langchain/docs/modules/models/llms/examples/llm_caching.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "f36d938c",
   "metadata": {},
   "source": [
    "# How to cache LLM calls\n",
    "This notebook covers how to cache results of individual LLM calls."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "10ad9224",
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.llms import OpenAI"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b50f0598",
   "metadata": {},
   "source": [
    "## In Memory Cache"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "426ff912",
   "metadata": {},
   "outputs": [],
   "source": [
    "import langchain\n",
    "from langchain.cache import InMemoryCache\n",
    "langchain.llm_cache = InMemoryCache()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "f69f6283",
   "metadata": {},
   "outputs": [],
   "source": [
    "# To make the caching really obvious, lets use a slower model.\n",
    "llm = OpenAI(model_name=\"text-davinci-002\", n=2, best_of=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "64005d1f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 30.7 ms, sys: 18.6 ms, total: 49.3 ms\n",
      "Wall time: 791 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "\"\\n\\nWhy couldn't the bicycle stand up by itself? Because it was...two tired!\""
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "# The first time, it is not yet in cache, so it should take longer\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "c8a1cb2b",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 80 µs, sys: 0 ns, total: 80 µs\n",
      "Wall time: 83.9 µs\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "\"\\n\\nWhy couldn't the bicycle stand up by itself? Because it was...two tired!\""
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "# The second time it is, so it goes faster\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4bf59c12",
   "metadata": {},
   "source": [
    "## SQLite Cache"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "3ff65b00",
   "metadata": {},
   "outputs": [],
   "source": [
    "!rm .langchain.db"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "5f036236",
   "metadata": {},
   "outputs": [],
   "source": [
    "# We can do the same thing with a SQLite cache\n",
    "from langchain.cache import SQLiteCache\n",
    "langchain.llm_cache = SQLiteCache(database_path=\".langchain.db\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "fa18e3af",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 17 ms, sys: 9.76 ms, total: 26.7 ms\n",
      "Wall time: 825 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'\\n\\nWhy did the chicken cross the road?\\n\\nTo get to the other side.'"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "# The first time, it is not yet in cache, so it should take longer\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "5bf2f6fd",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 2.46 ms, sys: 1.23 ms, total: 3.7 ms\n",
      "Wall time: 2.67 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'\\n\\nWhy did the chicken cross the road?\\n\\nTo get to the other side.'"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "# The second time it is, so it goes faster\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "278ad7ae",
   "metadata": {},
   "source": [
    "## Redis Cache"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "39f6eb0b",
   "metadata": {},
   "outputs": [],
   "source": [
    "# We can do the same thing with a Redis cache\n",
    "# (make sure your local Redis instance is running first before running this example)\n",
    "from redis import Redis\n",
    "from langchain.cache import RedisCache\n",
    "langchain.llm_cache = RedisCache(redis_=Redis())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "28920749",
   "metadata": {},
   "outputs": [],
   "source": [
    "%%time\n",
    "# The first time, it is not yet in cache, so it should take longer\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "94bf9415",
   "metadata": {},
   "outputs": [],
   "source": [
    "%%time\n",
    "# The second time it is, so it goes faster\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "934943dc",
   "metadata": {},
   "source": [
    "## SQLAlchemy Cache"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "acccff40",
   "metadata": {},
   "outputs": [],
   "source": [
    "# You can use SQLAlchemyCache to cache with any SQL database supported by SQLAlchemy.\n",
    "\n",
    "# from langchain.cache import SQLAlchemyCache\n",
    "# from sqlalchemy import create_engine\n",
    "\n",
    "# engine = create_engine(\"postgresql://postgres:postgres@localhost:5432/postgres\")\n",
    "# langchain.llm_cache = SQLAlchemyCache(engine)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0959d640",
   "metadata": {},
   "source": [
    "### Custom SQLAlchemy Schemas"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ac967b39",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 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:\n",
    "\n",
    "from sqlalchemy import Column, Integer, String, Computed, Index, Sequence\n",
    "from sqlalchemy import create_engine\n",
    "from sqlalchemy.ext.declarative import declarative_base\n",
    "from sqlalchemy_utils import TSVectorType\n",
    "from langchain.cache import SQLAlchemyCache\n",
    "\n",
    "Base = declarative_base()\n",
    "\n",
    "\n",
    "class FulltextLLMCache(Base):  # type: ignore\n",
    "    \"\"\"Postgres table for fulltext-indexed LLM Cache\"\"\"\n",
    "\n",
    "    __tablename__ = \"llm_cache_fulltext\"\n",
    "    id = Column(Integer, Sequence('cache_id'), primary_key=True)\n",
    "    prompt = Column(String, nullable=False)\n",
    "    llm = Column(String, nullable=False)\n",
    "    idx = Column(Integer)\n",
    "    response = Column(String)\n",
    "    prompt_tsv = Column(TSVectorType(), Computed(\"to_tsvector('english', llm || ' ' || prompt)\", persisted=True))\n",
    "    __table_args__ = (\n",
    "        Index(\"idx_fulltext_prompt_tsv\", prompt_tsv, postgresql_using=\"gin\"),\n",
    "    )\n",
    "\n",
    "engine = create_engine(\"postgresql://postgres:postgres@localhost:5432/postgres\")\n",
    "langchain.llm_cache = SQLAlchemyCache(engine, FulltextLLMCache)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0c69d84d",
   "metadata": {},
   "source": [
    "## Optional Caching\n",
    "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"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "6af46e2b",
   "metadata": {},
   "outputs": [],
   "source": [
    "llm = OpenAI(model_name=\"text-davinci-002\", n=2, best_of=2, cache=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "26c4fd8f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 5.8 ms, sys: 2.71 ms, total: 8.51 ms\n",
      "Wall time: 745 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'\\n\\nWhy did the chicken cross the road?\\n\\nTo get to the other side!'"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "46846b20",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 4.91 ms, sys: 2.64 ms, total: 7.55 ms\n",
      "Wall time: 623 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'\\n\\nTwo guys stole a calendar. They got six months each.'"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "llm(\"Tell me a joke\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5da41b77",
   "metadata": {},
   "source": [
    "## Optional Caching in Chains\n",
    "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.\n",
    "\n",
    "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."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "9afa3f7a",
   "metadata": {},
   "outputs": [],
   "source": [
    "llm = OpenAI(model_name=\"text-davinci-002\")\n",
    "no_cache_llm = OpenAI(model_name=\"text-davinci-002\", cache=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "98a78e8e",
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.text_splitter import CharacterTextSplitter\n",
    "from langchain.chains.mapreduce import MapReduceChain\n",
    "\n",
    "text_splitter = CharacterTextSplitter()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "2bfb099b",
   "metadata": {},
   "outputs": [],
   "source": [
    "with open('../../../state_of_the_union.txt') as f:\n",
    "    state_of_the_union = f.read()\n",
    "texts = text_splitter.split_text(state_of_the_union)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "f78b7f51",
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.docstore.document import Document\n",
    "docs = [Document(page_content=t) for t in texts[:3]]\n",
    "from langchain.chains.summarize import load_summarize_chain"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "a2a30822",
   "metadata": {},
   "outputs": [],
   "source": [
    "chain = load_summarize_chain(llm, chain_type=\"map_reduce\", reduce_llm=no_cache_llm)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "a545b743",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 452 ms, sys: 60.3 ms, total: 512 ms\n",
      "Wall time: 5.09 s\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'\\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.'"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "chain.run(docs)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3ed85e9d",
   "metadata": {},
   "source": [
    "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."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "39cbb282",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 11.5 ms, sys: 4.33 ms, total: 15.8 ms\n",
      "Wall time: 1.04 s\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "'\\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.'"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "chain.run(docs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9df0dab8",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}