Model caches
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)
In Memory Cache
from langchain_community.cache import InMemoryCache
set_llm_cache(InMemoryCache())
%%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_community.cache import SQLiteCache
set_llm_cache(SQLiteCache(database_path=".langchain.db"))
%%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_community.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))
%%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_community.cache import RedisCache
from redis import Redis
set_llm_cache(RedisCache(redis_=Redis()))
%%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_community.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_community.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))
%%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_community.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))
%%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.'
MongoDB Atlas Cache
MongoDB Atlas is a fully-managed cloud database available in AWS, Azure, and GCP. It has native support for Vector Search on the MongoDB document data. Use MongoDB Atlas Vector Search to semantically cache prompts and responses.
MongoDBCache
An abstraction to store a simple cache in MongoDB. This does not use Semantic Caching, nor does it require an index to be made on the collection before generation.
To import this cache:
from langchain_mongodb.cache import MongoDBCache
To use this cache with your LLMs:
from langchain_core.globals import set_llm_cache
# use any embedding provider...
from tests.integration_tests.vectorstores.fake_embeddings import FakeEmbeddings
mongodb_atlas_uri = "<YOUR_CONNECTION_STRING>"
COLLECTION_NAME="<YOUR_CACHE_COLLECTION_NAME>"
DATABASE_NAME="<YOUR_DATABASE_NAME>"
set_llm_cache(MongoDBCache(
connection_string=mongodb_atlas_uri,
collection_name=COLLECTION_NAME,
database_name=DATABASE_NAME,
))
MongoDBAtlasSemanticCache
Semantic caching allows users to retrieve cached prompts based on semantic similarity between the user input and previously cached results. Under the hood it blends MongoDBAtlas as both a cache and a vectorstore.
The MongoDBAtlasSemanticCache inherits from MongoDBAtlasVectorSearch and needs an Atlas Vector Search Index defined to work. Please look at the usage example on how to set up the index.
To import this cache:
from langchain_mongodb.cache import MongoDBAtlasSemanticCache
To use this cache with your LLMs:
from langchain_core.globals import set_llm_cache
# use any embedding provider...
from tests.integration_tests.vectorstores.fake_embeddings import FakeEmbeddings
mongodb_atlas_uri = "<YOUR_CONNECTION_STRING>"
COLLECTION_NAME="<YOUR_CACHE_COLLECTION_NAME>"
DATABASE_NAME="<YOUR_DATABASE_NAME>"
set_llm_cache(MongoDBAtlasSemanticCache(
embedding=FakeEmbeddings(),
connection_string=mongodb_atlas_uri,
collection_name=COLLECTION_NAME,
database_name=DATABASE_NAME,
))
To find more resources about using MongoDBSemanticCache visit here
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_community.cache import MomentoCache
cache_name = "langchain"
ttl = timedelta(days=1)
set_llm_cache(MomentoCache.from_client_params(cache_name, ttl))
%%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))
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_community.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))
Cassandra caches
Apache Cassandra® is a NoSQL, row-oriented, highly scalable and highly available database. Starting with version 5.0, the database ships with vector search capabilities.
You can use Cassandra for caching LLM responses, choosing from the exact-match CassandraCache or the (vector-similarity-based) CassandraSemanticCache.
Let's see both in action. The next cells guide you through the (little) required setup, and the following cells showcase the two available cache classes.
Required dependency
%pip install --upgrade --quiet "cassio>=0.1.4"
Connect to the DB
The Cassandra caches shown in this page can be used with Cassandra as well as other derived databases, such as Astra DB, which use the CQL (Cassandra Query Language) protocol.
DataStax Astra DB is a managed serverless database built on Cassandra, offering the same interface and strengths.
Depending on whether you connect to a Cassandra cluster or to Astra DB through CQL, you will provide different parameters when instantiating the cache (through initialization of a CassIO connection).
Connecting to a Cassandra cluster
You first need to create a cassandra.cluster.Session object, as described in the Cassandra driver documentation. The details vary (e.g. with network settings and authentication), but this might be something like:
from cassandra.cluster import Cluster
cluster = Cluster(["127.0.0.1"])
session = cluster.connect()
You can now set the session, along with your desired keyspace name, as a global CassIO parameter:
import cassio
CASSANDRA_KEYSPACE = input("CASSANDRA_KEYSPACE = ")
cassio.init(session=session, keyspace=CASSANDRA_KEYSPACE)
CASSANDRA_KEYSPACE = demo_keyspace
Connecting to Astra DB through CQL
In this case you initialize CassIO with the following connection parameters:
- the Database ID, e.g.
01234567-89ab-cdef-0123-456789abcdef - the Token, e.g.
AstraCS:6gBhNmsk135....(it must be a "Database Administrator" token) - Optionally a Keyspace name (if omitted, the default one for the database will be used)
import getpass
ASTRA_DB_ID = input("ASTRA_DB_ID = ")
ASTRA_DB_APPLICATION_TOKEN = getpass.getpass("ASTRA_DB_APPLICATION_TOKEN = ")
desired_keyspace = input("ASTRA_DB_KEYSPACE (optional, can be left empty) = ")
if desired_keyspace:
ASTRA_DB_KEYSPACE = desired_keyspace
else:
ASTRA_DB_KEYSPACE = None
ASTRA_DB_ID = 01234567-89ab-cdef-0123-456789abcdef
ASTRA_DB_APPLICATION_TOKEN = ········
ASTRA_DB_KEYSPACE (optional, can be left empty) = my_keyspace
import cassio
cassio.init(
database_id=ASTRA_DB_ID,
token=ASTRA_DB_APPLICATION_TOKEN,
keyspace=ASTRA_DB_KEYSPACE,
)
Cassandra: Exact cache
This will avoid invoking the LLM when the supplied prompt is exactly the same as one encountered already:
from langchain_community.cache import CassandraCache
from langchain_core.globals import set_llm_cache
set_llm_cache(CassandraCache())
%%time
print(llm.invoke("Why is the Moon always showing the same side?"))
The Moon is tidally locked with the Earth, which means that its rotation on its own axis is synchronized with its orbit around the Earth. This results in the Moon always showing the same side to the Earth. This is because the gravitational forces between the Earth and the Moon have caused the Moon's rotation to slow down over time, until it reached a point where it takes the same amount of time for the Moon to rotate on its axis as it does to orbit around the Earth. This phenomenon is common among satellites in close orbits around their parent planets and is known as tidal locking.
CPU times: user 92.5 ms, sys: 8.89 ms, total: 101 ms
Wall time: 1.98 s
%%time
print(llm.invoke("Why is the Moon always showing the same side?"))
The Moon is tidally locked with the Earth, which means that its rotation on its own axis is synchronized with its orbit around the Earth. This results in the Moon always showing the same side to the Earth. This is because the gravitational forces between the Earth and the Moon have caused the Moon's rotation to slow down over time, until it reached a point where it takes the same amount of time for the Moon to rotate on its axis as it does to orbit around the Earth. This phenomenon is common among satellites in close orbits around their parent planets and is known as tidal locking.
CPU times: user 5.51 ms, sys: 0 ns, total: 5.51 ms
Wall time: 5.78 ms
Cassandra: 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()
from langchain_community.cache import CassandraSemanticCache
from langchain_core.globals import set_llm_cache
set_llm_cache(
CassandraSemanticCache(
embedding=embedding,
table_name="my_semantic_cache",
)
)
%%time
print(llm.invoke("Why is the Moon always showing the same side?"))
The Moon is always showing the same side because of a phenomenon called synchronous rotation. This means that the Moon rotates on its axis at the same rate that it orbits around the Earth, which takes approximately 27.3 days. This results in the same side of the Moon always facing the Earth. This is due to the gravitational forces between the Earth and the Moon, which have caused the Moon's rotation to gradually slow down and become synchronized with its orbit. This is a common occurrence among many moons in our solar system.
CPU times: user 49.5 ms, sys: 7.38 ms, total: 56.9 ms
Wall time: 2.55 s
%%time
print(llm.invoke("How come we always see one face of the moon?"))
The Moon is always showing the same side because of a phenomenon called synchronous rotation. This means that the Moon rotates on its axis at the same rate that it orbits around the Earth, which takes approximately 27.3 days. This results in the same side of the Moon always facing the Earth. This is due to the gravitational forces between the Earth and the Moon, which have caused the Moon's rotation to gradually slow down and become synchronized with its orbit. This is a common occurrence among many moons in our solar system.
CPU times: user 21.2 ms, sys: 3.38 ms, total: 24.6 ms
Wall time: 532 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.globals import set_llm_cache
from langchain_astradb import AstraDBCache
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()
from langchain_astradb import AstraDBSemanticCache
set_llm_cache(
AstraDBSemanticCache(
api_endpoint=ASTRA_DB_API_ENDPOINT,
token=ASTRA_DB_APPLICATION_TOKEN,
embedding=embedding,
collection_name="demo_semantic_cache",
)
)
%%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()
with open("../../how_to/state_of_the_union.txt") as f:
state_of_the_union = f.read()
texts = text_splitter.split_text(state_of_the_union)
from langchain_core.documents 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."
Cache classes: summary table
Cache classes are implemented by inheriting the BaseCache class.
This table lists all 20 derived classes with links to the API Reference.
| Namespace 🔻 | Class |
|---|---|
| langchain_astradb.cache | AstraDBCache |
| langchain_astradb.cache | AstraDBSemanticCache |
| langchain_community.cache | AstraDBCache |
| langchain_community.cache | AstraDBSemanticCache |
| langchain_community.cache | AzureCosmosDBSemanticCache |
| langchain_community.cache | CassandraCache |
| langchain_community.cache | CassandraSemanticCache |
| langchain_community.cache | GPTCache |
| langchain_community.cache | InMemoryCache |
| langchain_community.cache | MomentoCache |
| langchain_community.cache | OpenSearchSemanticCache |
| langchain_community.cache | RedisSemanticCache |
| langchain_community.cache | SQLAlchemyCache |
| langchain_community.cache | SQLAlchemyMd5Cache |
| langchain_community.cache | UpstashRedisCache |
| langchain_core.caches | InMemoryCache |
| langchain_elasticsearch.cache | ElasticsearchCache |
| langchain_mongodb.cache | MongoDBAtlasSemanticCache |
| langchain_mongodb.cache | MongoDBCache |