diff --git a/litellm-rust/crates/cache-valkey-semantic/src/lib.rs b/litellm-rust/crates/cache-valkey-semantic/src/lib.rs index 85c4c9af15c..ef028f0a7b2 100644 --- a/litellm-rust/crates/cache-valkey-semantic/src/lib.rs +++ b/litellm-rust/crates/cache-valkey-semantic/src/lib.rs @@ -62,6 +62,14 @@ enum Connections { Fixed(Mutex), } +#[derive(Clone)] +struct IndexState { + name: String, + prefix: String, + dimension: Arc>>, + similarity_threshold: f64, +} + struct ConnectionRef<'a>(&'a mut dyn redis::ConnectionLike); impl redis::ConnectionLike for ConnectionRef<'_> { @@ -187,18 +195,13 @@ where &self.config.index_name } - fn key_prefix(&self) -> String { - format!("{}:", self.config.index_name) - } - - fn ensure_index(&self, dimension: usize) -> Result<(), Error> { - ensure_index( - &self.connections, - &self.config.index_name, - &self.key_prefix(), - &self.index_dimension, - dimension, - ) + fn index_state(&self) -> IndexState { + IndexState { + name: self.config.index_name.clone(), + prefix: format!("{}:", self.config.index_name), + dimension: Arc::clone(&self.index_dimension), + similarity_threshold: self.config.similarity_threshold, + } } } @@ -225,37 +228,19 @@ where return Ok(()); }; let embedding = self.embedder.embed(&prompt, context.metadata.as_ref())?; - self.ensure_index(embedding.len())?; let scope = scope_tag(key); - let document = format!("{}{}:{}", self.key_prefix(), scope, Uuid::new_v4()); let response = self.codec.encode(&value)?; let vector = embedding_bytes(&embedding); - let ttl = self.get_ttl(context); - self.connections.execute(|connection| { - let mut pipeline = redis::pipe(); - pipeline - .cmd("HSET") - .arg(&document) - .arg("litellm_cache_key") - .arg(&scope) - .arg("prompt") - .arg(prompt) - .arg("response") - .arg(response) - .arg("embedding") - .arg(vector) - .ignore(); - if let Some(ttl) = ttl { - pipeline - .cmd("EXPIRE") - .arg(&document) - .arg(ttl.as_secs()) - .ignore(); - } - pipeline - .query::<()>(connection) - .map_err(|_| Error::Unavailable) - }) + let index = self.index_state(); + write_document( + &self.connections, + &index, + &scope, + &prompt, + response, + vector, + self.get_ttl(context), + ) } fn get_cache(&self, key: &str, context: &Self::Context) -> Result, Error> { @@ -263,43 +248,14 @@ where return Ok(None); }; let embedding = self.embedder.embed(&prompt, context.metadata.as_ref())?; - self.ensure_index(embedding.len())?; let scope = scope_tag(key); - let query = - format!("(@litellm_cache_key:{{{scope}}})=>[KNN 1 @embedding $vec AS vector_distance]"); let vector = embedding_bytes(&embedding); - let response = self.connections.execute(|connection| { - redis::cmd("FT.SEARCH") - .arg(&self.config.index_name) - .arg(query) - .arg("PARAMS") - .arg(2) - .arg("vec") - .arg(vector) - .arg("RETURN") - .arg(2) - .arg("response") - .arg("vector_distance") - .arg("DIALECT") - .arg(2) - .query::(connection) - .map_err(|_| Error::Unavailable) - })?; - let Some(fields) = search_fields(response)? else { + let index = self.index_state(); + let Some(response) = + search_document(&self.connections, &index, &scope, vector, embedding.len())? + else { return Ok(None); }; - let response = fields - .iter() - .find_map(|(name, value)| (name == "response").then(|| value.clone())) - .ok_or(Error::InvalidEntry)?; - let distance = fields - .iter() - .find_map(|(name, value)| (name == "vector_distance").then(|| value.clone())) - .ok_or(Error::InvalidEntry)?; - let distance = parse_f64(&distance)?; - if 1.0 - distance < self.config.similarity_threshold { - return Ok(None); - } self.codec.decode(&response).map(Some) } @@ -321,47 +277,13 @@ where .async_embed(&prompt, metadata.as_ref()) .await?; let connections = Arc::clone(&self.connections); - let config = self.config.clone(); - let index_dimension = Arc::clone(&self.index_dimension); + let index = self.index_state(); let response = self.codec.encode(&value)?; let vector = embedding_bytes(&embedding); - let prefix = format!("{}:", config.index_name); let scope = scope_tag(&key); - let document = format!("{prefix}{scope}:{}", Uuid::new_v4()); let ttl = context.ttl; tokio::task::spawn_blocking(move || { - ensure_index( - &connections, - &config.index_name, - &prefix, - &index_dimension, - embedding.len(), - )?; - connections.execute(|connection| { - let mut pipeline = redis::pipe(); - pipeline - .cmd("HSET") - .arg(&document) - .arg("litellm_cache_key") - .arg(&scope) - .arg("prompt") - .arg(prompt) - .arg("response") - .arg(response) - .arg("embedding") - .arg(vector) - .ignore(); - if let Some(ttl) = ttl { - pipeline - .cmd("EXPIRE") - .arg(&document) - .arg(ttl.as_secs()) - .ignore(); - } - pipeline - .query::<()>(connection) - .map_err(|_| Error::Unavailable) - }) + write_document(&connections, &index, &scope, &prompt, response, vector, ttl) }) .await .map_err(|_| Error::Unavailable)? @@ -385,56 +307,11 @@ where .async_embed(&prompt, metadata.as_ref()) .await?; let connections = Arc::clone(&self.connections); - let config = self.config.clone(); - let index_dimension = Arc::clone(&self.index_dimension); - let threshold = config.similarity_threshold; + let index = self.index_state(); tokio::task::spawn_blocking(move || { - let prefix = format!("{}:", config.index_name); - ensure_index( - &connections, - &config.index_name, - &prefix, - &index_dimension, - embedding.len(), - )?; let scope = scope_tag(&key); - let query = format!( - "(@litellm_cache_key:{{{scope}}})=>[KNN 1 @embedding $vec AS vector_distance]" - ); let vector = embedding_bytes(&embedding); - let response = connections.execute(|connection| { - redis::cmd("FT.SEARCH") - .arg(&config.index_name) - .arg(query) - .arg("PARAMS") - .arg(2) - .arg("vec") - .arg(vector) - .arg("RETURN") - .arg(2) - .arg("response") - .arg("vector_distance") - .arg("DIALECT") - .arg(2) - .query::(connection) - .map_err(|_| Error::Unavailable) - })?; - let Some(fields) = search_fields(response)? else { - return Ok(None); - }; - let response = fields - .iter() - .find_map(|(name, value)| (name == "response").then(|| value.clone())) - .ok_or(Error::InvalidEntry)?; - let distance = fields - .iter() - .find_map(|(name, value)| (name == "vector_distance").then(|| value.clone())) - .ok_or(Error::InvalidEntry)?; - let distance = parse_f64(&distance)?; - if 1.0 - distance < threshold { - return Ok(None); - } - Ok(Some(response)) + search_document(&connections, &index, &scope, vector, embedding.len()) }) .await .map_err(|_| Error::Unavailable)? @@ -560,6 +437,108 @@ fn embedding_bytes(embedding: &[f32]) -> Vec { .collect() } +fn write_document( + connections: &Connections, + index: &IndexState, + scope: &str, + prompt: &str, + response: Vec, + vector: Vec, + ttl: Option, +) -> Result<(), Error> +where + C: redis::ConnectionLike + Send + 'static, +{ + let dimension = vector.len() / std::mem::size_of::(); + ensure_index( + connections, + &index.name, + &index.prefix, + &index.dimension, + dimension, + )?; + let document = format!("{}{scope}:{}", index.prefix, Uuid::new_v4()); + connections.execute(|connection| { + let mut pipeline = redis::pipe(); + pipeline + .cmd("HSET") + .arg(&document) + .arg("litellm_cache_key") + .arg(scope) + .arg("prompt") + .arg(prompt) + .arg("response") + .arg(response) + .arg("embedding") + .arg(vector) + .ignore(); + if let Some(ttl) = ttl { + pipeline + .cmd("EXPIRE") + .arg(&document) + .arg(ttl.as_secs()) + .ignore(); + } + pipeline + .query::<()>(connection) + .map_err(|_| Error::Unavailable) + }) +} + +fn search_document( + connections: &Connections, + index: &IndexState, + scope: &str, + vector: Vec, + dimension: usize, +) -> Result>, Error> +where + C: redis::ConnectionLike + Send + 'static, +{ + ensure_index( + connections, + &index.name, + &index.prefix, + &index.dimension, + dimension, + )?; + let query = + format!("(@litellm_cache_key:{{{scope}}})=>[KNN 1 @embedding $vec AS vector_distance]"); + let response = connections.execute(|connection| { + redis::cmd("FT.SEARCH") + .arg(&index.name) + .arg(query) + .arg("PARAMS") + .arg(2) + .arg("vec") + .arg(vector) + .arg("RETURN") + .arg(2) + .arg("response") + .arg("vector_distance") + .arg("DIALECT") + .arg(2) + .query::(connection) + .map_err(|_| Error::Unavailable) + })?; + let Some(fields) = search_fields(response)? else { + return Ok(None); + }; + let response = fields + .iter() + .find_map(|(name, value)| (name == "response").then(|| value.clone())) + .ok_or(Error::InvalidEntry)?; + let distance = fields + .iter() + .find_map(|(name, value)| (name == "vector_distance").then(|| value.clone())) + .ok_or(Error::InvalidEntry)?; + let distance = parse_f64(&distance)?; + if 1.0 - distance < index.similarity_threshold { + return Ok(None); + } + Ok(Some(response)) +} + fn ensure_index( connections: &Connections, index_name: &str, @@ -712,10 +691,16 @@ fn value_bytes(value: &redis::Value) -> Result, Error> { #[cfg(test)] mod tests { - use std::sync::{Arc, Mutex}; + use std::{ + collections::VecDeque, + sync::{Arc, Mutex}, + time::Duration, + }; - use litellm_cache::BaseCache; - use litellm_cache_response::ResponseCacheCodec; + use litellm_cache::{BaseCache, CacheCodec}; + use litellm_cache_response::{ + CacheEntry, CacheKeyInput, ResponseCache, ResponseCacheCodec, ResponseCacheRequest, + }; use redis_test::MockRedisConnection; use rstest::rstest; use serde_json::{Value, json}; @@ -732,6 +717,9 @@ mod tests { } type EmbedderCalls = Arc)>>>; + type RecordingCache = + ValkeySemanticCache; + type RecordingSetup = (RecordingCache, Arc>>>, EmbedderCalls); impl Embedder for FixedEmbedder { fn embed(&self, prompt: &str, metadata: Option<&Value>) -> Result, super::Error> { @@ -751,6 +739,61 @@ mod tests { } } + struct RecordingConnection { + requests: Arc>>>, + replies: Mutex>>, + } + + impl RecordingConnection { + fn new(replies: impl IntoIterator>) -> Self { + Self { + requests: Arc::default(), + replies: Mutex::new(replies.into_iter().collect()), + } + } + + fn requests(&self) -> Arc>>> { + Arc::clone(&self.requests) + } + + fn reply(&self) -> redis::RedisResult { + self.replies + .lock() + .unwrap() + .pop_front() + .unwrap_or_else(|| Ok(redis::Value::SimpleString("OK".into()))) + } + } + + impl redis::ConnectionLike for RecordingConnection { + fn req_packed_command(&mut self, command: &[u8]) -> redis::RedisResult { + self.requests.lock().unwrap().push(command.to_vec()); + self.reply() + } + + fn req_packed_commands( + &mut self, + command: &[u8], + _offset: usize, + count: usize, + ) -> redis::RedisResult> { + self.requests.lock().unwrap().push(command.to_vec()); + (0..count).map(|_| self.reply()).collect() + } + + fn get_db(&self) -> i64 { + 0 + } + + fn check_connection(&mut self) -> bool { + true + } + + fn is_open(&self) -> bool { + true + } + } + fn context( messages: Option, input: Option, @@ -841,4 +884,302 @@ mod tests { assert_eq!(cache.get_cache("key", &context(None, None)).unwrap(), None); assert_eq!(cache.get_ttl(&context(None, None)), None); } + + fn semantic_context(ttl: Option) -> litellm_cache::SemanticCacheContext { + litellm_cache::SemanticCacheContext { + messages: Some(json!([{"role": "user", "content": "hello"}])), + metadata: Some(json!({"source": "test"})), + ttl, + ..Default::default() + } + } + + fn cache_with_recording( + replies: impl IntoIterator>, + vector: Vec, + threshold: f64, + ) -> RecordingSetup { + let connection = RecordingConnection::new(replies); + let requests = connection.requests(); + let calls: EmbedderCalls = Arc::default(); + let cache = ValkeySemanticCache::with_connection( + connection, + FixedEmbedder { + vector, + calls: Arc::clone(&calls), + }, + ResponseCacheCodec, + ValkeySemanticConfig { + similarity_threshold: threshold, + index_name: "test".into(), + }, + ); + (cache, requests, calls) + } + + fn ok() -> redis::RedisResult { + Ok(redis::Value::SimpleString("OK".into())) + } + + fn already_exists() -> redis::RedisResult { + Err(redis::RedisError::from(( + redis::ErrorKind::Io, + "already exists", + ))) + } + + fn info_dimension(dimension: usize) -> redis::Value { + redis::Value::Array(vec![ + redis::Value::SimpleString("attributes".into()), + redis::Value::Array(vec![redis::Value::Array(vec![ + redis::Value::SimpleString("embedding".into()), + redis::Value::Array(vec![ + redis::Value::SimpleString("dimensions".into()), + redis::Value::Int(dimension as i64), + ]), + ])]), + ]) + } + + fn search_hit(response: Vec, distance: &str) -> redis::Value { + redis::Value::Array(vec![ + redis::Value::Int(1), + redis::Value::BulkString(b"test:document".to_vec()), + redis::Value::Array(vec![ + redis::Value::BulkString(b"response".to_vec()), + redis::Value::BulkString(response), + redis::Value::BulkString(b"vector_distance".to_vec()), + redis::Value::BulkString(distance.as_bytes().to_vec()), + ]), + ]) + } + + fn requests_text(requests: &Arc>>>) -> String { + requests + .lock() + .unwrap() + .iter() + .map(|request| String::from_utf8_lossy(request)) + .collect::>() + .join("\n") + } + + #[test] + fn set_without_ttl_writes_hset_without_expire() { + let (cache, requests, calls) = cache_with_recording([ok()], vec![1.0, 0.0], 0.8); + cache + .set_cache( + "key", + CacheEntry { + timestamp: None, + response: json!({"answer": "ok"}), + }, + &semantic_context(None), + ) + .unwrap(); + let text = requests_text(&requests); + assert!(text.contains("FT.CREATE")); + assert!(text.contains("HSET")); + assert!( + text.contains("test:2c70e12b7a0646f92279f427c7b38e7334d8e5389cff167a1dc30e73f826b683:") + ); + assert!(!text.contains("EXPIRE")); + assert_eq!( + *calls.lock().unwrap(), + vec![("hello".into(), Some(json!({"source": "test"})))] + ); + } + + #[test] + fn set_with_ttl_truncates_expire_seconds() { + let (cache, requests, _) = cache_with_recording([ok()], vec![1.0, 0.0], 0.8); + cache + .set_cache( + "key", + CacheEntry { + timestamp: None, + response: json!({"answer": "ok"}), + }, + &semantic_context(Some(Duration::from_millis(1900))), + ) + .unwrap(); + let text = requests_text(&requests); + assert!(text.contains("EXPIRE")); + assert!(text.contains("\r\n$1\r\n1\r\n")); + } + + #[test] + fn second_set_skips_create_after_dimension_is_cached() { + let (cache, requests, _) = cache_with_recording([ok()], vec![1.0, 0.0], 0.8); + let context = semantic_context(None); + let entry = CacheEntry { + timestamp: None, + response: json!({"answer": "ok"}), + }; + cache.set_cache("key", entry.clone(), &context).unwrap(); + cache.set_cache("key", entry, &context).unwrap(); + let text = requests_text(&requests); + assert_eq!(text.matches("FT.CREATE").count(), 1); + assert_eq!(text.matches("HSET").count(), 2); + } + + #[test] + fn existing_index_dimension_must_match_embedding() { + let (cache, _, _) = cache_with_recording( + [already_exists(), Ok(info_dimension(2))], + vec![1.0, 0.0], + 0.8, + ); + cache + .set_cache( + "key", + CacheEntry { + timestamp: None, + response: json!({"answer": "ok"}), + }, + &semantic_context(None), + ) + .unwrap(); + + let (cache, _, _) = cache_with_recording( + [already_exists(), Ok(info_dimension(3))], + vec![1.0, 0.0], + 0.8, + ); + assert_eq!( + cache.set_cache( + "key", + CacheEntry { + timestamp: None, + response: json!({"answer": "ok"}), + }, + &semantic_context(None), + ), + Err(super::Error::Unavailable) + ); + } + + #[test] + fn get_applies_threshold_and_decodes_entry() { + let entry = CacheEntry { + timestamp: Some(1.0), + response: json!({"answer": "ok"}), + }; + let encoded = ResponseCacheCodec.encode(&entry).unwrap(); + let (cache, _, _) = cache_with_recording( + [ok(), Ok(search_hit(encoded.clone(), "0.1"))], + vec![1.0, 0.0], + 0.8, + ); + assert_eq!( + cache.get_cache("key", &semantic_context(None)).unwrap(), + Some(entry) + ); + + let (cache, _, _) = + cache_with_recording([ok(), Ok(search_hit(encoded, "0.5"))], vec![1.0, 0.0], 0.8); + assert_eq!( + cache.get_cache("key", &semantic_context(None)).unwrap(), + None + ); + } + + #[test] + fn get_zero_docs_is_a_miss() { + let (cache, _, _) = cache_with_recording( + [ok(), Ok(redis::Value::Array(vec![redis::Value::Int(0)]))], + vec![1.0, 0.0], + 0.8, + ); + assert_eq!( + cache.get_cache("key", &semantic_context(None)).unwrap(), + None + ); + } + + #[rstest] + #[case(redis::Value::Array(vec![ + redis::Value::Int(1), + redis::Value::BulkString(b"document".to_vec()), + redis::Value::Array(vec![ + redis::Value::BulkString(b"vector_distance".to_vec()), + redis::Value::BulkString(b"0.1".to_vec()), + ]), + ]))] + #[case(redis::Value::Array(vec![ + redis::Value::Int(1), + redis::Value::BulkString(b"document".to_vec()), + redis::Value::Array(vec![ + redis::Value::BulkString(b"response".to_vec()), + redis::Value::BulkString(b"not-json".to_vec()), + redis::Value::BulkString(b"vector_distance".to_vec()), + redis::Value::BulkString(b"abc".to_vec()), + ]), + ]))] + fn malformed_entries_are_invalid(#[case] search: redis::Value) { + let (cache, _, _) = cache_with_recording([ok(), Ok(search)], vec![1.0, 0.0], 0.8); + assert_eq!( + cache.get_cache("key", &semantic_context(None)), + Err(super::Error::InvalidEntry) + ); + } + + #[test] + fn response_cache_turns_invalid_entries_into_misses() { + let (cache, _, _) = cache_with_recording( + [ + ok(), + Ok(redis::Value::Array(vec![ + redis::Value::Int(1), + redis::Value::BulkString(b"document".to_vec()), + redis::Value::Array(vec![ + redis::Value::BulkString(b"response".to_vec()), + redis::Value::BulkString(b"not-json".to_vec()), + redis::Value::BulkString(b"vector_distance".to_vec()), + redis::Value::BulkString(b"0.1".to_vec()), + ]), + ])), + ], + vec![1.0, 0.0], + 0.8, + ); + let service = ResponseCache::new(Arc::new(cache)); + let request = ResponseCacheRequest { + key: CacheKeyInput { + preset: Some("key".into()), + ..Default::default() + }, + context: semantic_context(None), + ..ResponseCacheRequest::new(CacheKeyInput::default()) + }; + assert_eq!(service.lookup(&request, Duration::ZERO).unwrap(), None); + } + + #[tokio::test] + async fn async_set_and_get_use_shared_document_helpers() { + let entry = CacheEntry { + timestamp: Some(1.0), + response: json!({"answer": "ok"}), + }; + let encoded = ResponseCacheCodec.encode(&entry).unwrap(); + let (cache, requests, calls) = cache_with_recording( + [ok(), ok(), ok(), Ok(search_hit(encoded, "0.1"))], + vec![1.0, 0.0], + 0.8, + ); + let context = semantic_context(Some(Duration::from_millis(1900))); + cache + .async_set_cache("key", entry.clone(), context.clone()) + .await + .unwrap(); + assert_eq!( + cache.async_get_cache("key", &context).await.unwrap(), + Some(entry) + ); + let text = requests_text(&requests); + assert!(text.contains("FT.CREATE")); + assert!(text.contains("HSET")); + assert!(text.contains("EXPIRE")); + assert_eq!(calls.lock().unwrap().len(), 2); + } } diff --git a/litellm-rust/crates/python-bridge/src/cache/config.rs b/litellm-rust/crates/python-bridge/src/cache/config.rs index 7218805b8df..e074c5e2f5d 100644 --- a/litellm-rust/crates/python-bridge/src/cache/config.rs +++ b/litellm-rust/crates/python-bridge/src/cache/config.rs @@ -276,25 +276,15 @@ fn project_redis( let client = backend.getattr("redis_client")?; let pool = client.getattr("connection_pool")?; - if !instance_class_is(&pool, "redis.connection", "ConnectionPool")? { + let Ok((resolved, is_tls)) = project_connection_pool(&pool)? else { return Ok(Err(UnsupportedCacheConfig::RedisConnection)); - } - let resolved = pool.getattr("connection_kwargs")?.cast_into::()?; + }; for key in ["credential_provider", "redis_connect_func"] { if has_value(&resolved, key)? { return Ok(Err(UnsupportedCacheConfig::RedisCredentials)); } } - let connection_class = resolved - .get_item("connection_class")? - .unwrap_or(pool.getattr("connection_class")?); - let tls = if class_is(&connection_class, "redis.connection", "Connection")? { - None - } else if class_is(&connection_class, "redis.connection", "SSLConnection")? { - Some(project_tls(&resolved)?) - } else { - return Ok(Err(UnsupportedCacheConfig::RedisConnection)); - }; + let tls = is_tls.then(|| project_tls(&resolved)).transpose()?; let protocol = match optional_i64(&resolved, "protocol")?.unwrap_or(2) { 2 => RedisProtocol::Resp2, @@ -332,18 +322,9 @@ fn project_valkey_semantic( ) -> PyResult> { let client = backend.getattr("sync_client")?; let pool = client.getattr("connection_pool")?; - if !instance_class_is(&pool, "redis.connection", "ConnectionPool")? { + let Ok((resolved, _is_tls)) = project_connection_pool(&pool)? else { return Ok(Err(UnsupportedCacheConfig::RedisConnection)); - } - let resolved = pool.getattr("connection_kwargs")?.cast_into::()?; - let connection_class = resolved - .get_item("connection_class")? - .unwrap_or(pool.getattr("connection_class")?); - if !class_is(&connection_class, "redis.connection", "Connection")? - && !class_is(&connection_class, "redis.connection", "SSLConnection")? - { - return Ok(Err(UnsupportedCacheConfig::RedisConnection)); - } + }; let connection = RedisConnectionConfig { host: required_string(&resolved, "host")?, port: u16::try_from(required_i64(&resolved, "port")?) @@ -371,6 +352,27 @@ fn project_valkey_semantic( })) } +#[inline(never)] +fn project_connection_pool<'py>( + pool: &Bound<'py, PyAny>, +) -> PyResult, bool), UnsupportedCacheConfig>> { + if !instance_class_is(pool, "redis.connection", "ConnectionPool")? { + return Ok(Err(UnsupportedCacheConfig::RedisConnection)); + } + let resolved = pool.getattr("connection_kwargs")?.cast_into::()?; + let connection_class = resolved + .get_item("connection_class")? + .unwrap_or(pool.getattr("connection_class")?); + let is_tls = if class_is(&connection_class, "redis.connection", "Connection")? { + false + } else if class_is(&connection_class, "redis.connection", "SSLConnection")? { + true + } else { + return Ok(Err(UnsupportedCacheConfig::RedisConnection)); + }; + Ok(Ok((resolved, is_tls))) +} + #[inline(never)] fn project_tls(values: &Bound<'_, PyDict>) -> PyResult { Ok(RedisTlsConfig { @@ -646,6 +648,40 @@ mod tests { }); } + #[test] + fn projects_valkey_semantic_configuration() { + Python::initialize(); + Python::attach(|py| { + let facade = facade( + py, + "pool = ConnectionPool()\n\ + pool.connection_class = Connection\n\ + pool.max_connections = 12\n\ + pool.connection_kwargs = {'host': 'cache.internal', 'port': 6390, 'db': 2}\n\ + client = SimpleNamespace(connection_pool=pool)\n\ + backend = SimpleNamespace(similarity_threshold=0.85, index_name='semantic_idx', embedding_model='text-embedding-3-small', sync_client=client)\n\ + facade = SimpleNamespace(type='valkey-semantic', mode='default-on', ttl=None, namespace=None, supported_call_types=None, redis_flush_size=None, semantic_cache_scope='key', cache=backend)", + ); + let CacheConfigProjection::Native(config) = + NativeCacheConfig::project(&facade).unwrap() + else { + panic!("Valkey semantic cache should be supported"); + }; + let CacheBackendConfig::ValkeySemantic(valkey) = config.backend else { + panic!("expected Valkey semantic configuration"); + }; + assert_eq!(valkey.similarity_threshold, 0.85); + assert_eq!(valkey.index_name, "semantic_idx"); + assert_eq!(valkey.embedding_model, "text-embedding-3-small"); + assert_eq!(valkey.connection.host, "cache.internal"); + assert_eq!(valkey.connection.port, 6390); + assert_eq!(valkey.connection.database, 2); + assert_eq!(valkey.connection.pool_size, 12); + assert_eq!(valkey.connection.protocol, RedisProtocol::Resp2); + assert!(valkey.connection.tls.is_none()); + }); + } + #[test] fn dynamic_redis_auth_stays_on_python() { Python::initialize(); diff --git a/tests/test_litellm_rust/test_valkey_semantic_cache_native.py b/tests/test_litellm_rust/test_valkey_semantic_cache_native.py index 00037e6e29f..dc4ede8ea30 100644 --- a/tests/test_litellm_rust/test_valkey_semantic_cache_native.py +++ b/tests/test_litellm_rust/test_valkey_semantic_cache_native.py @@ -1,4 +1,7 @@ +import hashlib import os +import struct +import time from collections.abc import Generator, Mapping from types import SimpleNamespace from typing import Final, cast @@ -36,24 +39,32 @@ def index_name(valkey_url: str) -> Generator[str]: client.close() -def _request() -> dict[str, object]: +def _request(prompt: str = "semantic cache prompt") -> dict[str, object]: return { "key": {"preset": "key"}, - "messages": [{"role": "user", "content": "semantic cache prompt"}], + "messages": [{"role": "user", "content": prompt}], } -def _backend(url: str, index_name: str) -> ValkeySemanticCache: +def _backend( + url: str, + index_name: str, + embeddings: Mapping[str, list[float]] | None = None, +) -> ValkeySemanticCache: + vectors: Final = embeddings or {"semantic cache prompt": [1.0, 0.0]} backend: Final = ValkeySemanticCache( redis_url=url, similarity_threshold=0.8, index_name=index_name, ) - backend._get_embedding = lambda prompt, metadata=None: [1.0, 0.0] + + def embed(prompt: str, metadata: Mapping[str, object] | None = None) -> list[float]: + return vectors[prompt] async def async_embedding(prompt: str, metadata: dict[str, object] | None = None) -> list[float]: - return [1.0, 0.0] + return vectors[prompt] + backend._get_embedding = embed backend._get_async_embedding = async_embedding return backend @@ -147,3 +158,122 @@ def test_batch_lookup_is_unsupported( binding: Final = _native._CacheTestResolver(SimpleNamespace(cache=handle)).resolve() with pytest.raises(NotImplementedError): binding.lookup_batch([_request()]) + + +def test_ttl_expiry( + valkey_url: str, + index_name: str, +) -> None: + backend: Final = _backend(valkey_url, index_name) + handle: Final = _native._CacheTestHandle.valkey_semantic(valkey_url, 0.8, index_name, backend) + binding: Final = _native._CacheTestResolver(SimpleNamespace(cache=handle)).resolve() + binding.store({**_request(), "ttl_seconds": 1.0}, {"answer": "expires"}) + client: Final = redis.Redis.from_url(valkey_url) + documents: Final = list(client.scan_iter(f"{index_name}:*")) + assert len(documents) == 1 + assert client.ttl(documents[0]) > 0 + time.sleep(1.5) + assert binding.lookup(_request()) is None + + +def test_no_ttl_is_persistent_and_python_reads_native_value( + valkey_url: str, + index_name: str, +) -> None: + backend: Final = _backend(valkey_url, index_name) + handle: Final = _native._CacheTestHandle.valkey_semantic(valkey_url, 0.8, index_name, backend) + binding: Final = _native._CacheTestResolver(SimpleNamespace(cache=handle)).resolve() + response: Final = {"answer": "persistent"} + binding.store(_request(), response) + client: Final = redis.Redis.from_url(valkey_url) + documents: Final = list(client.scan_iter(f"{index_name}:*")) + assert len(documents) == 1 + assert client.ttl(documents[0]) == -1 + cached: Final = cast(Mapping[str, object], backend.get_cache("key", messages=_request()["messages"])) + assert cached["response"] == response + + +def test_below_threshold_misses_on_native_and_python( + valkey_url: str, + index_name: str, +) -> None: + backend: Final = _backend( + valkey_url, + index_name, + {"prompt A": [1.0, 0.0], "prompt B": [0.0, 1.0]}, + ) + handle: Final = _native._CacheTestHandle.valkey_semantic(valkey_url, 0.8, index_name, backend) + binding: Final = _native._CacheTestResolver(SimpleNamespace(cache=handle)).resolve() + binding.store(_request("prompt A"), {"answer": "A"}) + assert binding.lookup(_request("prompt B")) is None + assert backend.get_cache("key", messages=_request("prompt B")["messages"]) is None + + +def test_malformed_entry_is_a_miss_on_native_and_python( + valkey_url: str, + index_name: str, +) -> None: + backend: Final = _backend(valkey_url, index_name) + client: Final = redis.Redis.from_url(valkey_url) + scope: Final = hashlib.sha256(b"key").hexdigest() + document: Final = f"{index_name}:{scope}:{uuid4().hex}" + client.hset( + document, + mapping={ + "litellm_cache_key": scope, + "prompt": "semantic cache prompt", + "response": "not json", + "embedding": struct.pack("<2f", 1.0, 0.0), + }, + ) + handle: Final = _native._CacheTestHandle.valkey_semantic(valkey_url, 0.8, index_name, backend) + binding: Final = _native._CacheTestResolver(SimpleNamespace(cache=handle)).resolve() + assert binding.lookup(_request()) is None + assert backend.get_cache("key", messages=_request()["messages"]) is None + + +async def test_async_store_batch_and_lookup( + valkey_url: str, + index_name: str, +) -> None: + backend: Final = _backend( + valkey_url, + index_name, + {"prompt A": [1.0, 0.0], "prompt B": [0.0, 1.0]}, + ) + handle: Final = _native._CacheTestHandle.valkey_semantic(valkey_url, 0.8, index_name, backend) + binding: Final = _native._CacheTestResolver(SimpleNamespace(cache=handle)).resolve() + requests: Final = [_request("prompt A"), _request("prompt B")] + responses: Final = [{"answer": "A"}, {"answer": "B"}] + await binding.async_store_batch(requests, responses) + assert await binding.async_lookup(requests[0]) == responses[0] + assert await binding.async_lookup(requests[1]) == responses[1] + + +def test_subclass_backend_falls_back_to_python( + valkey_url: str, + index_name: str, +) -> None: + class Custom(ValkeySemanticCache): + pass + + facade: Final = Cache( + type=LiteLLMCacheType.VALKEY_SEMANTIC, + redis_url=valkey_url, + similarity_threshold=0.8, + valkey_semantic_cache_index_name=index_name, + ) + facade.cache = Custom(redis_url=valkey_url, similarity_threshold=0.8, index_name=index_name) + resolver: Final = _native._CacheTestResolver(SimpleNamespace(cache=facade)) + assert resolver.resolve().kind == "python_callback" + + +async def test_ping_maps_unsupported_native_operation_to_not_implemented( + valkey_url: str, + index_name: str, +) -> None: + backend: Final = _backend(valkey_url, index_name) + handle: Final = _native._CacheTestHandle.valkey_semantic(valkey_url, 0.8, index_name, backend) + binding: Final = _native._CacheTestResolver(SimpleNamespace(cache=handle)).resolve() + with pytest.raises(NotImplementedError): + await binding.ping()