"""
.. warning::
Beta Feature!
**Cache** provides an optional caching layer for LLMs.
Cache is useful for two reasons:
- It can save you money by reducing the number of API calls you make to the LLM
provider if you're often requesting the same completion multiple times.
- It can speed up your application by reducing the number of API calls you make
to the LLM provider.
Cache directly competes with Memory. See documentation for Pros and Cons.
**Class hierarchy:**
.. code-block::
BaseCache --> <name>Cache # Examples: InMemoryCache, RedisCache, GPTCache
"""
from __future__ import annotations
import hashlib
import inspect
import json
import logging
import uuid
import warnings
from datetime import timedelta
from functools import lru_cache
from typing import (
TYPE_CHECKING,
Any,
Callable,
Dict,
List,
Optional,
Tuple,
Type,
Union,
cast,
)
from sqlalchemy import Column, Integer, Row, String, create_engine, select
from sqlalchemy.engine.base import Engine
from sqlalchemy.orm import Session
try:
from sqlalchemy.orm import declarative_base
except ImportError:
from sqlalchemy.ext.declarative import declarative_base
from langchain_core.caches import RETURN_VAL_TYPE, BaseCache
from langchain_core.embeddings import Embeddings
from langchain_core.language_models.llms import LLM, get_prompts
from langchain_core.load.dump import dumps
from langchain_core.load.load import loads
from langchain_core.outputs import ChatGeneration, Generation
from langchain_core.utils import get_from_env
from langchain_community.vectorstores.redis import Redis as RedisVectorstore
logger = logging.getLogger(__file__)
if TYPE_CHECKING:
import momento
from cassandra.cluster import Session as CassandraSession
def _hash(_input: str) -> str:
"""Use a deterministic hashing approach."""
return hashlib.md5(_input.encode()).hexdigest()
def _dump_generations_to_json(generations: RETURN_VAL_TYPE) -> str:
"""Dump generations to json.
Args:
generations (RETURN_VAL_TYPE): A list of language model generations.
Returns:
str: Json representing a list of generations.
Warning: would not work well with arbitrary subclasses of `Generation`
"""
return json.dumps([generation.dict() for generation in generations])
def _load_generations_from_json(generations_json: str) -> RETURN_VAL_TYPE:
"""Load generations from json.
Args:
generations_json (str): A string of json representing a list of generations.
Raises:
ValueError: Could not decode json string to list of generations.
Returns:
RETURN_VAL_TYPE: A list of generations.
Warning: would not work well with arbitrary subclasses of `Generation`
"""
try:
results = json.loads(generations_json)
return [Generation(**generation_dict) for generation_dict in results]
except json.JSONDecodeError:
raise ValueError(
f"Could not decode json to list of generations: {generations_json}"
)
def _dumps_generations(generations: RETURN_VAL_TYPE) -> str:
"""
Serialization for generic RETURN_VAL_TYPE, i.e. sequence of `Generation`
Args:
generations (RETURN_VAL_TYPE): A list of language model generations.
Returns:
str: a single string representing a list of generations.
This function (+ its counterpart `_loads_generations`) rely on
the dumps/loads pair with Reviver, so are able to deal
with all subclasses of Generation.
Each item in the list can be `dumps`ed to a string,
then we make the whole list of strings into a json-dumped.
"""
return json.dumps([dumps(_item) for _item in generations])
def _loads_generations(generations_str: str) -> Union[RETURN_VAL_TYPE, None]:
"""
Deserialization of a string into a generic RETURN_VAL_TYPE
(i.e. a sequence of `Generation`).
See `_dumps_generations`, the inverse of this function.
Args:
generations_str (str): A string representing a list of generations.
Compatible with the legacy cache-blob format
Does not raise exceptions for malformed entries, just logs a warning
and returns none: the caller should be prepared for such a cache miss.
Returns:
RETURN_VAL_TYPE: A list of generations.
"""
try:
generations = [loads(_item_str) for _item_str in json.loads(generations_str)]
return generations
except (json.JSONDecodeError, TypeError):
# deferring the (soft) handling to after the legacy-format attempt
pass
try:
gen_dicts = json.loads(generations_str)
# not relying on `_load_generations_from_json` (which could disappear):
generations = [Generation(**generation_dict) for generation_dict in gen_dicts]
logger.warning(
f"Legacy 'Generation' cached blob encountered: '{generations_str}'"
)
return generations
except (json.JSONDecodeError, TypeError):
logger.warning(
f"Malformed/unparsable cached blob encountered: '{generations_str}'"
)
return None
[docs]class InMemoryCache(BaseCache):
"""Cache that stores things in memory."""
[docs] def __init__(self) -> None:
"""Initialize with empty cache."""
self._cache: Dict[Tuple[str, str], RETURN_VAL_TYPE] = {}
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
return self._cache.get((prompt, llm_string), None)
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
self._cache[(prompt, llm_string)] = return_val
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear cache."""
self._cache = {}
Base = declarative_base()
[docs]class FullLLMCache(Base): # type: ignore
"""SQLite table for full LLM Cache (all generations)."""
__tablename__ = "full_llm_cache"
prompt = Column(String, primary_key=True)
llm = Column(String, primary_key=True)
idx = Column(Integer, primary_key=True)
response = Column(String)
[docs]class SQLAlchemyCache(BaseCache):
"""Cache that uses SQAlchemy as a backend."""
[docs] def __init__(self, engine: Engine, cache_schema: Type[FullLLMCache] = FullLLMCache):
"""Initialize by creating all tables."""
self.engine = engine
self.cache_schema = cache_schema
self.cache_schema.metadata.create_all(self.engine)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
stmt = (
select(self.cache_schema.response)
.where(self.cache_schema.prompt == prompt) # type: ignore
.where(self.cache_schema.llm == llm_string)
.order_by(self.cache_schema.idx)
)
with Session(self.engine) as session:
rows = session.execute(stmt).fetchall()
if rows:
try:
return [loads(row[0]) for row in rows]
except Exception:
logger.warning(
"Retrieving a cache value that could not be deserialized "
"properly. This is likely due to the cache being in an "
"older format. Please recreate your cache to avoid this "
"error."
)
# In a previous life we stored the raw text directly
# in the table, so assume it's in that format.
return [Generation(text=row[0]) for row in rows]
return None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update based on prompt and llm_string."""
items = [
self.cache_schema(prompt=prompt, llm=llm_string, response=dumps(gen), idx=i)
for i, gen in enumerate(return_val)
]
with Session(self.engine) as session, session.begin():
for item in items:
session.merge(item)
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear cache."""
with Session(self.engine) as session:
session.query(self.cache_schema).delete()
session.commit()
[docs]class SQLiteCache(SQLAlchemyCache):
"""Cache that uses SQLite as a backend."""
[docs] def __init__(self, database_path: str = ".langchain.db"):
"""Initialize by creating the engine and all tables."""
engine = create_engine(f"sqlite:///{database_path}")
super().__init__(engine)
[docs]class UpstashRedisCache(BaseCache):
"""Cache that uses Upstash Redis as a backend."""
[docs] def __init__(self, redis_: Any, *, ttl: Optional[int] = None):
"""
Initialize an instance of UpstashRedisCache.
This method initializes an object with Upstash Redis caching capabilities.
It takes a `redis_` parameter, which should be an instance of an Upstash Redis
client class, allowing the object to interact with Upstash Redis
server for caching purposes.
Parameters:
redis_: An instance of Upstash Redis client class
(e.g., Redis) used for caching.
This allows the object to communicate with
Redis server for caching operations on.
ttl (int, optional): Time-to-live (TTL) for cached items in seconds.
If provided, it sets the time duration for how long cached
items will remain valid. If not provided, cached items will not
have an automatic expiration.
"""
try:
from upstash_redis import Redis
except ImportError:
raise ValueError(
"Could not import upstash_redis python package. "
"Please install it with `pip install upstash_redis`."
)
if not isinstance(redis_, Redis):
raise ValueError("Please pass in Upstash Redis object.")
self.redis = redis_
self.ttl = ttl
def _key(self, prompt: str, llm_string: str) -> str:
"""Compute key from prompt and llm_string"""
return _hash(prompt + llm_string)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
generations = []
# Read from a HASH
results = self.redis.hgetall(self._key(prompt, llm_string))
if results:
for _, text in results.items():
generations.append(Generation(text=text))
return generations if generations else None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
for gen in return_val:
if not isinstance(gen, Generation):
raise ValueError(
"UpstashRedisCache supports caching of normal LLM generations, "
f"got {type(gen)}"
)
if isinstance(gen, ChatGeneration):
warnings.warn(
"NOTE: Generation has not been cached. UpstashRedisCache does not"
" support caching ChatModel outputs."
)
return
# Write to a HASH
key = self._key(prompt, llm_string)
mapping = {
str(idx): generation.text for idx, generation in enumerate(return_val)
}
self.redis.hset(key=key, values=mapping)
if self.ttl is not None:
self.redis.expire(key, self.ttl)
[docs] def clear(self, **kwargs: Any) -> None:
"""
Clear cache. If `asynchronous` is True, flush asynchronously.
This flushes the *whole* db.
"""
asynchronous = kwargs.get("asynchronous", False)
if asynchronous:
asynchronous = "ASYNC"
else:
asynchronous = "SYNC"
self.redis.flushdb(flush_type=asynchronous)
[docs]class RedisCache(BaseCache):
"""Cache that uses Redis as a backend."""
[docs] def __init__(self, redis_: Any, *, ttl: Optional[int] = None):
"""
Initialize an instance of RedisCache.
This method initializes an object with Redis caching capabilities.
It takes a `redis_` parameter, which should be an instance of a Redis
client class, allowing the object to interact with a Redis
server for caching purposes.
Parameters:
redis_ (Any): An instance of a Redis client class
(e.g., redis.Redis) used for caching.
This allows the object to communicate with a
Redis server for caching operations.
ttl (int, optional): Time-to-live (TTL) for cached items in seconds.
If provided, it sets the time duration for how long cached
items will remain valid. If not provided, cached items will not
have an automatic expiration.
"""
try:
from redis import Redis
except ImportError:
raise ValueError(
"Could not import redis python package. "
"Please install it with `pip install redis`."
)
if not isinstance(redis_, Redis):
raise ValueError("Please pass in Redis object.")
self.redis = redis_
self.ttl = ttl
def _key(self, prompt: str, llm_string: str) -> str:
"""Compute key from prompt and llm_string"""
return _hash(prompt + llm_string)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
generations = []
# Read from a Redis HASH
results = self.redis.hgetall(self._key(prompt, llm_string))
if results:
for _, text in results.items():
try:
generations.append(loads(text))
except Exception:
logger.warning(
"Retrieving a cache value that could not be deserialized "
"properly. This is likely due to the cache being in an "
"older format. Please recreate your cache to avoid this "
"error."
)
# In a previous life we stored the raw text directly
# in the table, so assume it's in that format.
generations.append(Generation(text=text))
return generations if generations else None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
for gen in return_val:
if not isinstance(gen, Generation):
raise ValueError(
"RedisCache only supports caching of normal LLM generations, "
f"got {type(gen)}"
)
# Write to a Redis HASH
key = self._key(prompt, llm_string)
with self.redis.pipeline() as pipe:
pipe.hset(
key,
mapping={
str(idx): dumps(generation)
for idx, generation in enumerate(return_val)
},
)
if self.ttl is not None:
pipe.expire(key, self.ttl)
pipe.execute()
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear cache. If `asynchronous` is True, flush asynchronously."""
asynchronous = kwargs.get("asynchronous", False)
self.redis.flushdb(asynchronous=asynchronous, **kwargs)
[docs]class RedisSemanticCache(BaseCache):
"""Cache that uses Redis as a vector-store backend."""
# TODO - implement a TTL policy in Redis
DEFAULT_SCHEMA = {
"content_key": "prompt",
"text": [
{"name": "prompt"},
],
"extra": [{"name": "return_val"}, {"name": "llm_string"}],
}
[docs] def __init__(
self, redis_url: str, embedding: Embeddings, score_threshold: float = 0.2
):
"""Initialize by passing in the `init` GPTCache func
Args:
redis_url (str): URL to connect to Redis.
embedding (Embedding): Embedding provider for semantic encoding and search.
score_threshold (float, 0.2):
Example:
.. code-block:: python
from langchain_community.globals import set_llm_cache
from langchain_community.cache import RedisSemanticCache
from langchain_community.embeddings import OpenAIEmbeddings
set_llm_cache(RedisSemanticCache(
redis_url="redis://localhost:6379",
embedding=OpenAIEmbeddings()
))
"""
self._cache_dict: Dict[str, RedisVectorstore] = {}
self.redis_url = redis_url
self.embedding = embedding
self.score_threshold = score_threshold
def _index_name(self, llm_string: str) -> str:
hashed_index = _hash(llm_string)
return f"cache:{hashed_index}"
def _get_llm_cache(self, llm_string: str) -> RedisVectorstore:
index_name = self._index_name(llm_string)
# return vectorstore client for the specific llm string
if index_name in self._cache_dict:
return self._cache_dict[index_name]
# create new vectorstore client for the specific llm string
try:
self._cache_dict[index_name] = RedisVectorstore.from_existing_index(
embedding=self.embedding,
index_name=index_name,
redis_url=self.redis_url,
schema=cast(Dict, self.DEFAULT_SCHEMA),
)
except ValueError:
redis = RedisVectorstore(
embedding=self.embedding,
index_name=index_name,
redis_url=self.redis_url,
index_schema=cast(Dict, self.DEFAULT_SCHEMA),
)
_embedding = self.embedding.embed_query(text="test")
redis._create_index_if_not_exist(dim=len(_embedding))
self._cache_dict[index_name] = redis
return self._cache_dict[index_name]
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear semantic cache for a given llm_string."""
index_name = self._index_name(kwargs["llm_string"])
if index_name in self._cache_dict:
self._cache_dict[index_name].drop_index(
index_name=index_name, delete_documents=True, redis_url=self.redis_url
)
del self._cache_dict[index_name]
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
llm_cache = self._get_llm_cache(llm_string)
generations: List = []
# Read from a Hash
results = llm_cache.similarity_search(
query=prompt,
k=1,
distance_threshold=self.score_threshold,
)
if results:
for document in results:
try:
generations.extend(loads(document.metadata["return_val"]))
except Exception:
logger.warning(
"Retrieving a cache value that could not be deserialized "
"properly. This is likely due to the cache being in an "
"older format. Please recreate your cache to avoid this "
"error."
)
# In a previous life we stored the raw text directly
# in the table, so assume it's in that format.
generations.extend(
_load_generations_from_json(document.metadata["return_val"])
)
return generations if generations else None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
for gen in return_val:
if not isinstance(gen, Generation):
raise ValueError(
"RedisSemanticCache only supports caching of "
f"normal LLM generations, got {type(gen)}"
)
llm_cache = self._get_llm_cache(llm_string)
metadata = {
"llm_string": llm_string,
"prompt": prompt,
"return_val": dumps([g for g in return_val]),
}
llm_cache.add_texts(texts=[prompt], metadatas=[metadata])
[docs]class GPTCache(BaseCache):
"""Cache that uses GPTCache as a backend."""
[docs] def __init__(
self,
init_func: Union[
Callable[[Any, str], None], Callable[[Any], None], None
] = None,
):
"""Initialize by passing in init function (default: `None`).
Args:
init_func (Optional[Callable[[Any], None]]): init `GPTCache` function
(default: `None`)
Example:
.. code-block:: python
# Initialize GPTCache with a custom init function
import gptcache
from gptcache.processor.pre import get_prompt
from gptcache.manager.factory import get_data_manager
from langchain_community.globals import set_llm_cache
# Avoid multiple caches using the same file,
causing different llm model caches to affect each other
def init_gptcache(cache_obj: gptcache.Cache, llm str):
cache_obj.init(
pre_embedding_func=get_prompt,
data_manager=manager_factory(
manager="map",
data_dir=f"map_cache_{llm}"
),
)
set_llm_cache(GPTCache(init_gptcache))
"""
try:
import gptcache # noqa: F401
except ImportError:
raise ImportError(
"Could not import gptcache python package. "
"Please install it with `pip install gptcache`."
)
self.init_gptcache_func: Union[
Callable[[Any, str], None], Callable[[Any], None], None
] = init_func
self.gptcache_dict: Dict[str, Any] = {}
def _new_gptcache(self, llm_string: str) -> Any:
"""New gptcache object"""
from gptcache import Cache
from gptcache.manager.factory import get_data_manager
from gptcache.processor.pre import get_prompt
_gptcache = Cache()
if self.init_gptcache_func is not None:
sig = inspect.signature(self.init_gptcache_func)
if len(sig.parameters) == 2:
self.init_gptcache_func(_gptcache, llm_string) # type: ignore[call-arg]
else:
self.init_gptcache_func(_gptcache) # type: ignore[call-arg]
else:
_gptcache.init(
pre_embedding_func=get_prompt,
data_manager=get_data_manager(data_path=llm_string),
)
self.gptcache_dict[llm_string] = _gptcache
return _gptcache
def _get_gptcache(self, llm_string: str) -> Any:
"""Get a cache object.
When the corresponding llm model cache does not exist, it will be created."""
_gptcache = self.gptcache_dict.get(llm_string, None)
if not _gptcache:
_gptcache = self._new_gptcache(llm_string)
return _gptcache
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up the cache data.
First, retrieve the corresponding cache object using the `llm_string` parameter,
and then retrieve the data from the cache based on the `prompt`.
"""
from gptcache.adapter.api import get
_gptcache = self._get_gptcache(llm_string)
res = get(prompt, cache_obj=_gptcache)
if res:
return [
Generation(**generation_dict) for generation_dict in json.loads(res)
]
return None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache.
First, retrieve the corresponding cache object using the `llm_string` parameter,
and then store the `prompt` and `return_val` in the cache object.
"""
for gen in return_val:
if not isinstance(gen, Generation):
raise ValueError(
"GPTCache only supports caching of normal LLM generations, "
f"got {type(gen)}"
)
from gptcache.adapter.api import put
_gptcache = self._get_gptcache(llm_string)
handled_data = json.dumps([generation.dict() for generation in return_val])
put(prompt, handled_data, cache_obj=_gptcache)
return None
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear cache."""
from gptcache import Cache
for gptcache_instance in self.gptcache_dict.values():
gptcache_instance = cast(Cache, gptcache_instance)
gptcache_instance.flush()
self.gptcache_dict.clear()
def _ensure_cache_exists(cache_client: momento.CacheClient, cache_name: str) -> None:
"""Create cache if it doesn't exist.
Raises:
SdkException: Momento service or network error
Exception: Unexpected response
"""
from momento.responses import CreateCache
create_cache_response = cache_client.create_cache(cache_name)
if isinstance(create_cache_response, CreateCache.Success) or isinstance(
create_cache_response, CreateCache.CacheAlreadyExists
):
return None
elif isinstance(create_cache_response, CreateCache.Error):
raise create_cache_response.inner_exception
else:
raise Exception(f"Unexpected response cache creation: {create_cache_response}")
def _validate_ttl(ttl: Optional[timedelta]) -> None:
if ttl is not None and ttl <= timedelta(seconds=0):
raise ValueError(f"ttl must be positive but was {ttl}.")
[docs]class MomentoCache(BaseCache):
"""Cache that uses Momento as a backend. See https://gomomento.com/"""
[docs] def __init__(
self,
cache_client: momento.CacheClient,
cache_name: str,
*,
ttl: Optional[timedelta] = None,
ensure_cache_exists: bool = True,
):
"""Instantiate a prompt cache using Momento as a backend.
Note: to instantiate the cache client passed to MomentoCache,
you must have a Momento account. See https://gomomento.com/.
Args:
cache_client (CacheClient): The Momento cache client.
cache_name (str): The name of the cache to use to store the data.
ttl (Optional[timedelta], optional): The time to live for the cache items.
Defaults to None, ie use the client default TTL.
ensure_cache_exists (bool, optional): Create the cache if it doesn't
exist. Defaults to True.
Raises:
ImportError: Momento python package is not installed.
TypeError: cache_client is not of type momento.CacheClientObject
ValueError: ttl is non-null and non-negative
"""
try:
from momento import CacheClient
except ImportError:
raise ImportError(
"Could not import momento python package. "
"Please install it with `pip install momento`."
)
if not isinstance(cache_client, CacheClient):
raise TypeError("cache_client must be a momento.CacheClient object.")
_validate_ttl(ttl)
if ensure_cache_exists:
_ensure_cache_exists(cache_client, cache_name)
self.cache_client = cache_client
self.cache_name = cache_name
self.ttl = ttl
[docs] @classmethod
def from_client_params(
cls,
cache_name: str,
ttl: timedelta,
*,
configuration: Optional[momento.config.Configuration] = None,
api_key: Optional[str] = None,
auth_token: Optional[str] = None, # for backwards compatibility
**kwargs: Any,
) -> MomentoCache:
"""Construct cache from CacheClient parameters."""
try:
from momento import CacheClient, Configurations, CredentialProvider
except ImportError:
raise ImportError(
"Could not import momento python package. "
"Please install it with `pip install momento`."
)
if configuration is None:
configuration = Configurations.Laptop.v1()
# Try checking `MOMENTO_AUTH_TOKEN` first for backwards compatibility
try:
api_key = auth_token or get_from_env("auth_token", "MOMENTO_AUTH_TOKEN")
except ValueError:
api_key = api_key or get_from_env("api_key", "MOMENTO_API_KEY")
credentials = CredentialProvider.from_string(api_key)
cache_client = CacheClient(configuration, credentials, default_ttl=ttl)
return cls(cache_client, cache_name, ttl=ttl, **kwargs)
def __key(self, prompt: str, llm_string: str) -> str:
"""Compute cache key from prompt and associated model and settings.
Args:
prompt (str): The prompt run through the language model.
llm_string (str): The language model version and settings.
Returns:
str: The cache key.
"""
return _hash(prompt + llm_string)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Lookup llm generations in cache by prompt and associated model and settings.
Args:
prompt (str): The prompt run through the language model.
llm_string (str): The language model version and settings.
Raises:
SdkException: Momento service or network error
Returns:
Optional[RETURN_VAL_TYPE]: A list of language model generations.
"""
from momento.responses import CacheGet
generations: RETURN_VAL_TYPE = []
get_response = self.cache_client.get(
self.cache_name, self.__key(prompt, llm_string)
)
if isinstance(get_response, CacheGet.Hit):
value = get_response.value_string
generations = _load_generations_from_json(value)
elif isinstance(get_response, CacheGet.Miss):
pass
elif isinstance(get_response, CacheGet.Error):
raise get_response.inner_exception
return generations if generations else None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Store llm generations in cache.
Args:
prompt (str): The prompt run through the language model.
llm_string (str): The language model string.
return_val (RETURN_VAL_TYPE): A list of language model generations.
Raises:
SdkException: Momento service or network error
Exception: Unexpected response
"""
for gen in return_val:
if not isinstance(gen, Generation):
raise ValueError(
"Momento only supports caching of normal LLM generations, "
f"got {type(gen)}"
)
key = self.__key(prompt, llm_string)
value = _dump_generations_to_json(return_val)
set_response = self.cache_client.set(self.cache_name, key, value, self.ttl)
from momento.responses import CacheSet
if isinstance(set_response, CacheSet.Success):
pass
elif isinstance(set_response, CacheSet.Error):
raise set_response.inner_exception
else:
raise Exception(f"Unexpected response: {set_response}")
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear the cache.
Raises:
SdkException: Momento service or network error
"""
from momento.responses import CacheFlush
flush_response = self.cache_client.flush_cache(self.cache_name)
if isinstance(flush_response, CacheFlush.Success):
pass
elif isinstance(flush_response, CacheFlush.Error):
raise flush_response.inner_exception
CASSANDRA_CACHE_DEFAULT_TABLE_NAME = "langchain_llm_cache"
CASSANDRA_CACHE_DEFAULT_TTL_SECONDS = None
[docs]class CassandraCache(BaseCache):
"""
Cache that uses Cassandra / Astra DB as a backend.
It uses a single Cassandra table.
The lookup keys (which get to form the primary key) are:
- prompt, a string
- llm_string, a deterministic str representation of the model parameters.
(needed to prevent collisions same-prompt-different-model collisions)
"""
[docs] def __init__(
self,
session: Optional[CassandraSession] = None,
keyspace: Optional[str] = None,
table_name: str = CASSANDRA_CACHE_DEFAULT_TABLE_NAME,
ttl_seconds: Optional[int] = CASSANDRA_CACHE_DEFAULT_TTL_SECONDS,
skip_provisioning: bool = False,
):
"""
Initialize with a ready session and a keyspace name.
Args:
session (cassandra.cluster.Session): an open Cassandra session
keyspace (str): the keyspace to use for storing the cache
table_name (str): name of the Cassandra table to use as cache
ttl_seconds (optional int): time-to-live for cache entries
(default: None, i.e. forever)
"""
try:
from cassio.table import ElasticCassandraTable
except (ImportError, ModuleNotFoundError):
raise ValueError(
"Could not import cassio python package. "
"Please install it with `pip install cassio`."
)
self.session = session
self.keyspace = keyspace
self.table_name = table_name
self.ttl_seconds = ttl_seconds
self.kv_cache = ElasticCassandraTable(
session=self.session,
keyspace=self.keyspace,
table=self.table_name,
keys=["llm_string", "prompt"],
primary_key_type=["TEXT", "TEXT"],
ttl_seconds=self.ttl_seconds,
skip_provisioning=skip_provisioning,
)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
item = self.kv_cache.get(
llm_string=_hash(llm_string),
prompt=_hash(prompt),
)
if item is not None:
generations = _loads_generations(item["body_blob"])
# this protects against malformed cached items:
if generations is not None:
return generations
else:
return None
else:
return None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
blob = _dumps_generations(return_val)
self.kv_cache.put(
llm_string=_hash(llm_string),
prompt=_hash(prompt),
body_blob=blob,
)
[docs] def delete_through_llm(
self, prompt: str, llm: LLM, stop: Optional[List[str]] = None
) -> None:
"""
A wrapper around `delete` with the LLM being passed.
In case the llm(prompt) calls have a `stop` param, you should pass it here
"""
llm_string = get_prompts(
{**llm.dict(), **{"stop": stop}},
[],
)[1]
return self.delete(prompt, llm_string=llm_string)
[docs] def delete(self, prompt: str, llm_string: str) -> None:
"""Evict from cache if there's an entry."""
return self.kv_cache.delete(
llm_string=_hash(llm_string),
prompt=_hash(prompt),
)
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear cache. This is for all LLMs at once."""
self.kv_cache.clear()
CASSANDRA_SEMANTIC_CACHE_DEFAULT_DISTANCE_METRIC = "dot"
CASSANDRA_SEMANTIC_CACHE_DEFAULT_SCORE_THRESHOLD = 0.85
CASSANDRA_SEMANTIC_CACHE_DEFAULT_TABLE_NAME = "langchain_llm_semantic_cache"
CASSANDRA_SEMANTIC_CACHE_DEFAULT_TTL_SECONDS = None
CASSANDRA_SEMANTIC_CACHE_EMBEDDING_CACHE_SIZE = 16
[docs]class CassandraSemanticCache(BaseCache):
"""
Cache that uses Cassandra as a vector-store backend for semantic
(i.e. similarity-based) lookup.
It uses a single (vector) Cassandra table and stores, in principle,
cached values from several LLMs, so the LLM's llm_string is part
of the rows' primary keys.
The similarity is based on one of several distance metrics (default: "dot").
If choosing another metric, the default threshold is to be re-tuned accordingly.
"""
[docs] def __init__(
self,
session: Optional[CassandraSession],
keyspace: Optional[str],
embedding: Embeddings,
table_name: str = CASSANDRA_SEMANTIC_CACHE_DEFAULT_TABLE_NAME,
distance_metric: str = CASSANDRA_SEMANTIC_CACHE_DEFAULT_DISTANCE_METRIC,
score_threshold: float = CASSANDRA_SEMANTIC_CACHE_DEFAULT_SCORE_THRESHOLD,
ttl_seconds: Optional[int] = CASSANDRA_SEMANTIC_CACHE_DEFAULT_TTL_SECONDS,
skip_provisioning: bool = False,
):
"""
Initialize the cache with all relevant parameters.
Args:
session (cassandra.cluster.Session): an open Cassandra session
keyspace (str): the keyspace to use for storing the cache
embedding (Embedding): Embedding provider for semantic
encoding and search.
table_name (str): name of the Cassandra (vector) table
to use as cache
distance_metric (str, 'dot'): which measure to adopt for
similarity searches
score_threshold (optional float): numeric value to use as
cutoff for the similarity searches
ttl_seconds (optional int): time-to-live for cache entries
(default: None, i.e. forever)
The default score threshold is tuned to the default metric.
Tune it carefully yourself if switching to another distance metric.
"""
try:
from cassio.table import MetadataVectorCassandraTable
except (ImportError, ModuleNotFoundError):
raise ValueError(
"Could not import cassio python package. "
"Please install it with `pip install cassio`."
)
self.session = session
self.keyspace = keyspace
self.embedding = embedding
self.table_name = table_name
self.distance_metric = distance_metric
self.score_threshold = score_threshold
self.ttl_seconds = ttl_seconds
# The contract for this class has separate lookup and update:
# in order to spare some embedding calculations we cache them between
# the two calls.
# Note: each instance of this class has its own `_get_embedding` with
# its own lru.
@lru_cache(maxsize=CASSANDRA_SEMANTIC_CACHE_EMBEDDING_CACHE_SIZE)
def _cache_embedding(text: str) -> List[float]:
return self.embedding.embed_query(text=text)
self._get_embedding = _cache_embedding
self.embedding_dimension = self._get_embedding_dimension()
self.table = MetadataVectorCassandraTable(
session=self.session,
keyspace=self.keyspace,
table=self.table_name,
primary_key_type=["TEXT"],
vector_dimension=self.embedding_dimension,
ttl_seconds=self.ttl_seconds,
metadata_indexing=("allow", {"_llm_string_hash"}),
skip_provisioning=skip_provisioning,
)
def _get_embedding_dimension(self) -> int:
return len(self._get_embedding(text="This is a sample sentence."))
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
embedding_vector = self._get_embedding(text=prompt)
llm_string_hash = _hash(llm_string)
body = _dumps_generations(return_val)
metadata = {
"_prompt": prompt,
"_llm_string_hash": llm_string_hash,
}
row_id = f"{_hash(prompt)}-{llm_string_hash}"
#
self.table.put(
body_blob=body,
vector=embedding_vector,
row_id=row_id,
metadata=metadata,
)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
hit_with_id = self.lookup_with_id(prompt, llm_string)
if hit_with_id is not None:
return hit_with_id[1]
else:
return None
[docs] def lookup_with_id(
self, prompt: str, llm_string: str
) -> Optional[Tuple[str, RETURN_VAL_TYPE]]:
"""
Look up based on prompt and llm_string.
If there are hits, return (document_id, cached_entry)
"""
prompt_embedding: List[float] = self._get_embedding(text=prompt)
hits = list(
self.table.metric_ann_search(
vector=prompt_embedding,
metadata={"_llm_string_hash": _hash(llm_string)},
n=1,
metric=self.distance_metric,
metric_threshold=self.score_threshold,
)
)
if hits:
hit = hits[0]
generations = _loads_generations(hit["body_blob"])
if generations is not None:
# this protects against malformed cached items:
return (
hit["row_id"],
generations,
)
else:
return None
else:
return None
[docs] def lookup_with_id_through_llm(
self, prompt: str, llm: LLM, stop: Optional[List[str]] = None
) -> Optional[Tuple[str, RETURN_VAL_TYPE]]:
llm_string = get_prompts(
{**llm.dict(), **{"stop": stop}},
[],
)[1]
return self.lookup_with_id(prompt, llm_string=llm_string)
[docs] def delete_by_document_id(self, document_id: str) -> None:
"""
Given this is a "similarity search" cache, an invalidation pattern
that makes sense is first a lookup to get an ID, and then deleting
with that ID. This is for the second step.
"""
self.table.delete(row_id=document_id)
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear the *whole* semantic cache."""
self.table.clear()
[docs]class FullMd5LLMCache(Base): # type: ignore
"""SQLite table for full LLM Cache (all generations)."""
__tablename__ = "full_md5_llm_cache"
id = Column(String, primary_key=True)
prompt_md5 = Column(String, index=True)
llm = Column(String, index=True)
idx = Column(Integer, index=True)
prompt = Column(String)
response = Column(String)
[docs]class SQLAlchemyMd5Cache(BaseCache):
"""Cache that uses SQAlchemy as a backend."""
[docs] def __init__(
self, engine: Engine, cache_schema: Type[FullMd5LLMCache] = FullMd5LLMCache
):
"""Initialize by creating all tables."""
self.engine = engine
self.cache_schema = cache_schema
self.cache_schema.metadata.create_all(self.engine)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
rows = self._search_rows(prompt, llm_string)
if rows:
return [loads(row[0]) for row in rows]
return None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update based on prompt and llm_string."""
self._delete_previous(prompt, llm_string)
prompt_md5 = self.get_md5(prompt)
items = [
self.cache_schema(
id=str(uuid.uuid1()),
prompt=prompt,
prompt_md5=prompt_md5,
llm=llm_string,
response=dumps(gen),
idx=i,
)
for i, gen in enumerate(return_val)
]
with Session(self.engine) as session, session.begin():
for item in items:
session.merge(item)
def _delete_previous(self, prompt: str, llm_string: str) -> None:
stmt = (
select(self.cache_schema.response)
.where(self.cache_schema.prompt_md5 == self.get_md5(prompt)) # type: ignore
.where(self.cache_schema.llm == llm_string)
.where(self.cache_schema.prompt == prompt)
.order_by(self.cache_schema.idx)
)
with Session(self.engine) as session, session.begin():
rows = session.execute(stmt).fetchall()
for item in rows:
session.delete(item)
def _search_rows(self, prompt: str, llm_string: str) -> List[Row]:
prompt_pd5 = self.get_md5(prompt)
stmt = (
select(self.cache_schema.response)
.where(self.cache_schema.prompt_md5 == prompt_pd5) # type: ignore
.where(self.cache_schema.llm == llm_string)
.where(self.cache_schema.prompt == prompt)
.order_by(self.cache_schema.idx)
)
with Session(self.engine) as session:
return session.execute(stmt).fetchall()
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear cache."""
with Session(self.engine) as session:
session.execute(self.cache_schema.delete())
[docs] @staticmethod
def get_md5(input_string: str) -> str:
return hashlib.md5(input_string.encode()).hexdigest()
ASTRA_DB_CACHE_DEFAULT_COLLECTION_NAME = "langchain_astradb_cache"
[docs]class AstraDBCache(BaseCache):
"""
Cache that uses Astra DB as a backend.
It uses a single collection as a kv store
The lookup keys, combined in the _id of the documents, are:
- prompt, a string
- llm_string, a deterministic str representation of the model parameters.
(needed to prevent same-prompt-different-model collisions)
"""
[docs] def __init__(
self,
*,
collection_name: str = ASTRA_DB_CACHE_DEFAULT_COLLECTION_NAME,
token: Optional[str] = None,
api_endpoint: Optional[str] = None,
astra_db_client: Optional[Any] = None, # 'astrapy.db.AstraDB' if passed
namespace: Optional[str] = None,
):
"""
Create an AstraDB cache using a collection for storage.
Args (only keyword-arguments accepted):
collection_name (str): name of the Astra DB collection to create/use.
token (Optional[str]): API token for Astra DB usage.
api_endpoint (Optional[str]): full URL to the API endpoint,
such as "https://<DB-ID>-us-east1.apps.astra.datastax.com".
astra_db_client (Optional[Any]): *alternative to token+api_endpoint*,
you can pass an already-created 'astrapy.db.AstraDB' instance.
namespace (Optional[str]): namespace (aka keyspace) where the
collection is created. Defaults to the database's "default namespace".
"""
try:
from astrapy.db import (
AstraDB as LibAstraDB,
)
except (ImportError, ModuleNotFoundError):
raise ImportError(
"Could not import a recent astrapy python package. "
"Please install it with `pip install --upgrade astrapy`."
)
# Conflicting-arg checks:
if astra_db_client is not None:
if token is not None or api_endpoint is not None:
raise ValueError(
"You cannot pass 'astra_db_client' to AstraDB if passing "
"'token' and 'api_endpoint'."
)
self.collection_name = collection_name
self.token = token
self.api_endpoint = api_endpoint
self.namespace = namespace
if astra_db_client is not None:
self.astra_db = astra_db_client
else:
self.astra_db = LibAstraDB(
token=self.token,
api_endpoint=self.api_endpoint,
namespace=self.namespace,
)
self.collection = self.astra_db.create_collection(
collection_name=self.collection_name,
)
@staticmethod
def _make_id(prompt: str, llm_string: str) -> str:
return f"{_hash(prompt)}#{_hash(llm_string)}"
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
doc_id = self._make_id(prompt, llm_string)
item = self.collection.find_one(
filter={
"_id": doc_id,
},
projection={
"body_blob": 1,
},
)["data"]["document"]
if item is not None:
generations = _loads_generations(item["body_blob"])
# this protects against malformed cached items:
if generations is not None:
return generations
else:
return None
else:
return None
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
doc_id = self._make_id(prompt, llm_string)
blob = _dumps_generations(return_val)
self.collection.upsert(
{
"_id": doc_id,
"body_blob": blob,
},
)
[docs] def delete_through_llm(
self, prompt: str, llm: LLM, stop: Optional[List[str]] = None
) -> None:
"""
A wrapper around `delete` with the LLM being passed.
In case the llm(prompt) calls have a `stop` param, you should pass it here
"""
llm_string = get_prompts(
{**llm.dict(), **{"stop": stop}},
[],
)[1]
return self.delete(prompt, llm_string=llm_string)
[docs] def delete(self, prompt: str, llm_string: str) -> None:
"""Evict from cache if there's an entry."""
doc_id = self._make_id(prompt, llm_string)
return self.collection.delete_one(doc_id)
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear cache. This is for all LLMs at once."""
self.astra_db.truncate_collection(self.collection_name)
ASTRA_DB_SEMANTIC_CACHE_DEFAULT_THRESHOLD = 0.85
ASTRA_DB_CACHE_DEFAULT_COLLECTION_NAME = "langchain_astradb_semantic_cache"
ASTRA_DB_SEMANTIC_CACHE_EMBEDDING_CACHE_SIZE = 16
[docs]class AstraDBSemanticCache(BaseCache):
"""
Cache that uses Astra DB as a vector-store backend for semantic
(i.e. similarity-based) lookup.
It uses a single (vector) collection and can store
cached values from several LLMs, so the LLM's 'llm_string' is stored
in the document metadata.
You can choose the preferred similarity (or use the API default) --
remember the threshold might require metric-dependend tuning.
"""
[docs] def __init__(
self,
*,
collection_name: str = ASTRA_DB_CACHE_DEFAULT_COLLECTION_NAME,
token: Optional[str] = None,
api_endpoint: Optional[str] = None,
astra_db_client: Optional[Any] = None, # 'astrapy.db.AstraDB' if passed
namespace: Optional[str] = None,
embedding: Embeddings,
metric: Optional[str] = None,
similarity_threshold: float = ASTRA_DB_SEMANTIC_CACHE_DEFAULT_THRESHOLD,
):
"""
Initialize the cache with all relevant parameters.
Args:
collection_name (str): name of the Astra DB collection to create/use.
token (Optional[str]): API token for Astra DB usage.
api_endpoint (Optional[str]): full URL to the API endpoint,
such as "https://<DB-ID>-us-east1.apps.astra.datastax.com".
astra_db_client (Optional[Any]): *alternative to token+api_endpoint*,
you can pass an already-created 'astrapy.db.AstraDB' instance.
namespace (Optional[str]): namespace (aka keyspace) where the
collection is created. Defaults to the database's "default namespace".
embedding (Embedding): Embedding provider for semantic
encoding and search.
metric: the function to use for evaluating similarity of text embeddings.
Defaults to 'cosine' (alternatives: 'euclidean', 'dot_product')
similarity_threshold (float, optional): the minimum similarity
for accepting a (semantic-search) match.
The default score threshold is tuned to the default metric.
Tune it carefully yourself if switching to another distance metric.
"""
try:
from astrapy.db import (
AstraDB as LibAstraDB,
)
except (ImportError, ModuleNotFoundError):
raise ImportError(
"Could not import a recent astrapy python package. "
"Please install it with `pip install --upgrade astrapy`."
)
# Conflicting-arg checks:
if astra_db_client is not None:
if token is not None or api_endpoint is not None:
raise ValueError(
"You cannot pass 'astra_db_client' to AstraDB if passing "
"'token' and 'api_endpoint'."
)
self.embedding = embedding
self.metric = metric
self.similarity_threshold = similarity_threshold
# The contract for this class has separate lookup and update:
# in order to spare some embedding calculations we cache them between
# the two calls.
# Note: each instance of this class has its own `_get_embedding` with
# its own lru.
@lru_cache(maxsize=ASTRA_DB_SEMANTIC_CACHE_EMBEDDING_CACHE_SIZE)
def _cache_embedding(text: str) -> List[float]:
return self.embedding.embed_query(text=text)
self._get_embedding = _cache_embedding
self.embedding_dimension = self._get_embedding_dimension()
self.collection_name = collection_name
self.token = token
self.api_endpoint = api_endpoint
self.namespace = namespace
if astra_db_client is not None:
self.astra_db = astra_db_client
else:
self.astra_db = LibAstraDB(
token=self.token,
api_endpoint=self.api_endpoint,
namespace=self.namespace,
)
self.collection = self.astra_db.create_collection(
collection_name=self.collection_name,
dimension=self.embedding_dimension,
metric=self.metric,
)
def _get_embedding_dimension(self) -> int:
return len(self._get_embedding(text="This is a sample sentence."))
@staticmethod
def _make_id(prompt: str, llm_string: str) -> str:
return f"{_hash(prompt)}#{_hash(llm_string)}"
[docs] def update(self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE) -> None:
"""Update cache based on prompt and llm_string."""
doc_id = self._make_id(prompt, llm_string)
llm_string_hash = _hash(llm_string)
embedding_vector = self._get_embedding(text=prompt)
body = _dumps_generations(return_val)
#
self.collection.upsert(
{
"_id": doc_id,
"body_blob": body,
"llm_string_hash": llm_string_hash,
"$vector": embedding_vector,
}
)
[docs] def lookup(self, prompt: str, llm_string: str) -> Optional[RETURN_VAL_TYPE]:
"""Look up based on prompt and llm_string."""
hit_with_id = self.lookup_with_id(prompt, llm_string)
if hit_with_id is not None:
return hit_with_id[1]
else:
return None
[docs] def lookup_with_id(
self, prompt: str, llm_string: str
) -> Optional[Tuple[str, RETURN_VAL_TYPE]]:
"""
Look up based on prompt and llm_string.
If there are hits, return (document_id, cached_entry) for the top hit
"""
prompt_embedding: List[float] = self._get_embedding(text=prompt)
llm_string_hash = _hash(llm_string)
hit = self.collection.vector_find_one(
vector=prompt_embedding,
filter={
"llm_string_hash": llm_string_hash,
},
fields=["body_blob", "_id"],
include_similarity=True,
)
if hit is None or hit["$similarity"] < self.similarity_threshold:
return None
else:
generations = _loads_generations(hit["body_blob"])
if generations is not None:
# this protects against malformed cached items:
return (hit["_id"], generations)
else:
return None
[docs] def lookup_with_id_through_llm(
self, prompt: str, llm: LLM, stop: Optional[List[str]] = None
) -> Optional[Tuple[str, RETURN_VAL_TYPE]]:
llm_string = get_prompts(
{**llm.dict(), **{"stop": stop}},
[],
)[1]
return self.lookup_with_id(prompt, llm_string=llm_string)
[docs] def delete_by_document_id(self, document_id: str) -> None:
"""
Given this is a "similarity search" cache, an invalidation pattern
that makes sense is first a lookup to get an ID, and then deleting
with that ID. This is for the second step.
"""
self.collection.delete_one(document_id)
[docs] def clear(self, **kwargs: Any) -> None:
"""Clear the *whole* semantic cache."""
self.astra_db.truncate_collection(self.collection_name)