How to cache LLM calls#

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

from langchain.llms import OpenAI

In Memory Cache#

import langchain
from langchain.cache import InMemoryCache
langchain.llm_cache = InMemoryCache()

# To make the caching really obvious, lets use a slower model.
llm = OpenAI(model_name="text-davinci-002", n=2, best_of=2)

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")

CPU times: user 14.2 ms, sys: 4.9 ms, total: 19.1 ms
Wall time: 1.1 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 162 µs, sys: 7 µs, total: 169 µs
Wall time: 175 µs

'\n\nWhy did the chicken cross the road?\n\nTo get to the other side.'

SQLite Cache#

!rm .langchain.db

# We can do the same thing with a SQLite cache
from langchain.cache import SQLiteCache
langchain.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 17 ms, sys: 9.76 ms, total: 26.7 ms
Wall time: 825 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 2.46 ms, sys: 1.23 ms, total: 3.7 ms
Wall time: 2.67 ms

'\n\nWhy did the chicken cross the road?\n\nTo get to the other side.'

Redis Cache#

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

%%time
# The second time it is, so it goes faster
llm("Tell me a joke")

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

# Avoid multiple caches using the same file, causing different llm model caches to affect each other
i = 0
file_prefix = "data_map"

def init_gptcache_map(cache_obj: gptcache.Cache):
    global i
    cache_path = f'{file_prefix}_{i}.txt'
    cache_obj.init(
        pre_embedding_func=get_prompt,
        data_manager=get_data_manager(data_path=cache_path),
    )
    i += 1

langchain.llm_cache = GPTCache(init_gptcache_map)

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")

CPU times: user 8.6 ms, sys: 3.82 ms, total: 12.4 ms
Wall time: 881 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 286 µs, sys: 21 µs, total: 307 µs
Wall time: 316 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 gptcache
from gptcache.processor.pre import get_prompt
from gptcache.manager.factory import get_data_manager
from langchain.cache import GPTCache
from gptcache.manager import get_data_manager, CacheBase, VectorBase
from gptcache import Cache
from gptcache.embedding import Onnx
from gptcache.similarity_evaluation.distance import SearchDistanceEvaluation

# Avoid multiple caches using the same file, causing different llm model caches to affect each other
i = 0
file_prefix = "data_map"
llm_cache = Cache()


def init_gptcache_map(cache_obj: gptcache.Cache):
    global i
    cache_path = f'{file_prefix}_{i}.txt'
    onnx = Onnx()
    cache_base = CacheBase('sqlite')
    vector_base = VectorBase('faiss', dimension=onnx.dimension)
    data_manager = get_data_manager(cache_base, vector_base, max_size=10, clean_size=2)
    cache_obj.init(
        pre_embedding_func=get_prompt,
        embedding_func=onnx.to_embeddings,
        data_manager=data_manager,
        similarity_evaluation=SearchDistanceEvaluation(),
    )
    i += 1

langchain.llm_cache = GPTCache(init_gptcache_map)

%%time
# The first time, it is not yet in cache, so it should take longer
llm("Tell me a joke")

CPU times: user 1.01 s, sys: 153 ms, total: 1.16 s
Wall time: 2.49 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 745 ms, sys: 13.2 ms, total: 758 ms
Wall time: 136 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 737 ms, sys: 7.79 ms, total: 745 ms
Wall time: 135 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")
# langchain.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 sqlalchemy import Column, Integer, String, Computed, Index, Sequence
from sqlalchemy import create_engine
from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy_utils import TSVectorType
from langchain.cache import SQLAlchemyCache

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")
langchain.llm_cache = SQLAlchemyCache(engine, FulltextLLMCache)

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="text-davinci-002", 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="text-davinci-002")
no_cache_llm = OpenAI(model_name="text-davinci-002", cache=False)

from langchain.text_splitter import CharacterTextSplitter
from langchain.chains.mapreduce import MapReduceChain

text_splitter = CharacterTextSplitter()

with open('../../../state_of_the_union.txt') as f:
    state_of_the_union = f.read()
texts = text_splitter.split_text(state_of_the_union)

from langchain.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.'

How to cache LLM calls

Contents