| 1 | //! Small in-process cache for deterministic non-streaming chat responses. |
| 2 | |
| 3 | use std::num::NonZeroUsize; |
| 4 | use std::sync::{Mutex, OnceLock}; |
| 5 | |
| 6 | use lru::LruCache; |
| 7 | use sha2::{Digest, Sha256}; |
| 8 | |
| 9 | use crate::models::{MessageRequest, MessageResponse, Usage}; |
| 10 | |
| 11 | const DEFAULT_CAPACITY: usize = 256; |
| 12 | |
| 13 | static RESPONSE_CACHE: OnceLock<ResponseCache> = OnceLock::new(); |
| 14 | |
| 15 | pub(crate) fn response_cache() -> &'static ResponseCache { |
| 16 | RESPONSE_CACHE.get_or_init(ResponseCache::new) |
| 17 | } |
| 18 | |
| 19 | pub(crate) fn request_is_cacheable(request: &MessageRequest) -> bool { |
| 20 | request.stream != Some(true) |
| 21 | && request.tools.as_ref().is_none_or(Vec::is_empty) |
| 22 | && request.tool_choice.is_none() |
| 23 | && request.temperature == Some(0.0) |
| 24 | && request.top_p.is_none_or(|top_p| top_p == 1.0) |
| 25 | } |
| 26 | |
| 27 | pub(crate) struct ResponseCache { |
| 28 | inner: Mutex<LruCache<[u8; 32], MessageResponse>>, |
| 29 | } |
| 30 | |
| 31 | impl ResponseCache { |
| 32 | fn new() -> Self { |
| 33 | Self::with_capacity(NonZeroUsize::new(DEFAULT_CAPACITY).expect("non-zero capacity")) |
| 34 | } |
| 35 | |
| 36 | fn with_capacity(capacity: NonZeroUsize) -> Self { |
| 37 | Self { |
| 38 | inner: Mutex::new(LruCache::new(capacity)), |
| 39 | } |
| 40 | } |
| 41 | |
| 42 | pub(crate) fn make_key( |
| 43 | provider: &str, |
| 44 | base_url: &str, |
| 45 | path_suffix: Option<&str>, |
| 46 | api_key: &str, |
| 47 | wire_body: &[u8], |
| 48 | ) -> [u8; 32] { |
| 49 | let mut hasher = Sha256::new(); |
| 50 | update_field(&mut hasher, provider.as_bytes()); |
| 51 | update_field(&mut hasher, base_url.as_bytes()); |
| 52 | update_field(&mut hasher, path_suffix.unwrap_or("").as_bytes()); |
| 53 | update_field(&mut hasher, &Sha256::digest(api_key.as_bytes())); |
| 54 | update_field(&mut hasher, wire_body); |
| 55 | hasher.finalize().into() |
| 56 | } |
| 57 | |
| 58 | pub(crate) fn get(&self, key: &[u8; 32]) -> Option<MessageResponse> { |
| 59 | let mut cache = self.inner.lock().ok()?; |
| 60 | cache.get(key).cloned().map(|mut response| { |
| 61 | response.usage = Usage::default(); |
| 62 | response |
| 63 | }) |
| 64 | } |
| 65 | |
| 66 | pub(crate) fn put(&self, key: [u8; 32], value: MessageResponse) { |
| 67 | if let Ok(mut cache) = self.inner.lock() { |
| 68 | cache.put(key, value); |
| 69 | } |
| 70 | } |
| 71 | } |
| 72 | |
| 73 | fn update_field(hasher: &mut Sha256, bytes: &[u8]) { |
| 74 | hasher.update((bytes.len() as u64).to_le_bytes()); |
| 75 | hasher.update(bytes); |
| 76 | } |
| 77 | |
| 78 | #[cfg(test)] |
| 79 | mod tests { |
| 80 | use super::*; |
| 81 | |
| 82 | fn response_with_usage(id: &str) -> MessageResponse { |
| 83 | MessageResponse { |
| 84 | id: id.to_string(), |
| 85 | r#type: "message".to_string(), |
| 86 | role: "assistant".to_string(), |
| 87 | content: Vec::new(), |
| 88 | model: "test-model".to_string(), |
| 89 | stop_reason: Some("end_turn".to_string()), |
| 90 | stop_sequence: None, |
| 91 | container: None, |
| 92 | usage: Usage { |
| 93 | input_tokens: 42, |
| 94 | output_tokens: 7, |
| 95 | prompt_cache_hit_tokens: Some(3), |
| 96 | prompt_cache_miss_tokens: Some(39), |
| 97 | prompt_cache_write_tokens: None, |
| 98 | reasoning_tokens: Some(5), |
| 99 | reasoning_replay_tokens: Some(2), |
| 100 | server_tool_use: None, |
| 101 | }, |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | fn request() -> MessageRequest { |
| 106 | MessageRequest { |
| 107 | model: "test-model".to_string(), |
| 108 | messages: Vec::new(), |
| 109 | max_tokens: 16, |
| 110 | system: None, |
| 111 | tools: None, |
| 112 | tool_choice: None, |
| 113 | metadata: None, |
| 114 | thinking: None, |
| 115 | reasoning_effort: None, |
| 116 | stream: None, |
| 117 | temperature: Some(0.0), |
| 118 | top_p: None, |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | #[test] |
| 123 | fn cache_key_separates_provider_route_account_and_wire_body() { |
| 124 | let base = ResponseCache::make_key( |
| 125 | "deepseek", |
| 126 | "https://api.example.com/v1", |
| 127 | None, |
| 128 | "key-a", |
| 129 | br#"{"model":"m","messages":[]}"#, |
| 130 | ); |
| 131 | |
| 132 | assert_ne!( |
| 133 | base, |
| 134 | ResponseCache::make_key( |
| 135 | "openai", |
| 136 | "https://api.example.com/v1", |
| 137 | None, |
| 138 | "key-a", |
| 139 | br#"{"model":"m","messages":[]}"# |
| 140 | ) |
| 141 | ); |
| 142 | assert_ne!( |
| 143 | base, |
| 144 | ResponseCache::make_key( |
| 145 | "deepseek", |
| 146 | "https://proxy.example.com/v1", |
| 147 | None, |
| 148 | "key-a", |
| 149 | br#"{"model":"m","messages":[]}"# |
| 150 | ) |
| 151 | ); |
| 152 | assert_ne!( |
| 153 | base, |
| 154 | ResponseCache::make_key( |
| 155 | "deepseek", |
| 156 | "https://api.example.com/v1", |
| 157 | Some("responses"), |
| 158 | "key-a", |
| 159 | br#"{"model":"m","messages":[]}"# |
| 160 | ) |
| 161 | ); |
| 162 | assert_ne!( |
| 163 | base, |
| 164 | ResponseCache::make_key( |
| 165 | "deepseek", |
| 166 | "https://api.example.com/v1", |
| 167 | None, |
| 168 | "key-b", |
| 169 | br#"{"model":"m","messages":[]}"# |
| 170 | ) |
| 171 | ); |
| 172 | assert_ne!( |
| 173 | base, |
| 174 | ResponseCache::make_key( |
| 175 | "deepseek", |
| 176 | "https://api.example.com/v1", |
| 177 | None, |
| 178 | "key-a", |
| 179 | br#"{"model":"m","messages":[],"reasoning_effort":"high"}"# |
| 180 | ) |
| 181 | ); |
| 182 | } |
| 183 | |
| 184 | #[test] |
| 185 | fn cache_hit_zeroes_usage_to_avoid_fake_spend() { |
| 186 | let cache = ResponseCache::with_capacity(NonZeroUsize::new(2).unwrap()); |
| 187 | let key = |
| 188 | ResponseCache::make_key("deepseek", "https://api.example.com", None, "key", b"{}"); |
| 189 | |
| 190 | cache.put(key, response_with_usage("cached")); |
| 191 | |
| 192 | let hit = cache.get(&key).expect("cache hit"); |
| 193 | assert_eq!(hit.id, "cached"); |
| 194 | assert_eq!(hit.usage, Usage::default()); |
| 195 | } |
| 196 | |
| 197 | #[test] |
| 198 | fn capacity_evicts_oldest_entry() { |
| 199 | let cache = ResponseCache::with_capacity(NonZeroUsize::new(2).unwrap()); |
| 200 | let key1 = |
| 201 | ResponseCache::make_key("deepseek", "https://api.example.com", None, "key", b"one"); |
| 202 | let key2 = |
| 203 | ResponseCache::make_key("deepseek", "https://api.example.com", None, "key", b"two"); |
| 204 | let key3 = |
| 205 | ResponseCache::make_key("deepseek", "https://api.example.com", None, "key", b"three"); |
| 206 | |
| 207 | cache.put(key1, response_with_usage("one")); |
| 208 | cache.put(key2, response_with_usage("two")); |
| 209 | cache.put(key3, response_with_usage("three")); |
| 210 | |
| 211 | assert!(cache.get(&key1).is_none()); |
| 212 | assert!(cache.get(&key2).is_some()); |
| 213 | assert!(cache.get(&key3).is_some()); |
| 214 | } |
| 215 | |
| 216 | #[test] |
| 217 | fn cacheability_requires_deterministic_tool_free_non_streaming_request() { |
| 218 | let mut req = request(); |
| 219 | assert!(request_is_cacheable(&req)); |
| 220 | |
| 221 | req.temperature = None; |
| 222 | assert!(!request_is_cacheable(&req)); |
| 223 | |
| 224 | req = request(); |
| 225 | req.temperature = Some(0.2); |
| 226 | assert!(!request_is_cacheable(&req)); |
| 227 | |
| 228 | req = request(); |
| 229 | req.stream = Some(true); |
| 230 | assert!(!request_is_cacheable(&req)); |
| 231 | |
| 232 | req = request(); |
| 233 | req.top_p = Some(0.5); |
| 234 | assert!(!request_is_cacheable(&req)); |
| 235 | |
| 236 | req = request(); |
| 237 | req.tool_choice = Some(serde_json::json!("auto")); |
| 238 | assert!(!request_is_cacheable(&req)); |
| 239 | } |
| 240 | } |
| 241 |