fn mode(&self) -> BedrockModelMode {
286 match self {
287 Model::Claude3_7SonnetThinking => BedrockModelMode::Thinking {
288 budget_tokens: Some(4096),
289 },
290 _ => BedrockModelMode::Default,
291 }
292 }
293
294 pub fn cross_region_inference_id(&self, region: &str) -> Result<String, anyhow::Error> {
295 let region_group = if region.starts_with("us-gov-") {
296 "us-gov"
297 } else if region.starts_with("us-") {
298 "us"
299 } else if region.starts_with("eu-") {
300 "eu"
301 } else if region.starts_with("ap-") || region == "me-central-1" || region == "me-south-1" {
302 "apac"
303 } else if region.starts_with("ca-") || region.starts_with("sa-") {
304 // Canada and South America regions - default to US profiles
305 "us"
306 } else {
307 // Unknown region
308 return Err(anyhow!("Unsupported Region"));
309 };
310
311 let model_id = self.id();
312
313 match (self, region_group) {
314 // Custom models can't have CRI IDs
315 (Model::Custom { .. }, _) => Ok(self.id().into()),
316
317 // Models with US Gov only
318 (Model::Claude3_5Sonnet, "us-gov") | (Model::Claude3Haiku, "us-gov") => {
319 Ok(format!("{}.{}", region_group, model_id))
320 }
321
322 // Models available only in US
323 (Model::Claude3Opus, "us")
324 | (Model::Claude3_5Sonnet, "us")
325 | (Model::Claude3_7Sonnet, "us")
326 | (Model::Claude3_7SonnetThinking, "us")
327 | (Model::AmazonNovaPremier, "us") => Ok(format!("{}.{}", region_group, model_id)),
328
329 // Models available in US, EU, and APAC
330 (Model::Claude3_5SonnetV2, "us")
331 | (Model::Claude3_5SonnetV2, "apac")
332 | (Model::Claude3_5Sonnet, _)
333 | (Model::Claude3Haiku, _)
334 | (Model::Claude3Sonnet, _)
335 | (Model::AmazonNovaLite, _)
336 | (Model::AmazonNovaMicro, _)
337 | (Model::AmazonNovaPro, _) => Ok(format!("{}.{}", region_group, model_id)),
338
339 // Models with limited EU availability
340 (Model::MetaLlama321BInstructV1, "us")
341 | (Model::MetaLlama321BInstructV1, "eu")
342 | (Model::MetaLlama323BInstructV1, "us")
343 | (Model::MetaLlama323BInstructV1, "eu") => {
344 Ok(format!("{}.{}", region_group, model_id))
345 }
346
347 // US-only models (all remaining Meta models)
348 (Model::MetaLlama38BInstructV1, "us")
349 | (Model::MetaLlama370BInstructV1, "us")
350 | (Model::MetaLlama318BInstructV1, "us")
351 | (Model::MetaLlama318BInstructV1_128k, "us")
352 | (Model::MetaLlama3170BInstructV1, "us")
353 | (Model::MetaLlama3170BInstructV1_128k, "us")
354 | (Model::MetaLlama3211BInstructV1, "us")
355 | (Model::MetaLlama3290BInstructV1, "us") => {
356 Ok(format!("{}.{}", region_group, model_id))
357 }
358
359 // Writer models only available in the US
360 (Model::PalmyraWriterX4, "us") | (Model::PalmyraWriterX5, "us") => {
361 // They have some goofiness
362 Ok(format!("{}.{}", region_group, model_id))
363 }
364
365 // Any other combination is not supported
366 _ => Ok(self.id().into()),
367 }
368 }
369}
370
371#[cfg(test)]
372mod tests {
373 use super::*;
374
375 #[test]
376 fn test_us_region_inference_ids() -> anyhow::Result<()> {
377 // Test US regions
378 assert_eq!(
379 Model::Claude3_5SonnetV2.cross_region_inference_id("us-east-1")?,
380 "us.anthropic.claude-3-5-sonnet-20241022-v2:0"
381 );
382 assert_eq!(
383 Model::Claude3_5SonnetV2.cross_region_inference_id("us-west-2")?,
384 "us.anthropic.claude-3-5-sonnet-20241022-v2:0"
385 );
386 assert_eq!(
387 Model::AmazonNovaPro.cross_region_inference_id("us-east-2")?,
388 "us.amazon.nova-pro-v1:0"
389 );
390 Ok(())
391 }
392
393 #[test]
394 fn test_eu_region_inference_ids() -> anyhow::Result<()> {
395 // Test European regions
396 assert_eq!(
397 Model::Claude3Sonnet.cross_region_inference_id("eu-west-1")?,
398 "eu.anthropic.claude-3-sonnet-20240229-v1:0"
399 );
400 assert_eq!(
401 Model::AmazonNovaMicro.cross_region_inference_id("eu-north-1")?,
402 "eu.amazon.nova-micro-v1:0"
403 );
404 Ok(())
405 }
406
407 #[test]
408 fn test_apac_region_inference_ids() -> anyhow::Result<()> {
409 // Test Asia-Pacific regions
410 assert_eq!(
411 Model::Claude3_5SonnetV2.cross_region_inference_id("ap-northeast-1")?,
412 "apac.anthropic.claude-3-5-sonnet-20241022-v2:0"
413 );
414 assert_eq!(
415 Model::AmazonNovaLite.cross_region_inference_id("ap-south-1")?,
416 "apac.amazon.nova-lite-v1:0"
417 );
418 Ok(())
419 }
420
421 #[test]
422 fn test_gov_region_inference_ids() -> anyhow::Result<()> {
423 // Test Government regions
424 assert_eq!(
425 Model::Claude3_5Sonnet.cross_region_inference_id("us-gov-east-1")?,
426 "us-gov.anthropic.claude-3-5-sonnet-20240620-v1:0"
427 );
428 assert_eq!(
429 Model::Claude3Haiku.cross_region_inference_id("us-gov-west-1")?,
430 "us-gov.anthropic.claude-3-haiku-20240307-v1:0"
431 );
432 Ok(())
433 }
434
435 #[test]
436 fn test_meta_models_inference_ids() -> anyhow::Result<()> {
437 // Test Meta models
438 assert_eq!(
439 Model::MetaLlama370BInstructV1.cross_region_inference_id("us-east-1")?,
440 "us.meta.llama3-70b-instruct-v1:0"
441 );
442 assert_eq!(
443 Model::MetaLlama321BInstructV1.cross_region_inference_id("eu-west-1")?,
444 "eu.meta.llama3-2-1b-instruct-v1:0"
445 );
446 Ok(())
447 }
448
449 #[test]
450 fn test_mistral_models_inference_ids() -> anyhow::Result<()> {
451 // Mistral models don't follow the regional prefix pattern,
452 // so they should return their original IDs
453 assert_eq!(
454 Model::MistralMistralLarge2402V1.cross_region_inference_id("us-east-1")?,
455 "mistral.mistral-large-2402-v1:0"
456 );
457 assert_eq!(
458 Model::MistralMixtral8x7BInstructV0.cross_region_inference_id("eu-west-1")?,
459 "mistral.mixtral-8x7b-instruct-v0:1"
460 );
461 Ok(())
462 }
463
464 #[test]
465 fn test_ai21_models_inference_ids() -> anyhow::Result<()> {
466 // AI21 models don't follow the regional prefix pattern,
467 // so they should return their original IDs
468 assert_eq!(
469 Model::AI21J2UltraV1.cross_region_inference_id("us-east-1")?,
470 "ai21.j2-ultra-v1"
471 );
472 assert_eq!(
473 Model::AI21JambaInstructV1.cross_region_inference_id("eu-west-1")?,
474 "ai21.jamba-instruct-v1:0"
475 );
476 Ok(())
477 }
478
479 #[test]
480 fn test_cohere_models_inference_ids() -> anyhow::Result<()> {
481 // Cohere models don't follow the regional prefix pattern,
482 // so they should return their original IDs
483 assert_eq!(
484 Model::CohereCommandRV1.cross_region_inference_id("us-east-1")?,
485 "cohere.command-r-v1:0"
486 );
487 assert_eq!(
488 Model::CohereCommandTextV14_4k.cross_region_inference_id("ap-southeast-1")?,
489 "cohere.command-text-v14:7:4k"
490 );
491 Ok(())
492 }
493
494 #[test]
495 fn test_custom_model_inference_ids() -> anyhow::Result<()> {
496 // Test custom models
497 let custom_model = Model::Custom {
498 name: "custom.my-model-v1:0".to_string(),
499 max_tokens: 100000,
500 display_name: Some("My Custom Model".to_string()),
501 max_output_tokens: Some(8192),
502 default_temperature: Some(0.7),
503 };
504
505 // Custom model should return its name unchanged
506 assert_eq!(
507 custom_model.cross_region_inference_id("us-east-1")?,
508 "custom.my-model-v1:0"
509 );
510
511 Ok(())
512 }
513}