Skip to main content

LLM Caching integrations

This notebook covers how to cache results of individual LLM calls using different caches.

from langchain.globals import set_llm_cache
from langchain_openai import OpenAI

# To make the caching really obvious, lets use a slower model.
llm = OpenAI(model_name="gpt-3.5-turbo-instruct", n=2, best_of=2)

API Reference:

In Memory Cache​

from langchain.cache import InMemoryCache

set_llm_cache(InMemoryCache())

API Reference:

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 52.2 ms, sys: 15.2 ms, total: 67.4 ms
Wall time: 1.19 s
"\n\nWhy couldn't the bicycle stand up by itself? Because it was...two tired!"
%%time
# The second time it is, so it goes faster
llm("Tell me a joke")
CPU times: user 191 µs, sys: 11 µs, total: 202 µs
Wall time: 205 µs
"\n\nWhy couldn't the bicycle stand up by itself? Because it was...two tired!"

SQLite Cache​

!rm .langchain.db
# We can do the same thing with a SQLite cache
from langchain.cache import SQLiteCache

set_llm_cache(SQLiteCache(database_path=".langchain.db"))

API Reference:

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 33.2 ms, sys: 18.1 ms, total: 51.2 ms
Wall time: 667 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side.'
%%time
# The second time it is, so it goes faster
llm("Tell me a joke")
CPU times: user 4.86 ms, sys: 1.97 ms, total: 6.83 ms
Wall time: 5.79 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side.'

Upstash Redis Cache​

Standard Cache​

Use Upstash Redis to cache prompts and responses with a serverless HTTP API.

import langchain
from langchain.cache import UpstashRedisCache
from upstash_redis import Redis

URL = "<UPSTASH_REDIS_REST_URL>"
TOKEN = "<UPSTASH_REDIS_REST_TOKEN>"

langchain.llm_cache = UpstashRedisCache(redis_=Redis(url=URL, token=TOKEN))

API Reference:

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 7.56 ms, sys: 2.98 ms, total: 10.5 ms
Wall time: 1.14 s
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'
%%time
# The second time it is, so it goes faster
llm("Tell me a joke")
CPU times: user 2.78 ms, sys: 1.95 ms, total: 4.73 ms
Wall time: 82.9 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'

Redis Cache​

Standard Cache​

Use Redis to cache prompts and responses.

# We can do the same thing with a Redis cache
# (make sure your local Redis instance is running first before running this example)
from langchain.cache import RedisCache
from redis import Redis

set_llm_cache(RedisCache(redis_=Redis()))

API Reference:

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 6.88 ms, sys: 8.75 ms, total: 15.6 ms
Wall time: 1.04 s
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'
%%time
# The second time it is, so it goes faster
llm("Tell me a joke")
CPU times: user 1.59 ms, sys: 610 µs, total: 2.2 ms
Wall time: 5.58 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'

Semantic Cache​

Use Redis to cache prompts and responses and evaluate hits based on semantic similarity.

from langchain.cache import RedisSemanticCache
from langchain_openai import OpenAIEmbeddings

set_llm_cache(
    RedisSemanticCache(redis_url="redis://localhost:6379", embedding=OpenAIEmbeddings())
)
%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 351 ms, sys: 156 ms, total: 507 ms
Wall time: 3.37 s
"\n\nWhy don't scientists trust atoms?\nBecause they make up everything."
%%time
# The second time, while not a direct hit, the question is semantically similar to the original question,
# so it uses the cached result!
llm("Tell me one joke")
CPU times: user 6.25 ms, sys: 2.72 ms, total: 8.97 ms
Wall time: 262 ms
"\n\nWhy don't scientists trust atoms?\nBecause they make up everything."

GPTCache​

We can use GPTCache for exact match caching OR to cache results based on semantic similarity

Let's first start with an example of exact match

import hashlib

from gptcache import Cache
from gptcache.manager.factory import manager_factory
from gptcache.processor.pre import get_prompt
from langchain.cache import GPTCache


def get_hashed_name(name):
    return hashlib.sha256(name.encode()).hexdigest()


def init_gptcache(cache_obj: Cache, llm: str):
    hashed_llm = get_hashed_name(llm)
    cache_obj.init(
        pre_embedding_func=get_prompt,
        data_manager=manager_factory(manager="map", data_dir=f"map_cache_{hashed_llm}"),
    )


set_llm_cache(GPTCache(init_gptcache))

API Reference:

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 21.5 ms, sys: 21.3 ms, total: 42.8 ms
Wall time: 6.2 s
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'
%%time
# The second time it is, so it goes faster
llm("Tell me a joke")
CPU times: user 571 µs, sys: 43 µs, total: 614 µs
Wall time: 635 µs
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'

Let's now show an example of similarity caching

import hashlib

from gptcache import Cache
from gptcache.adapter.api import init_similar_cache
from langchain.cache import GPTCache


def get_hashed_name(name):
    return hashlib.sha256(name.encode()).hexdigest()


def init_gptcache(cache_obj: Cache, llm: str):
    hashed_llm = get_hashed_name(llm)
    init_similar_cache(cache_obj=cache_obj, data_dir=f"similar_cache_{hashed_llm}")


set_llm_cache(GPTCache(init_gptcache))

API Reference:

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 1.42 s, sys: 279 ms, total: 1.7 s
Wall time: 8.44 s
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side.'
%%time
# This is an exact match, so it finds it in the cache
llm("Tell me a joke")
CPU times: user 866 ms, sys: 20 ms, total: 886 ms
Wall time: 226 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side.'
%%time
# This is not an exact match, but semantically within distance so it hits!
llm("Tell me joke")
CPU times: user 853 ms, sys: 14.8 ms, total: 868 ms
Wall time: 224 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side.'

Momento Cache​

Use Momento to cache prompts and responses.

Requires momento to use, uncomment below to install:

%pip install --upgrade --quiet  momento

You'll need to get a Momento auth token to use this class. This can either be passed in to a momento.CacheClient if you'd like to instantiate that directly, as a named parameter auth_token to MomentoChatMessageHistory.from_client_params, or can just be set as an environment variable MOMENTO_AUTH_TOKEN.

from datetime import timedelta

from langchain.cache import MomentoCache

cache_name = "langchain"
ttl = timedelta(days=1)
set_llm_cache(MomentoCache.from_client_params(cache_name, ttl))

API Reference:

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 40.7 ms, sys: 16.5 ms, total: 57.2 ms
Wall time: 1.73 s
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'
%%time
# The second time it is, so it goes faster
# When run in the same region as the cache, latencies are single digit ms
llm("Tell me a joke")
CPU times: user 3.16 ms, sys: 2.98 ms, total: 6.14 ms
Wall time: 57.9 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'

SQLAlchemy Cache​

You can use SQLAlchemyCache to cache with any SQL database supported by SQLAlchemy.

# from langchain.cache import SQLAlchemyCache
# from sqlalchemy import create_engine

# engine = create_engine("postgresql://postgres:postgres@localhost:5432/postgres")
# set_llm_cache(SQLAlchemyCache(engine))

API Reference:

Custom SQLAlchemy Schemas​

# You can define your own declarative SQLAlchemyCache child class to customize the schema used for caching. For example, to support high-speed fulltext prompt indexing with Postgres, use:

from langchain.cache import SQLAlchemyCache
from sqlalchemy import Column, Computed, Index, Integer, Sequence, String, create_engine
from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy_utils import TSVectorType

Base = declarative_base()


class FulltextLLMCache(Base):  # type: ignore
    """Postgres table for fulltext-indexed LLM Cache"""

    __tablename__ = "llm_cache_fulltext"
    id = Column(Integer, Sequence("cache_id"), primary_key=True)
    prompt = Column(String, nullable=False)
    llm = Column(String, nullable=False)
    idx = Column(Integer)
    response = Column(String)
    prompt_tsv = Column(
        TSVectorType(),
        Computed("to_tsvector('english', llm || ' ' || prompt)", persisted=True),
    )
    __table_args__ = (
        Index("idx_fulltext_prompt_tsv", prompt_tsv, postgresql_using="gin"),
    )


engine = create_engine("postgresql://postgres:postgres@localhost:5432/postgres")
set_llm_cache(SQLAlchemyCache(engine, FulltextLLMCache))

API Reference:

Cassandra caches​

You can use Cassandra / Astra DB through CQL for caching LLM responses, choosing from the exact-match CassandraCache or the (vector-similarity-based) CassandraSemanticCache.

Let's see both in action in the following cells.

Connect to the DB​

First you need to establish a Session to the DB and to specify a keyspace for the cache table(s). The following gets you connected to Astra DB through CQL (see e.g. here for more backends and connection options).

import getpass

keyspace = input("\nKeyspace name? ")
ASTRA_DB_APPLICATION_TOKEN = getpass.getpass('\nAstra DB Token ("AstraCS:...") ')
ASTRA_DB_SECURE_BUNDLE_PATH = input("Full path to your Secure Connect Bundle? ")

Keyspace name? my_keyspace

Astra DB Token ("AstraCS:...") ········
Full path to your Secure Connect Bundle? /path/to/secure-connect-databasename.zip
from cassandra.auth import PlainTextAuthProvider
from cassandra.cluster import Cluster

cluster = Cluster(
    cloud={
        "secure_connect_bundle": ASTRA_DB_SECURE_BUNDLE_PATH,
    },
    auth_provider=PlainTextAuthProvider("token", ASTRA_DB_APPLICATION_TOKEN),
)
session = cluster.connect()

Exact cache​

This will avoid invoking the LLM when the supplied prompt is exactly the same as one encountered already:

from langchain.cache import CassandraCache
from langchain.globals import set_llm_cache

set_llm_cache(CassandraCache(session=session, keyspace=keyspace))
%%time

print(llm.invoke("Why is the Moon always showing the same side?"))


The Moon always shows the same side because it is tidally locked to Earth.
CPU times: user 41.7 ms, sys: 153 µs, total: 41.8 ms
Wall time: 1.96 s
%%time

print(llm.invoke("Why is the Moon always showing the same side?"))


The Moon always shows the same side because it is tidally locked to Earth.
CPU times: user 4.09 ms, sys: 0 ns, total: 4.09 ms
Wall time: 119 ms

Semantic cache​

This cache will do a semantic similarity search and return a hit if it finds a cached entry that is similar enough, For this, you need to provide an Embeddings instance of your choice.

from langchain_openai import OpenAIEmbeddings

embedding = OpenAIEmbeddings()

API Reference:

from langchain.cache import CassandraSemanticCache

set_llm_cache(
    CassandraSemanticCache(
        session=session,
        keyspace=keyspace,
        embedding=embedding,
        table_name="cass_sem_cache",
    )
)
%%time

print(llm.invoke("Why is the Moon always showing the same side?"))


The Moon always shows the same side because it is tidally locked with Earth. This means that the same side of the Moon always faces Earth.
CPU times: user 21.3 ms, sys: 177 µs, total: 21.4 ms
Wall time: 3.09 s
%%time

print(llm.invoke("How come we always see one face of the moon?"))


The Moon always shows the same side because it is tidally locked with Earth. This means that the same side of the Moon always faces Earth.
CPU times: user 10.9 ms, sys: 17 µs, total: 10.9 ms
Wall time: 461 ms

Attribution statement​

Apache Cassandra, Cassandra and Apache are either registered trademarks or trademarks of the Apache Software Foundation in the United States and/or other countries.

Astra DB Caches​

You can easily use Astra DB as an LLM cache, with either the "exact" or the "semantic-based" cache.

Make sure you have a running database (it must be a Vector-enabled database to use the Semantic cache) and get the required credentials on your Astra dashboard:

  • the API Endpoint looks like https://01234567-89ab-cdef-0123-456789abcdef-us-east1.apps.astra.datastax.com
  • the Token looks like AstraCS:6gBhNmsk135....
import getpass

ASTRA_DB_API_ENDPOINT = input("ASTRA_DB_API_ENDPOINT = ")
ASTRA_DB_APPLICATION_TOKEN = getpass.getpass("ASTRA_DB_APPLICATION_TOKEN = ")
ASTRA_DB_API_ENDPOINT =  https://01234567-89ab-cdef-0123-456789abcdef-us-east1.apps.astra.datastax.com
ASTRA_DB_APPLICATION_TOKEN =  ········

Astra DB exact LLM cache​

This will avoid invoking the LLM when the supplied prompt is exactly the same as one encountered already:

from langchain.cache import AstraDBCache
from langchain.globals import set_llm_cache

set_llm_cache(
    AstraDBCache(
        api_endpoint=ASTRA_DB_API_ENDPOINT,
        token=ASTRA_DB_APPLICATION_TOKEN,
    )
)
%%time

print(llm.invoke("Is a true fakery the same as a fake truth?"))


There is no definitive answer to this question as it depends on the interpretation of the terms "true fakery" and "fake truth". However, one possible interpretation is that a true fakery is a counterfeit or imitation that is intended to deceive, whereas a fake truth is a false statement that is presented as if it were true.
CPU times: user 70.8 ms, sys: 4.13 ms, total: 74.9 ms
Wall time: 2.06 s
%%time

print(llm.invoke("Is a true fakery the same as a fake truth?"))


There is no definitive answer to this question as it depends on the interpretation of the terms "true fakery" and "fake truth". However, one possible interpretation is that a true fakery is a counterfeit or imitation that is intended to deceive, whereas a fake truth is a false statement that is presented as if it were true.
CPU times: user 15.1 ms, sys: 3.7 ms, total: 18.8 ms
Wall time: 531 ms

Astra DB Semantic cache​

This cache will do a semantic similarity search and return a hit if it finds a cached entry that is similar enough, For this, you need to provide an Embeddings instance of your choice.

from langchain_openai import OpenAIEmbeddings

embedding = OpenAIEmbeddings()

API Reference:

from langchain.cache import AstraDBSemanticCache

set_llm_cache(
    AstraDBSemanticCache(
        api_endpoint=ASTRA_DB_API_ENDPOINT,
        token=ASTRA_DB_APPLICATION_TOKEN,
        embedding=embedding,
        collection_name="demo_semantic_cache",
    )
)

API Reference:

%%time

print(llm.invoke("Are there truths that are false?"))


There is no definitive answer to this question since it presupposes a great deal about the nature of truth itself, which is a matter of considerable philosophical debate. It is possible, however, to construct scenarios in which something could be considered true despite being false, such as if someone sincerely believes something to be true even though it is not.
CPU times: user 65.6 ms, sys: 15.3 ms, total: 80.9 ms
Wall time: 2.72 s
%%time

print(llm.invoke("Is is possible that something false can be also true?"))


There is no definitive answer to this question since it presupposes a great deal about the nature of truth itself, which is a matter of considerable philosophical debate. It is possible, however, to construct scenarios in which something could be considered true despite being false, such as if someone sincerely believes something to be true even though it is not.
CPU times: user 29.3 ms, sys: 6.21 ms, total: 35.5 ms
Wall time: 1.03 s

Azure Cosmos DB Semantic Cache​

You can use this integrated vector database for caching.

from langchain_community.cache import AzureCosmosDBSemanticCache
from langchain_community.vectorstores.azure_cosmos_db import (
    CosmosDBSimilarityType,
    CosmosDBVectorSearchType,
)
from langchain_openai import OpenAIEmbeddings

# Read more about Azure CosmosDB Mongo vCore vector search here https://learn.microsoft.com/en-us/azure/cosmos-db/mongodb/vcore/vector-search

NAMESPACE = "langchain_test_db.langchain_test_collection"
CONNECTION_STRING = (
    "Please provide your azure cosmos mongo vCore vector db connection string"
)

DB_NAME, COLLECTION_NAME = NAMESPACE.split(".")

# Default value for these params
num_lists = 3
dimensions = 1536
similarity_algorithm = CosmosDBSimilarityType.COS
kind = CosmosDBVectorSearchType.VECTOR_IVF
m = 16
ef_construction = 64
ef_search = 40
score_threshold = 0.9
application_name = "LANGCHAIN_CACHING_PYTHON"


set_llm_cache(
    AzureCosmosDBSemanticCache(
        cosmosdb_connection_string=CONNECTION_STRING,
        cosmosdb_client=None,
        embedding=OpenAIEmbeddings(),
        database_name=DB_NAME,
        collection_name=COLLECTION_NAME,
        num_lists=num_lists,
        similarity=similarity_algorithm,
        kind=kind,
        dimensions=dimensions,
        m=m,
        ef_construction=ef_construction,
        ef_search=ef_search,
        score_threshold=score_threshold,
        application_name=application_name,
    )
)
%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 45.6 ms, sys: 19.7 ms, total: 65.3 ms
Wall time: 2.29 s
'\n\nWhy was the math book sad? Because it had too many problems.'
%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 9.61 ms, sys: 3.42 ms, total: 13 ms
Wall time: 474 ms
'\n\nWhy was the math book sad? Because it had too many problems.'

Elasticsearch Cache​

A caching layer for LLMs that uses Elasticsearch.

First install the LangChain integration with Elasticsearch.

%pip install -U langchain-elasticsearch

Use the class ElasticsearchCache.

Simple example:

from elasticsearch import Elasticsearch
from langchain.globals import set_llm_cache
from langchain_elasticsearch import ElasticsearchCache

es_client = Elasticsearch(hosts="http://localhost:9200")
set_llm_cache(
    ElasticsearchCache(
        es_connection=es_client,
        index_name="llm-chat-cache",
        metadata={"project": "my_chatgpt_project"},
    )
)

The index_name parameter can also accept aliases. This allows to use the ILM: Manage the index lifecycle that we suggest to consider for managing retention and controlling cache growth.

Look at the class docstring for all parameters.

Index the generated text​

The cached data won't be searchable by default. The developer can customize the building of the Elasticsearch document in order to add indexed text fields, where to put, for example, the text generated by the LLM.

This can be done by subclassing end overriding methods. The new cache class can be applied also to a pre-existing cache index:

import json
from typing import Any, Dict, List

from elasticsearch import Elasticsearch
from langchain.globals import set_llm_cache
from langchain_core.caches import RETURN_VAL_TYPE
from langchain_elasticsearch import ElasticsearchCache


class SearchableElasticsearchCache(ElasticsearchCache):
    @property
    def mapping(self) -> Dict[str, Any]:
        mapping = super().mapping
        mapping["mappings"]["properties"]["parsed_llm_output"] = {
            "type": "text",
            "analyzer": "english",
        }
        return mapping

    def build_document(
        self, prompt: str, llm_string: str, return_val: RETURN_VAL_TYPE
    ) -> Dict[str, Any]:
        body = super().build_document(prompt, llm_string, return_val)
        body["parsed_llm_output"] = self._parse_output(body["llm_output"])
        return body

    @staticmethod
    def _parse_output(data: List[str]) -> List[str]:
        return [
            json.loads(output)["kwargs"]["message"]["kwargs"]["content"]
            for output in data
        ]


es_client = Elasticsearch(hosts="http://localhost:9200")
set_llm_cache(
    SearchableElasticsearchCache(es_connection=es_client, index_name="llm-chat-cache")
)

When overriding the mapping and the document building, please only make additive modifications, keeping the base mapping intact.

Optional Caching​

You can also turn off caching for specific LLMs should you choose. In the example below, even though global caching is enabled, we turn it off for a specific LLM

llm = OpenAI(model_name="gpt-3.5-turbo-instruct", n=2, best_of=2, cache=False)
%%time
llm("Tell me a joke")
CPU times: user 5.8 ms, sys: 2.71 ms, total: 8.51 ms
Wall time: 745 ms
'\n\nWhy did the chicken cross the road?\n\nTo get to the other side!'
%%time
llm("Tell me a joke")
CPU times: user 4.91 ms, sys: 2.64 ms, total: 7.55 ms
Wall time: 623 ms
'\n\nTwo guys stole a calendar. They got six months each.'

Optional Caching in Chains​

You can also turn off caching for particular nodes in chains. Note that because of certain interfaces, its often easier to construct the chain first, and then edit the LLM afterwards.

As an example, we will load a summarizer map-reduce chain. We will cache results for the map-step, but then not freeze it for the combine step.

llm = OpenAI(model_name="gpt-3.5-turbo-instruct")
no_cache_llm = OpenAI(model_name="gpt-3.5-turbo-instruct", cache=False)
from langchain_text_splitters import CharacterTextSplitter

text_splitter = CharacterTextSplitter()

API Reference:

with open("../../modules/state_of_the_union.txt") as f:
    state_of_the_union = f.read()
texts = text_splitter.split_text(state_of_the_union)
from langchain_community.docstore.document import Document

docs = [Document(page_content=t) for t in texts[:3]]
from langchain.chains.summarize import load_summarize_chain
chain = load_summarize_chain(llm, chain_type="map_reduce", reduce_llm=no_cache_llm)
%%time
chain.run(docs)
CPU times: user 452 ms, sys: 60.3 ms, total: 512 ms
Wall time: 5.09 s
'\n\nPresident Biden is discussing the American Rescue Plan and the Bipartisan Infrastructure Law, which will create jobs and help Americans. He also talks about his vision for America, which includes investing in education and infrastructure. In response to Russian aggression in Ukraine, the United States is joining with European allies to impose sanctions and isolate Russia. American forces are being mobilized to protect NATO countries in the event that Putin decides to keep moving west. The Ukrainians are bravely fighting back, but the next few weeks will be hard for them. Putin will pay a high price for his actions in the long run. Americans should not be alarmed, as the United States is taking action to protect its interests and allies.'

When we run it again, we see that it runs substantially faster but the final answer is different. This is due to caching at the map steps, but not at the reduce step.

%%time
chain.run(docs)
CPU times: user 11.5 ms, sys: 4.33 ms, total: 15.8 ms
Wall time: 1.04 s
'\n\nPresident Biden is discussing the American Rescue Plan and the Bipartisan Infrastructure Law, which will create jobs and help Americans. He also talks about his vision for America, which includes investing in education and infrastructure.'
!rm .langchain.db sqlite.db

OpenSearch Semantic Cache​

Use OpenSearch as a semantic cache to cache prompts and responses and evaluate hits based on semantic similarity.

from langchain_community.cache import OpenSearchSemanticCache
from langchain_openai import OpenAIEmbeddings

set_llm_cache(
    OpenSearchSemanticCache(
        opensearch_url="http://localhost:9200", embedding=OpenAIEmbeddings()
    )
)
%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")
CPU times: user 39.4 ms, sys: 11.8 ms, total: 51.2 ms
Wall time: 1.55 s
"\n\nWhy don't scientists trust atoms?\n\nBecause they make up everything."
%%time
# The second time, while not a direct hit, the question is semantically similar to the original question,
# so it uses the cached result!
llm("Tell me one joke")
CPU times: user 4.66 ms, sys: 1.1 ms, total: 5.76 ms
Wall time: 113 ms
"\n\nWhy don't scientists trust atoms?\n\nBecause they make up everything."

Help us out by providing feedback on this documentation page: