上下文增强型函数调用代理¶

在本教程中，我们将向您展示如何使您的智能体具备上下文感知能力。

我们的索引/检索模块有助于简化操作，避免因功能过多而超出提示词容量限制。

初始设置¶

在这里我们设置一个普通的 FunctionAgent，并通过上下文对其进行增强。该代理将在调用任何工具之前首先执行检索操作。这有助于将代理的工具选择和回答能力基于上下文进行落地。

如果您在 Colab 上打开此 Notebook，可能需要安装 LlamaIndex 🦙。

In [ ]:

%pip install llama-index

%pip install llama-index

In [ ]:

import os

os.environ["OPENAI_API_KEY"] = "sk-..."

import os os.environ["OPENAI_API_KEY"] = "sk-..."

In [ ]:

from llama_index.llms.openai import OpenAI
from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.core.settings import Settings

Settings.llm = OpenAI(model="gpt-4o-mini")
Settings.embed_model = OpenAIEmbedding(model="text-embedding-3-small")

from llama_index.llms.openai import OpenAI from llama_index.embeddings.openai import OpenAIEmbedding from llama_index.core.settings import Settings Settings.llm = OpenAI(model="gpt-4o-mini") Settings.embed_model = OpenAIEmbedding(model="text-embedding-3-small")

In [ ]:

import json
from typing import Sequence

from llama_index.core import (
    SimpleDirectoryReader,
    VectorStoreIndex,
    StorageContext,
    load_index_from_storage,
)
from llama_index.core.tools import QueryEngineTool

import json from typing import Sequence from llama_index.core import ( SimpleDirectoryReader, VectorStoreIndex, StorageContext, load_index_from_storage, ) from llama_index.core.tools import QueryEngineTool

In [ ]:

try:
    storage_context = StorageContext.from_defaults(
        persist_dir="./storage/march"
    )
    march_index = load_index_from_storage(storage_context)

    storage_context = StorageContext.from_defaults(
        persist_dir="./storage/june"
    )
    june_index = load_index_from_storage(storage_context)

    storage_context = StorageContext.from_defaults(
        persist_dir="./storage/sept"
    )
    sept_index = load_index_from_storage(storage_context)

    index_loaded = True
except:
    index_loaded = False

try: storage_context = StorageContext.from_defaults( persist_dir="./storage/march" ) march_index = load_index_from_storage(storage_context) storage_context = StorageContext.from_defaults( persist_dir="./storage/june" ) june_index = load_index_from_storage(storage_context) storage_context = StorageContext.from_defaults( persist_dir="./storage/sept" ) sept_index = load_index_from_storage(storage_context) index_loaded = True except: index_loaded = False

下载数据

In [ ]:

!mkdir -p 'data/10q/'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/10q/uber_10q_march_2022.pdf' -O 'data/10q/uber_10q_march_2022.pdf'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/10q/uber_10q_june_2022.pdf' -O 'data/10q/uber_10q_june_2022.pdf'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/10q/uber_10q_sept_2022.pdf' -O 'data/10q/uber_10q_sept_2022.pdf'

!mkdir -p 'data/10q/' !wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/10q/uber_10q_march_2022.pdf' -O 'data/10q/uber_10q_march_2022.pdf' !wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/10q/uber_10q_june_2022.pdf' -O 'data/10q/uber_10q_june_2022.pdf' !wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/docs/examples/data/10q/uber_10q_sept_2022.pdf' -O 'data/10q/uber_10q_sept_2022.pdf'

In [ ]:

# build indexes across the three data sources
if not index_loaded:
    # load data
    march_docs = SimpleDirectoryReader(
        input_files=["./data/10q/uber_10q_march_2022.pdf"]
    ).load_data()
    june_docs = SimpleDirectoryReader(
        input_files=["./data/10q/uber_10q_june_2022.pdf"]
    ).load_data()
    sept_docs = SimpleDirectoryReader(
        input_files=["./data/10q/uber_10q_sept_2022.pdf"]
    ).load_data()

    # build index
    march_index = VectorStoreIndex.from_documents(march_docs)
    june_index = VectorStoreIndex.from_documents(june_docs)
    sept_index = VectorStoreIndex.from_documents(sept_docs)

    # persist index
    march_index.storage_context.persist(persist_dir="./storage/march")
    june_index.storage_context.persist(persist_dir="./storage/june")
    sept_index.storage_context.persist(persist_dir="./storage/sept")

# build indexes across the three data sources if not index_loaded: # load data march_docs = SimpleDirectoryReader( input_files=["./data/10q/uber_10q_march_2022.pdf"] ).load_data() june_docs = SimpleDirectoryReader( input_files=["./data/10q/uber_10q_june_2022.pdf"] ).load_data() sept_docs = SimpleDirectoryReader( input_files=["./data/10q/uber_10q_sept_2022.pdf"] ).load_data() # build index march_index = VectorStoreIndex.from_documents(march_docs) june_index = VectorStoreIndex.from_documents(june_docs) sept_index = VectorStoreIndex.from_documents(sept_docs) # persist index march_index.storage_context.persist(persist_dir="./storage/march") june_index.storage_context.persist(persist_dir="./storage/june") sept_index.storage_context.persist(persist_dir="./storage/sept")

In [ ]:

march_engine = march_index.as_query_engine(similarity_top_k=3)
june_engine = june_index.as_query_engine(similarity_top_k=3)
sept_engine = sept_index.as_query_engine(similarity_top_k=3)

march_engine = march_index.as_query_engine(similarity_top_k=3) june_engine = june_index.as_query_engine(similarity_top_k=3) sept_engine = sept_index.as_query_engine(similarity_top_k=3)

In [ ]:

query_engine_tools = [
    QueryEngineTool.from_defaults(
        query_engine=march_engine,
        name="uber_march_10q",
        description=(
            "Provides information about Uber 10Q filings for March 2022. "
            "Use a detailed plain text question as input to the tool."
        ),
    ),
    QueryEngineTool.from_defaults(
        query_engine=june_engine,
        name="uber_june_10q",
        description=(
            "Provides information about Uber financials for June 2021. "
            "Use a detailed plain text question as input to the tool."
        ),
    ),
    QueryEngineTool.from_defaults(
        query_engine=sept_engine,
        name="uber_sept_10q",
        description=(
            "Provides information about Uber financials for Sept 2021. "
            "Use a detailed plain text question as input to the tool."
        ),
    ),
]

query_engine_tools = [ QueryEngineTool.from_defaults( query_engine=march_engine, name="uber_march_10q", description=( "Provides information about Uber 10Q filings for March 2022. " "Use a detailed plain text question as input to the tool." ), ), QueryEngineTool.from_defaults( query_engine=june_engine, name="uber_june_10q", description=( "Provides information about Uber financials for June 2021. " "Use a detailed plain text question as input to the tool." ), ), QueryEngineTool.from_defaults( query_engine=sept_engine, name="uber_sept_10q", description=( "Provides information about Uber financials for Sept 2021. " "Use a detailed plain text question as input to the tool." ), ), ]

尝试上下文增强型代理¶

在此我们通过不同场景下的上下文来增强代理能力：

• 示例上下文：定义一些映射到金融术语的缩写（例如 R=Revenue）。我们将此作为上下文提供给代理

In [ ]:

from llama_index.core import Document
from llama_index.core.agent.workflow import FunctionAgent

from llama_index.core import Document from llama_index.core.agent.workflow import FunctionAgent

In [ ]:

# toy index - stores a list of abbreviations
texts = [
    "Abbreviation: 'Y' = Revenue",
    "Abbreviation: 'X' = Risk Factors",
    "Abbreviation: 'Z' = Costs",
]
docs = [Document(text=t) for t in texts]
context_index = VectorStoreIndex.from_documents(docs)
context_retriever = context_index.as_retriever(similarity_top_k=2)

# toy index - stores a list of abbreviations texts = [ "Abbreviation: 'Y' = Revenue", "Abbreviation: 'X' = Risk Factors", "Abbreviation: 'Z' = Costs", ] docs = [Document(text=t) for t in texts] context_index = VectorStoreIndex.from_documents(docs) context_retriever = context_index.as_retriever(similarity_top_k=2)

In [ ]:

from llama_index.core.tools import BaseTool

system_prompt_template = """You are a helpful assistant. 
Here is some context that you can use to answer the user's question and for help with picking the right tool:

{context}
"""


async def get_agent_with_context_awareness(
    query: str, context_retriever, tools: list[BaseTool]
) -> FunctionAgent:
    context_nodes = await context_retriever.aretrieve(query)
    context_text = "\n----\n".join([n.node.text for n in context_nodes])

    return FunctionAgent(
        tools=tools,
        llm=OpenAI(model="gpt-4o"),
        system_prompt=system_prompt_template.format(context=context_text),
    )

from llama_index.core.tools import BaseTool system_prompt_template = """You are a helpful assistant. Here is some context that you can use to answer the user's question and for help with picking the right tool: {context} """ async def get_agent_with_context_awareness( query: str, context_retriever, tools: list[BaseTool] ) -> FunctionAgent: context_nodes = await context_retriever.aretrieve(query) context_text = "\n----\n".join([n.node.text for n in context_nodes]) return FunctionAgent( tools=tools, llm=OpenAI(model="gpt-4o"), system_prompt=system_prompt_template.format(context=context_text), )

In [ ]:

query = "What is the 'X' of March 2022?"
agent = await get_agent_with_context_awareness(
    query, context_retriever, query_engine_tools
)

response = await agent.run(query)

query = "What is the 'X' of March 2022?" agent = await get_agent_with_context_awareness( query, context_retriever, query_engine_tools ) response = await agent.run(query)

In [ ]:

print(str(response))

print(str(response))

The risk factors mentioned in Uber's 10-Q filing for March 2022 include uncertainties related to the COVID-19 pandemic, such as the severity and duration of the outbreak, potential future waves or variants of the virus, the administration and efficacy of vaccines, and the impact of governmental actions. There are also concerns regarding the effects on drivers, merchants, consumers, and business partners, as well as other factors that may affect the company's business, results of operations, financial position, and cash flows.

In [ ]:

query = "What is the 'Y' and 'Z' in September 2022?"
agent = await get_agent_with_context_awareness(
    query, context_retriever, query_engine_tools
)

response = await agent.run(query)

query = "What is the 'Y' and 'Z' in September 2022?" agent = await get_agent_with_context_awareness( query, context_retriever, query_engine_tools ) response = await agent.run(query)

In [ ]:

print(str(response))

print(str(response))

In September 2022, Uber's revenue (Y) was $8,343 million, and the total costs (Z) were $8,839 million.

管理上下文/记忆¶

默认情况下，每次 .run() 调用都是无状态的。我们可以通过可序列化的 Context 对象来管理上下文。

In [ ]:

from llama_index.core.workflow import Context

ctx = Context(agent)

query = "What is the 'Y' and 'Z' in September 2022?"
agent = await get_agent_with_context_awareness(
    query, context_retriever, query_engine_tools
)
response = await agent.run(query, ctx=ctx)

query = "What did I just ask?"
agent = await get_agent_with_context_awareness(
    query, context_retriever, query_engine_tools
)
response = await agent.run(query, ctx=ctx)
print(str(response))

from llama_index.core.workflow import Context ctx = Context(agent) query = "What is the 'Y' and 'Z' in September 2022?" agent = await get_agent_with_context_awareness( query, context_retriever, query_engine_tools ) response = await agent.run(query, ctx=ctx) query = "What did I just ask?" agent = await get_agent_with_context_awareness( query, context_retriever, query_engine_tools ) response = await agent.run(query, ctx=ctx) print(str(response))

You asked for the revenue ('Y') and costs ('Z') for Uber in September 2022.

以Uber 10-Q文件为背景，使用计算器工具¶

In [ ]:

from llama_index.core.tools import FunctionTool


def magic_formula(revenue: int, cost: int) -> int:
    """Runs MAGIC_FORMULA on revenue and cost."""
    return revenue - cost


magic_tool = FunctionTool.from_defaults(magic_formula)

from llama_index.core.tools import FunctionTool def magic_formula(revenue: int, cost: int) -> int: """Runs MAGIC_FORMULA on revenue and cost.""" return revenue - cost magic_tool = FunctionTool.from_defaults(magic_formula)

In [ ]:

context_retriever = sept_index.as_retriever(similarity_top_k=3)

context_retriever = sept_index.as_retriever(similarity_top_k=3)

In [ ]:

query = "Can you run MAGIC_FORMULA on Uber's revenue and cost?"
agent = await get_agent_with_context_awareness(
    query, context_retriever, [magic_tool]
)
response = await agent.run(query)
print(str(response))

query = "Can you run MAGIC_FORMULA on Uber's revenue and cost?" agent = await get_agent_with_context_awareness( query, context_retriever, [magic_tool] ) response = await agent.run(query) print(str(response))

The result of running MAGIC_FORMULA on Uber's revenue of $8,343 million and cost of $5,173 million is 3,170.