1use serde::{Deserialize, Serialize};
2use strum::EnumIter;
3
4#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
5#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
6pub enum BedrockModelMode {
7 #[default]
8 Default,
9 Thinking {
10 budget_tokens: Option<u64>,
11 },
12}
13
14#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
15#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
16pub struct BedrockModelCacheConfiguration {
17 pub max_cache_anchors: usize,
18 pub min_total_token: u64,
19}
20
21#[cfg_attr(feature = "schemars", derive(schemars::JsonSchema))]
22#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, EnumIter)]
23pub enum Model {
24 // Anthropic models (already included)
25 #[default]
26 #[serde(rename = "claude-sonnet-4", alias = "claude-sonnet-4-latest")]
27 ClaudeSonnet4,
28 #[serde(
29 rename = "claude-sonnet-4-thinking",
30 alias = "claude-sonnet-4-thinking-latest"
31 )]
32 ClaudeSonnet4Thinking,
33 #[serde(rename = "claude-opus-4", alias = "claude-opus-4-latest")]
34 ClaudeOpus4,
35 #[serde(rename = "claude-opus-4-1", alias = "claude-opus-4-1-latest")]
36 ClaudeOpus4_1,
37 #[serde(
38 rename = "claude-opus-4-thinking",
39 alias = "claude-opus-4-thinking-latest"
40 )]
41 ClaudeOpus4Thinking,
42 #[serde(
43 rename = "claude-opus-4-1-thinking",
44 alias = "claude-opus-4-1-thinking-latest"
45 )]
46 ClaudeOpus4_1Thinking,
47 #[serde(rename = "claude-3-5-sonnet-v2", alias = "claude-3-5-sonnet-latest")]
48 Claude3_5SonnetV2,
49 #[serde(rename = "claude-3-7-sonnet", alias = "claude-3-7-sonnet-latest")]
50 Claude3_7Sonnet,
51 #[serde(
52 rename = "claude-3-7-sonnet-thinking",
53 alias = "claude-3-7-sonnet-thinking-latest"
54 )]
55 Claude3_7SonnetThinking,
56 #[serde(rename = "claude-3-opus", alias = "claude-3-opus-latest")]
57 Claude3Opus,
58 #[serde(rename = "claude-3-sonnet", alias = "claude-3-sonnet-latest")]
59 Claude3Sonnet,
60 #[serde(rename = "claude-3-5-haiku", alias = "claude-3-5-haiku-latest")]
61 Claude3_5Haiku,
62 Claude3_5Sonnet,
63 Claude3Haiku,
64 // Amazon Nova Models
65 AmazonNovaLite,
66 AmazonNovaMicro,
67 AmazonNovaPro,
68 AmazonNovaPremier,
69 // AI21 models
70 AI21J2GrandeInstruct,
71 AI21J2JumboInstruct,
72 AI21J2Mid,
73 AI21J2MidV1,
74 AI21J2Ultra,
75 AI21J2UltraV1_8k,
76 AI21J2UltraV1,
77 AI21JambaInstructV1,
78 AI21Jamba15LargeV1,
79 AI21Jamba15MiniV1,
80 // Cohere models
81 CohereCommandTextV14_4k,
82 CohereCommandRV1,
83 CohereCommandRPlusV1,
84 CohereCommandLightTextV14_4k,
85 // DeepSeek
86 DeepSeekR1,
87 // Meta models
88 MetaLlama38BInstructV1,
89 MetaLlama370BInstructV1,
90 MetaLlama318BInstructV1_128k,
91 MetaLlama318BInstructV1,
92 MetaLlama3170BInstructV1_128k,
93 MetaLlama3170BInstructV1,
94 MetaLlama31405BInstructV1,
95 MetaLlama321BInstructV1,
96 MetaLlama323BInstructV1,
97 MetaLlama3211BInstructV1,
98 MetaLlama3290BInstructV1,
99 MetaLlama3370BInstructV1,
100 #[allow(non_camel_case_types)]
101 MetaLlama4Scout17BInstructV1,
102 #[allow(non_camel_case_types)]
103 MetaLlama4Maverick17BInstructV1,
104 // Mistral models
105 MistralMistral7BInstructV0,
106 MistralMixtral8x7BInstructV0,
107 MistralMistralLarge2402V1,
108 MistralMistralSmall2402V1,
109 MistralPixtralLarge2502V1,
110 // Writer models
111 PalmyraWriterX5,
112 PalmyraWriterX4,
113 #[serde(rename = "custom")]
114 Custom {
115 name: String,
116 max_tokens: u64,
117 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
118 display_name: Option<String>,
119 max_output_tokens: Option<u64>,
120 default_temperature: Option<f32>,
121 cache_configuration: Option<BedrockModelCacheConfiguration>,
122 },
123}
124
125impl Model {
126 pub fn default_fast(region: &str) -> Self {
127 if region.starts_with("us-") {
128 Self::Claude3_5Haiku
129 } else {
130 Self::Claude3Haiku
131 }
132 }
133
134 pub fn from_id(id: &str) -> anyhow::Result<Self> {
135 if id.starts_with("claude-3-5-sonnet-v2") {
136 Ok(Self::Claude3_5SonnetV2)
137 } else if id.starts_with("claude-3-opus") {
138 Ok(Self::Claude3Opus)
139 } else if id.starts_with("claude-3-sonnet") {
140 Ok(Self::Claude3Sonnet)
141 } else if id.starts_with("claude-3-5-haiku") {
142 Ok(Self::Claude3_5Haiku)
143 } else if id.starts_with("claude-3-7-sonnet") {
144 Ok(Self::Claude3_7Sonnet)
145 } else if id.starts_with("claude-3-7-sonnet-thinking") {
146 Ok(Self::Claude3_7SonnetThinking)
147 } else {
148 anyhow::bail!("invalid model id {id}");
149 }
150 }
151
152 pub fn id(&self) -> &str {
153 match self {
154 Model::ClaudeSonnet4 => "claude-sonnet-4",
155 Model::ClaudeSonnet4Thinking => "claude-sonnet-4-thinking",
156 Model::ClaudeOpus4 => "claude-opus-4",
157 Model::ClaudeOpus4_1 => "claude-opus-4-1",
158 Model::ClaudeOpus4Thinking => "claude-opus-4-thinking",
159 Model::ClaudeOpus4_1Thinking => "claude-opus-4-1-thinking",
160 Model::Claude3_5SonnetV2 => "claude-3-5-sonnet-v2",
161 Model::Claude3_5Sonnet => "claude-3-5-sonnet",
162 Model::Claude3Opus => "claude-3-opus",
163 Model::Claude3Sonnet => "claude-3-sonnet",
164 Model::Claude3Haiku => "claude-3-haiku",
165 Model::Claude3_5Haiku => "claude-3-5-haiku",
166 Model::Claude3_7Sonnet => "claude-3-7-sonnet",
167 Model::Claude3_7SonnetThinking => "claude-3-7-sonnet-thinking",
168 Model::AmazonNovaLite => "amazon-nova-lite",
169 Model::AmazonNovaMicro => "amazon-nova-micro",
170 Model::AmazonNovaPro => "amazon-nova-pro",
171 Model::AmazonNovaPremier => "amazon-nova-premier",
172 Model::DeepSeekR1 => "deepseek-r1",
173 Model::AI21J2GrandeInstruct => "ai21-j2-grande-instruct",
174 Model::AI21J2JumboInstruct => "ai21-j2-jumbo-instruct",
175 Model::AI21J2Mid => "ai21-j2-mid",
176 Model::AI21J2MidV1 => "ai21-j2-mid-v1",
177 Model::AI21J2Ultra => "ai21-j2-ultra",
178 Model::AI21J2UltraV1_8k => "ai21-j2-ultra-v1-8k",
179 Model::AI21J2UltraV1 => "ai21-j2-ultra-v1",
180 Model::AI21JambaInstructV1 => "ai21-jamba-instruct-v1",
181 Model::AI21Jamba15LargeV1 => "ai21-jamba-1-5-large-v1",
182 Model::AI21Jamba15MiniV1 => "ai21-jamba-1-5-mini-v1",
183 Model::CohereCommandTextV14_4k => "cohere-command-text-v14-4k",
184 Model::CohereCommandRV1 => "cohere-command-r-v1",
185 Model::CohereCommandRPlusV1 => "cohere-command-r-plus-v1",
186 Model::CohereCommandLightTextV14_4k => "cohere-command-light-text-v14-4k",
187 Model::MetaLlama38BInstructV1 => "meta-llama3-8b-instruct-v1",
188 Model::MetaLlama370BInstructV1 => "meta-llama3-70b-instruct-v1",
189 Model::MetaLlama318BInstructV1_128k => "meta-llama3-1-8b-instruct-v1-128k",
190 Model::MetaLlama318BInstructV1 => "meta-llama3-1-8b-instruct-v1",
191 Model::MetaLlama3170BInstructV1_128k => "meta-llama3-1-70b-instruct-v1-128k",
192 Model::MetaLlama3170BInstructV1 => "meta-llama3-1-70b-instruct-v1",
193 Model::MetaLlama31405BInstructV1 => "meta-llama3-1-405b-instruct-v1",
194 Model::MetaLlama321BInstructV1 => "meta-llama3-2-1b-instruct-v1",
195 Model::MetaLlama323BInstructV1 => "meta-llama3-2-3b-instruct-v1",
196 Model::MetaLlama3211BInstructV1 => "meta-llama3-2-11b-instruct-v1",
197 Model::MetaLlama3290BInstructV1 => "meta-llama3-2-90b-instruct-v1",
198 Model::MetaLlama3370BInstructV1 => "meta-llama3-3-70b-instruct-v1",
199 Model::MetaLlama4Scout17BInstructV1 => "meta-llama4-scout-17b-instruct-v1",
200 Model::MetaLlama4Maverick17BInstructV1 => "meta-llama4-maverick-17b-instruct-v1",
201 Model::MistralMistral7BInstructV0 => "mistral-7b-instruct-v0",
202 Model::MistralMixtral8x7BInstructV0 => "mistral-mixtral-8x7b-instruct-v0",
203 Model::MistralMistralLarge2402V1 => "mistral-large-2402-v1",
204 Model::MistralMistralSmall2402V1 => "mistral-small-2402-v1",
205 Model::MistralPixtralLarge2502V1 => "mistral-pixtral-large-2502-v1",
206 Model::PalmyraWriterX4 => "palmyra-writer-x4",
207 Model::PalmyraWriterX5 => "palmyra-writer-x5",
208 Self::Custom { name, .. } => name,
209 }
210 }
211
212 pub fn request_id(&self) -> &str {
213 match self {
214 Model::ClaudeSonnet4 | Model::ClaudeSonnet4Thinking => {
215 "anthropic.claude-sonnet-4-20250514-v1:0"
216 }
217 Model::ClaudeOpus4 | Model::ClaudeOpus4Thinking => {
218 "anthropic.claude-opus-4-20250514-v1:0"
219 }
220 Model::ClaudeOpus4_1 | Model::ClaudeOpus4_1Thinking => {
221 "anthropic.claude-opus-4-1-20250805-v1:0"
222 }
223 Model::Claude3_5SonnetV2 => "anthropic.claude-3-5-sonnet-20241022-v2:0",
224 Model::Claude3_5Sonnet => "anthropic.claude-3-5-sonnet-20240620-v1:0",
225 Model::Claude3Opus => "anthropic.claude-3-opus-20240229-v1:0",
226 Model::Claude3Sonnet => "anthropic.claude-3-sonnet-20240229-v1:0",
227 Model::Claude3Haiku => "anthropic.claude-3-haiku-20240307-v1:0",
228 Model::Claude3_5Haiku => "anthropic.claude-3-5-haiku-20241022-v1:0",
229 Model::Claude3_7Sonnet | Model::Claude3_7SonnetThinking => {
230 "anthropic.claude-3-7-sonnet-20250219-v1:0"
231 }
232 Model::AmazonNovaLite => "amazon.nova-lite-v1:0",
233 Model::AmazonNovaMicro => "amazon.nova-micro-v1:0",
234 Model::AmazonNovaPro => "amazon.nova-pro-v1:0",
235 Model::AmazonNovaPremier => "amazon.nova-premier-v1:0",
236 Model::DeepSeekR1 => "deepseek.r1-v1:0",
237 Model::AI21J2GrandeInstruct => "ai21.j2-grande-instruct",
238 Model::AI21J2JumboInstruct => "ai21.j2-jumbo-instruct",
239 Model::AI21J2Mid => "ai21.j2-mid",
240 Model::AI21J2MidV1 => "ai21.j2-mid-v1",
241 Model::AI21J2Ultra => "ai21.j2-ultra",
242 Model::AI21J2UltraV1_8k => "ai21.j2-ultra-v1:0:8k",
243 Model::AI21J2UltraV1 => "ai21.j2-ultra-v1",
244 Model::AI21JambaInstructV1 => "ai21.jamba-instruct-v1:0",
245 Model::AI21Jamba15LargeV1 => "ai21.jamba-1-5-large-v1:0",
246 Model::AI21Jamba15MiniV1 => "ai21.jamba-1-5-mini-v1:0",
247 Model::CohereCommandTextV14_4k => "cohere.command-text-v14:7:4k",
248 Model::CohereCommandRV1 => "cohere.command-r-v1:0",
249 Model::CohereCommandRPlusV1 => "cohere.command-r-plus-v1:0",
250 Model::CohereCommandLightTextV14_4k => "cohere.command-light-text-v14:7:4k",
251 Model::MetaLlama38BInstructV1 => "meta.llama3-8b-instruct-v1:0",
252 Model::MetaLlama370BInstructV1 => "meta.llama3-70b-instruct-v1:0",
253 Model::MetaLlama318BInstructV1_128k => "meta.llama3-1-8b-instruct-v1:0",
254 Model::MetaLlama318BInstructV1 => "meta.llama3-1-8b-instruct-v1:0",
255 Model::MetaLlama3170BInstructV1_128k => "meta.llama3-1-70b-instruct-v1:0",
256 Model::MetaLlama3170BInstructV1 => "meta.llama3-1-70b-instruct-v1:0",
257 Model::MetaLlama31405BInstructV1 => "meta.llama3-1-405b-instruct-v1:0",
258 Model::MetaLlama3211BInstructV1 => "meta.llama3-2-11b-instruct-v1:0",
259 Model::MetaLlama3290BInstructV1 => "meta.llama3-2-90b-instruct-v1:0",
260 Model::MetaLlama321BInstructV1 => "meta.llama3-2-1b-instruct-v1:0",
261 Model::MetaLlama323BInstructV1 => "meta.llama3-2-3b-instruct-v1:0",
262 Model::MetaLlama3370BInstructV1 => "meta.llama3-3-70b-instruct-v1:0",
263 Model::MetaLlama4Scout17BInstructV1 => "meta.llama4-scout-17b-instruct-v1:0",
264 Model::MetaLlama4Maverick17BInstructV1 => "meta.llama4-maverick-17b-instruct-v1:0",
265 Model::MistralMistral7BInstructV0 => "mistral.mistral-7b-instruct-v0:2",
266 Model::MistralMixtral8x7BInstructV0 => "mistral.mixtral-8x7b-instruct-v0:1",
267 Model::MistralMistralLarge2402V1 => "mistral.mistral-large-2402-v1:0",
268 Model::MistralMistralSmall2402V1 => "mistral.mistral-small-2402-v1:0",
269 Model::MistralPixtralLarge2502V1 => "mistral.pixtral-large-2502-v1:0",
270 Model::PalmyraWriterX4 => "writer.palmyra-x4-v1:0",
271 Model::PalmyraWriterX5 => "writer.palmyra-x5-v1:0",
272 Self::Custom { name, .. } => name,
273 }
274 }
275
276 pub fn display_name(&self) -> &str {
277 match self {
278 Self::ClaudeSonnet4 => "Claude Sonnet 4",
279 Self::ClaudeSonnet4Thinking => "Claude Sonnet 4 Thinking",
280 Self::ClaudeOpus4 => "Claude Opus 4",
281 Self::ClaudeOpus4_1 => "Claude Opus 4.1",
282 Self::ClaudeOpus4Thinking => "Claude Opus 4 Thinking",
283 Self::ClaudeOpus4_1Thinking => "Claude Opus 4.1 Thinking",
284 Self::Claude3_5SonnetV2 => "Claude 3.5 Sonnet v2",
285 Self::Claude3_5Sonnet => "Claude 3.5 Sonnet",
286 Self::Claude3Opus => "Claude 3 Opus",
287 Self::Claude3Sonnet => "Claude 3 Sonnet",
288 Self::Claude3Haiku => "Claude 3 Haiku",
289 Self::Claude3_5Haiku => "Claude 3.5 Haiku",
290 Self::Claude3_7Sonnet => "Claude 3.7 Sonnet",
291 Self::Claude3_7SonnetThinking => "Claude 3.7 Sonnet Thinking",
292 Self::AmazonNovaLite => "Amazon Nova Lite",
293 Self::AmazonNovaMicro => "Amazon Nova Micro",
294 Self::AmazonNovaPro => "Amazon Nova Pro",
295 Self::AmazonNovaPremier => "Amazon Nova Premier",
296 Self::DeepSeekR1 => "DeepSeek R1",
297 Self::AI21J2GrandeInstruct => "AI21 Jurassic2 Grande Instruct",
298 Self::AI21J2JumboInstruct => "AI21 Jurassic2 Jumbo Instruct",
299 Self::AI21J2Mid => "AI21 Jurassic2 Mid",
300 Self::AI21J2MidV1 => "AI21 Jurassic2 Mid V1",
301 Self::AI21J2Ultra => "AI21 Jurassic2 Ultra",
302 Self::AI21J2UltraV1_8k => "AI21 Jurassic2 Ultra V1 8K",
303 Self::AI21J2UltraV1 => "AI21 Jurassic2 Ultra V1",
304 Self::AI21JambaInstructV1 => "AI21 Jamba Instruct",
305 Self::AI21Jamba15LargeV1 => "AI21 Jamba 1.5 Large",
306 Self::AI21Jamba15MiniV1 => "AI21 Jamba 1.5 Mini",
307 Self::CohereCommandTextV14_4k => "Cohere Command Text V14 4K",
308 Self::CohereCommandRV1 => "Cohere Command R V1",
309 Self::CohereCommandRPlusV1 => "Cohere Command R Plus V1",
310 Self::CohereCommandLightTextV14_4k => "Cohere Command Light Text V14 4K",
311 Self::MetaLlama38BInstructV1 => "Meta Llama 3 8B Instruct",
312 Self::MetaLlama370BInstructV1 => "Meta Llama 3 70B Instruct",
313 Self::MetaLlama318BInstructV1_128k => "Meta Llama 3.1 8B Instruct 128K",
314 Self::MetaLlama318BInstructV1 => "Meta Llama 3.1 8B Instruct",
315 Self::MetaLlama3170BInstructV1_128k => "Meta Llama 3.1 70B Instruct 128K",
316 Self::MetaLlama3170BInstructV1 => "Meta Llama 3.1 70B Instruct",
317 Self::MetaLlama31405BInstructV1 => "Meta Llama 3.1 405B Instruct",
318 Self::MetaLlama3211BInstructV1 => "Meta Llama 3.2 11B Instruct",
319 Self::MetaLlama3290BInstructV1 => "Meta Llama 3.2 90B Instruct",
320 Self::MetaLlama321BInstructV1 => "Meta Llama 3.2 1B Instruct",
321 Self::MetaLlama323BInstructV1 => "Meta Llama 3.2 3B Instruct",
322 Self::MetaLlama3370BInstructV1 => "Meta Llama 3.3 70B Instruct",
323 Self::MetaLlama4Scout17BInstructV1 => "Meta Llama 4 Scout 17B Instruct",
324 Self::MetaLlama4Maverick17BInstructV1 => "Meta Llama 4 Maverick 17B Instruct",
325 Self::MistralMistral7BInstructV0 => "Mistral 7B Instruct V0",
326 Self::MistralMixtral8x7BInstructV0 => "Mistral Mixtral 8x7B Instruct V0",
327 Self::MistralMistralLarge2402V1 => "Mistral Large 2402 V1",
328 Self::MistralMistralSmall2402V1 => "Mistral Small 2402 V1",
329 Self::MistralPixtralLarge2502V1 => "Pixtral Large 25.02 V1",
330 Self::PalmyraWriterX5 => "Writer Palmyra X5",
331 Self::PalmyraWriterX4 => "Writer Palmyra X4",
332 Self::Custom {
333 display_name, name, ..
334 } => display_name.as_deref().unwrap_or(name),
335 }
336 }
337
338 pub fn max_token_count(&self) -> u64 {
339 match self {
340 Self::Claude3_5SonnetV2
341 | Self::Claude3Opus
342 | Self::Claude3Sonnet
343 | Self::Claude3_5Haiku
344 | Self::Claude3_7Sonnet
345 | Self::ClaudeSonnet4
346 | Self::ClaudeOpus4
347 | Self::ClaudeOpus4_1
348 | Self::ClaudeSonnet4Thinking
349 | Self::ClaudeOpus4Thinking
350 | Self::ClaudeOpus4_1Thinking => 200_000,
351 Self::AmazonNovaPremier => 1_000_000,
352 Self::PalmyraWriterX5 => 1_000_000,
353 Self::PalmyraWriterX4 => 128_000,
354 Self::Custom { max_tokens, .. } => *max_tokens,
355 _ => 128_000,
356 }
357 }
358
359 pub fn max_output_tokens(&self) -> u64 {
360 match self {
361 Self::Claude3Opus | Self::Claude3Sonnet | Self::Claude3_5Haiku => 4_096,
362 Self::Claude3_7Sonnet | Self::Claude3_7SonnetThinking => 128_000,
363 Self::ClaudeSonnet4 | Self::ClaudeSonnet4Thinking => 64_000,
364 Self::ClaudeOpus4
365 | Self::ClaudeOpus4Thinking
366 | Self::ClaudeOpus4_1
367 | Self::ClaudeOpus4_1Thinking => 32_000,
368 Self::Claude3_5SonnetV2 | Self::PalmyraWriterX4 | Self::PalmyraWriterX5 => 8_192,
369 Self::Custom {
370 max_output_tokens, ..
371 } => max_output_tokens.unwrap_or(4_096),
372 _ => 4_096,
373 }
374 }
375
376 pub fn default_temperature(&self) -> f32 {
377 match self {
378 Self::Claude3_5SonnetV2
379 | Self::Claude3Opus
380 | Self::Claude3Sonnet
381 | Self::Claude3_5Haiku
382 | Self::Claude3_7Sonnet
383 | Self::ClaudeOpus4
384 | Self::ClaudeOpus4Thinking
385 | Self::ClaudeOpus4_1
386 | Self::ClaudeOpus4_1Thinking
387 | Self::ClaudeSonnet4
388 | Self::ClaudeSonnet4Thinking => 1.0,
389 Self::Custom {
390 default_temperature,
391 ..
392 } => default_temperature.unwrap_or(1.0),
393 _ => 1.0,
394 }
395 }
396
397 pub fn supports_tool_use(&self) -> bool {
398 match self {
399 // Anthropic Claude 3 models (all support tool use)
400 Self::Claude3Opus
401 | Self::Claude3Sonnet
402 | Self::Claude3_5Sonnet
403 | Self::Claude3_5SonnetV2
404 | Self::Claude3_7Sonnet
405 | Self::Claude3_7SonnetThinking
406 | Self::ClaudeOpus4
407 | Self::ClaudeOpus4Thinking
408 | Self::ClaudeOpus4_1
409 | Self::ClaudeOpus4_1Thinking
410 | Self::ClaudeSonnet4
411 | Self::ClaudeSonnet4Thinking
412 | Self::Claude3_5Haiku => true,
413
414 // Amazon Nova models (all support tool use)
415 Self::AmazonNovaPremier
416 | Self::AmazonNovaPro
417 | Self::AmazonNovaLite
418 | Self::AmazonNovaMicro => true,
419
420 // AI21 Jamba 1.5 models support tool use
421 Self::AI21Jamba15LargeV1 | Self::AI21Jamba15MiniV1 => true,
422
423 // Cohere Command R models support tool use
424 Self::CohereCommandRV1 | Self::CohereCommandRPlusV1 => true,
425
426 // All other models don't support tool use
427 // Including Meta Llama 3.2, AI21 Jurassic, and others
428 _ => false,
429 }
430 }
431
432 pub fn supports_caching(&self) -> bool {
433 match self {
434 // Only Claude models on Bedrock support caching
435 // Nova models support only text caching
436 // https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html#prompt-caching-models
437 Self::Claude3_5Haiku
438 | Self::Claude3_7Sonnet
439 | Self::Claude3_7SonnetThinking
440 | Self::ClaudeSonnet4
441 | Self::ClaudeSonnet4Thinking
442 | Self::ClaudeOpus4
443 | Self::ClaudeOpus4Thinking
444 | Self::ClaudeOpus4_1
445 | Self::ClaudeOpus4_1Thinking => true,
446
447 // Custom models - check if they have cache configuration
448 Self::Custom {
449 cache_configuration,
450 ..
451 } => cache_configuration.is_some(),
452
453 // All other models don't support caching
454 _ => false,
455 }
456 }
457
458 pub fn cache_configuration(&self) -> Option<BedrockModelCacheConfiguration> {
459 match self {
460 Self::Claude3_7Sonnet
461 | Self::Claude3_7SonnetThinking
462 | Self::ClaudeSonnet4
463 | Self::ClaudeSonnet4Thinking
464 | Self::ClaudeOpus4
465 | Self::ClaudeOpus4Thinking
466 | Self::ClaudeOpus4_1
467 | Self::ClaudeOpus4_1Thinking => Some(BedrockModelCacheConfiguration {
468 max_cache_anchors: 4,
469 min_total_token: 1024,
470 }),
471
472 Self::Claude3_5Haiku => Some(BedrockModelCacheConfiguration {
473 max_cache_anchors: 4,
474 min_total_token: 2048,
475 }),
476
477 Self::Custom {
478 cache_configuration,
479 ..
480 } => cache_configuration.clone(),
481
482 _ => None,
483 }
484 }
485
486 pub fn mode(&self) -> BedrockModelMode {
487 match self {
488 Model::Claude3_7SonnetThinking => BedrockModelMode::Thinking {
489 budget_tokens: Some(4096),
490 },
491 Model::ClaudeSonnet4Thinking => BedrockModelMode::Thinking {
492 budget_tokens: Some(4096),
493 },
494 Model::ClaudeOpus4Thinking | Model::ClaudeOpus4_1Thinking => {
495 BedrockModelMode::Thinking {
496 budget_tokens: Some(4096),
497 }
498 }
499 _ => BedrockModelMode::Default,
500 }
501 }
502
503 pub fn cross_region_inference_id(&self, region: &str) -> anyhow::Result<String> {
504 let region_group = if region.starts_with("us-gov-") {
505 "us-gov"
506 } else if region.starts_with("us-") {
507 "us"
508 } else if region.starts_with("eu-") {
509 "eu"
510 } else if region.starts_with("ap-") || region == "me-central-1" || region == "me-south-1" {
511 "apac"
512 } else if region.starts_with("ca-") || region.starts_with("sa-") {
513 // Canada and South America regions - default to US profiles
514 "us"
515 } else {
516 anyhow::bail!("Unsupported Region {region}");
517 };
518
519 let model_id = self.request_id();
520
521 match (self, region_group) {
522 // Custom models can't have CRI IDs
523 (Model::Custom { .. }, _) => Ok(self.request_id().into()),
524
525 // Models with US Gov only
526 (Model::Claude3_5Sonnet, "us-gov") | (Model::Claude3Haiku, "us-gov") => {
527 Ok(format!("{}.{}", region_group, model_id))
528 }
529
530 // Available everywhere
531 (Model::AmazonNovaLite | Model::AmazonNovaMicro | Model::AmazonNovaPro, _) => {
532 Ok(format!("{}.{}", region_group, model_id))
533 }
534
535 // Models in US
536 (
537 Model::AmazonNovaPremier
538 | Model::Claude3_5Haiku
539 | Model::Claude3_5Sonnet
540 | Model::Claude3_5SonnetV2
541 | Model::Claude3_7Sonnet
542 | Model::Claude3_7SonnetThinking
543 | Model::ClaudeSonnet4
544 | Model::ClaudeSonnet4Thinking
545 | Model::ClaudeOpus4
546 | Model::ClaudeOpus4Thinking
547 | Model::ClaudeOpus4_1
548 | Model::ClaudeOpus4_1Thinking
549 | Model::Claude3Haiku
550 | Model::Claude3Opus
551 | Model::Claude3Sonnet
552 | Model::DeepSeekR1
553 | Model::MetaLlama31405BInstructV1
554 | Model::MetaLlama3170BInstructV1_128k
555 | Model::MetaLlama3170BInstructV1
556 | Model::MetaLlama318BInstructV1_128k
557 | Model::MetaLlama318BInstructV1
558 | Model::MetaLlama3211BInstructV1
559 | Model::MetaLlama321BInstructV1
560 | Model::MetaLlama323BInstructV1
561 | Model::MetaLlama3290BInstructV1
562 | Model::MetaLlama3370BInstructV1
563 | Model::MetaLlama4Maverick17BInstructV1
564 | Model::MetaLlama4Scout17BInstructV1
565 | Model::MistralPixtralLarge2502V1
566 | Model::PalmyraWriterX4
567 | Model::PalmyraWriterX5,
568 "us",
569 ) => Ok(format!("{}.{}", region_group, model_id)),
570
571 // Models available in EU
572 (
573 Model::Claude3_5Sonnet
574 | Model::Claude3_7Sonnet
575 | Model::Claude3_7SonnetThinking
576 | Model::ClaudeSonnet4
577 | Model::ClaudeSonnet4Thinking
578 | Model::Claude3Haiku
579 | Model::Claude3Sonnet
580 | Model::MetaLlama321BInstructV1
581 | Model::MetaLlama323BInstructV1
582 | Model::MistralPixtralLarge2502V1,
583 "eu",
584 ) => Ok(format!("{}.{}", region_group, model_id)),
585
586 // Models available in APAC
587 (
588 Model::Claude3_5Sonnet
589 | Model::Claude3_5SonnetV2
590 | Model::Claude3Haiku
591 | Model::Claude3Sonnet
592 | Model::Claude3_7Sonnet
593 | Model::Claude3_7SonnetThinking
594 | Model::ClaudeSonnet4
595 | Model::ClaudeSonnet4Thinking,
596 "apac",
597 ) => Ok(format!("{}.{}", region_group, model_id)),
598
599 // Any other combination is not supported
600 _ => Ok(self.request_id().into()),
601 }
602 }
603}
604
605#[cfg(test)]
606mod tests {
607 use super::*;
608
609 #[test]
610 fn test_us_region_inference_ids() -> anyhow::Result<()> {
611 // Test US regions
612 assert_eq!(
613 Model::Claude3_5SonnetV2.cross_region_inference_id("us-east-1")?,
614 "us.anthropic.claude-3-5-sonnet-20241022-v2:0"
615 );
616 assert_eq!(
617 Model::Claude3_5SonnetV2.cross_region_inference_id("us-west-2")?,
618 "us.anthropic.claude-3-5-sonnet-20241022-v2:0"
619 );
620 assert_eq!(
621 Model::AmazonNovaPro.cross_region_inference_id("us-east-2")?,
622 "us.amazon.nova-pro-v1:0"
623 );
624 Ok(())
625 }
626
627 #[test]
628 fn test_eu_region_inference_ids() -> anyhow::Result<()> {
629 // Test European regions
630 assert_eq!(
631 Model::ClaudeSonnet4.cross_region_inference_id("eu-west-1")?,
632 "eu.anthropic.claude-sonnet-4-20250514-v1:0"
633 );
634 assert_eq!(
635 Model::Claude3Sonnet.cross_region_inference_id("eu-west-1")?,
636 "eu.anthropic.claude-3-sonnet-20240229-v1:0"
637 );
638 assert_eq!(
639 Model::AmazonNovaMicro.cross_region_inference_id("eu-north-1")?,
640 "eu.amazon.nova-micro-v1:0"
641 );
642 Ok(())
643 }
644
645 #[test]
646 fn test_apac_region_inference_ids() -> anyhow::Result<()> {
647 // Test Asia-Pacific regions
648 assert_eq!(
649 Model::Claude3_5SonnetV2.cross_region_inference_id("ap-northeast-1")?,
650 "apac.anthropic.claude-3-5-sonnet-20241022-v2:0"
651 );
652 assert_eq!(
653 Model::Claude3_5SonnetV2.cross_region_inference_id("ap-southeast-2")?,
654 "apac.anthropic.claude-3-5-sonnet-20241022-v2:0"
655 );
656 assert_eq!(
657 Model::AmazonNovaLite.cross_region_inference_id("ap-south-1")?,
658 "apac.amazon.nova-lite-v1:0"
659 );
660 Ok(())
661 }
662
663 #[test]
664 fn test_gov_region_inference_ids() -> anyhow::Result<()> {
665 // Test Government regions
666 assert_eq!(
667 Model::Claude3_5Sonnet.cross_region_inference_id("us-gov-east-1")?,
668 "us-gov.anthropic.claude-3-5-sonnet-20240620-v1:0"
669 );
670 assert_eq!(
671 Model::Claude3Haiku.cross_region_inference_id("us-gov-west-1")?,
672 "us-gov.anthropic.claude-3-haiku-20240307-v1:0"
673 );
674 Ok(())
675 }
676
677 #[test]
678 fn test_meta_models_inference_ids() -> anyhow::Result<()> {
679 // Test Meta models
680 assert_eq!(
681 Model::MetaLlama370BInstructV1.cross_region_inference_id("us-east-1")?,
682 "meta.llama3-70b-instruct-v1:0"
683 );
684 assert_eq!(
685 Model::MetaLlama3170BInstructV1.cross_region_inference_id("us-east-1")?,
686 "us.meta.llama3-1-70b-instruct-v1:0"
687 );
688 assert_eq!(
689 Model::MetaLlama321BInstructV1.cross_region_inference_id("eu-west-1")?,
690 "eu.meta.llama3-2-1b-instruct-v1:0"
691 );
692 Ok(())
693 }
694
695 #[test]
696 fn test_mistral_models_inference_ids() -> anyhow::Result<()> {
697 // Mistral models don't follow the regional prefix pattern,
698 // so they should return their original IDs
699 assert_eq!(
700 Model::MistralMistralLarge2402V1.cross_region_inference_id("us-east-1")?,
701 "mistral.mistral-large-2402-v1:0"
702 );
703 assert_eq!(
704 Model::MistralMixtral8x7BInstructV0.cross_region_inference_id("eu-west-1")?,
705 "mistral.mixtral-8x7b-instruct-v0:1"
706 );
707 Ok(())
708 }
709
710 #[test]
711 fn test_ai21_models_inference_ids() -> anyhow::Result<()> {
712 // AI21 models don't follow the regional prefix pattern,
713 // so they should return their original IDs
714 assert_eq!(
715 Model::AI21J2UltraV1.cross_region_inference_id("us-east-1")?,
716 "ai21.j2-ultra-v1"
717 );
718 assert_eq!(
719 Model::AI21JambaInstructV1.cross_region_inference_id("eu-west-1")?,
720 "ai21.jamba-instruct-v1:0"
721 );
722 Ok(())
723 }
724
725 #[test]
726 fn test_cohere_models_inference_ids() -> anyhow::Result<()> {
727 // Cohere models don't follow the regional prefix pattern,
728 // so they should return their original IDs
729 assert_eq!(
730 Model::CohereCommandRV1.cross_region_inference_id("us-east-1")?,
731 "cohere.command-r-v1:0"
732 );
733 assert_eq!(
734 Model::CohereCommandTextV14_4k.cross_region_inference_id("ap-southeast-1")?,
735 "cohere.command-text-v14:7:4k"
736 );
737 Ok(())
738 }
739
740 #[test]
741 fn test_custom_model_inference_ids() -> anyhow::Result<()> {
742 // Test custom models
743 let custom_model = Model::Custom {
744 name: "custom.my-model-v1:0".to_string(),
745 max_tokens: 100000,
746 display_name: Some("My Custom Model".to_string()),
747 max_output_tokens: Some(8192),
748 default_temperature: Some(0.7),
749 cache_configuration: None,
750 };
751
752 // Custom model should return its name unchanged
753 assert_eq!(
754 custom_model.cross_region_inference_id("us-east-1")?,
755 "custom.my-model-v1:0"
756 );
757
758 Ok(())
759 }
760
761 #[test]
762 fn test_friendly_id_vs_request_id() {
763 // Test that id() returns friendly identifiers
764 assert_eq!(Model::Claude3_5SonnetV2.id(), "claude-3-5-sonnet-v2");
765 assert_eq!(Model::AmazonNovaLite.id(), "amazon-nova-lite");
766 assert_eq!(Model::DeepSeekR1.id(), "deepseek-r1");
767 assert_eq!(
768 Model::MetaLlama38BInstructV1.id(),
769 "meta-llama3-8b-instruct-v1"
770 );
771
772 // Test that request_id() returns actual backend model IDs
773 assert_eq!(
774 Model::Claude3_5SonnetV2.request_id(),
775 "anthropic.claude-3-5-sonnet-20241022-v2:0"
776 );
777 assert_eq!(Model::AmazonNovaLite.request_id(), "amazon.nova-lite-v1:0");
778 assert_eq!(Model::DeepSeekR1.request_id(), "deepseek.r1-v1:0");
779 assert_eq!(
780 Model::MetaLlama38BInstructV1.request_id(),
781 "meta.llama3-8b-instruct-v1:0"
782 );
783
784 // Test thinking models have different friendly IDs but same request IDs
785 assert_eq!(Model::ClaudeSonnet4.id(), "claude-sonnet-4");
786 assert_eq!(
787 Model::ClaudeSonnet4Thinking.id(),
788 "claude-sonnet-4-thinking"
789 );
790 assert_eq!(
791 Model::ClaudeSonnet4.request_id(),
792 Model::ClaudeSonnet4Thinking.request_id()
793 );
794 }
795}