主题
在本 notebook 中，我们将展示如何针对 Pinecone 执行自动检索，它让你能够执行广泛的半结构化查询，超越标准 top-k 语义搜索的能力范围。
我们将展示如何设置基本的自动检索，以及如何扩展它（通过自定义提示词和动态元数据检索）。
如果你在 Colab 中打开这个 Notebook，可能需要先安装 LlamaIndex 🦙。
%pip install llama-index-vector-stores-pinecone# !pip install llama-index>=0.9.31 scikit-learn==1.2.2 arize-phoenix==2.4.1 pinecone-client>=3.0.0第 1 部分：设置自动检索
要设置自动检索，请执行以下操作：
- 我们将进行一些设置、加载数据并构建 Pinecone 向量索引。
- 我们将定义自动检索器并运行一些示例查询。
- 我们将使用 Phoenix 观察每个跟踪并可视化提示词的输入/输出。
- 我们将向你展示如何自定义自动检索提示词。
1.a 设置 Pinecone/Phoenix、加载数据并构建向量索引
在本节中，我们设置 Pinecone 并导入一些关于书籍/电影的玩具数据（包含文本数据和元数据）。
我们还设置 Phoenix，以便它捕获下游跟踪。
# setup Phoenix
import phoenix as px
import llama_index.core
px.launch_app()
llama_index.core.set_global_handler("arize_phoenix")🌍 To view the Phoenix app in your browser, visit http://127.0.0.1:6006/
📺 To view the Phoenix app in a notebook, run `px.active_session().view()`
📖 For more information on how to use Phoenix, check out https://docs.arize.com/phoeniximport os
os.environ[
"PINECONE_API_KEY"
] = "<Your Pinecone API key, from app.pinecone.io>"
# os.environ["OPENAI_API_KEY"] = "sk-..."from pinecone import Pinecone
from pinecone import ServerlessSpec
api_key = os.environ["PINECONE_API_KEY"]
pc = Pinecone(api_key=api_key)# delete if needed
# pc.delete_index("quickstart-index")# Dimensions are for text-embedding-ada-002
try:
pc.create_index(
"quickstart-index",
dimension=1536,
metric="euclidean",
spec=ServerlessSpec(cloud="aws", region="us-west-2"),
)
except Exception as e:
# Most likely index already exists
print(e)
passpinecone_index = pc.Index("quickstart-index")加载文档，构建 PineconeVectorStore 和 VectorStoreIndex
from llama_index.core import VectorStoreIndex, StorageContext
from llama_index.vector_stores.pinecone import PineconeVectorStorefrom llama_index.core.schema import TextNode
nodes = [
TextNode(
text="The Shawshank Redemption",
metadata={
"author": "Stephen King",
"theme": "Friendship",
"year": 1994,
},
),
TextNode(
text="The Godfather",
metadata={
"director": "Francis Ford Coppola",
"theme": "Mafia",
"year": 1972,
},
),
TextNode(
text="Inception",
metadata={
"director": "Christopher Nolan",
"theme": "Fiction",
"year": 2010,
},
),
TextNode(
text="To Kill a Mockingbird",
metadata={
"author": "Harper Lee",
"theme": "Fiction",
"year": 1960,
},
),
TextNode(
text="1984",
metadata={
"author": "George Orwell",
"theme": "Totalitarianism",
"year": 1949,
},
),
TextNode(
text="The Great Gatsby",
metadata={
"author": "F. Scott Fitzgerald",
"theme": "The American Dream",
"year": 1925,
},
),
TextNode(
text="Harry Potter and the Sorcerer's Stone",
metadata={
"author": "J.K. Rowling",
"theme": "Fiction",
"year": 1997,
},
),
]vector_store = PineconeVectorStore(
pinecone_index=pinecone_index,
namespace="test",
)
storage_context = StorageContext.from_defaults(vector_store=vector_store)index = VectorStoreIndex(nodes, storage_context=storage_context)Upserted vectors: 0%| | 0/7 [00:00<?, ?it/s]1.b 定义自动检索器，运行一些示例查询
设置 VectorIndexAutoRetriever
其中一个输入是描述向量存储集合包含哪些内容的 schema。这类似于描述 SQL 数据库中某张表的表 schema。然后，此 schema 信息会被注入提示词，并传给 LLM，以推断完整查询应该是什么（包括元数据过滤器）。
from llama_index.core.retrievers import VectorIndexAutoRetriever
from llama_index.core.vector_stores import MetadataInfo, VectorStoreInfo
vector_store_info = VectorStoreInfo(
content_info="famous books and movies",
metadata_info=[
MetadataInfo(
name="director",
type="str",
description=("Name of the director"),
),
MetadataInfo(
name="theme",
type="str",
description=("Theme of the book/movie"),
),
MetadataInfo(
name="year",
type="int",
description=("Year of the book/movie"),
),
],
)
retriever = VectorIndexAutoRetriever(
index,
vector_store_info=vector_store_info,
empty_query_top_k=10,
# this is a hack to allow for blank queries in pinecone
default_empty_query_vector=[0] * 1536,
verbose=True,
)运行一些查询
让我们运行一些利用结构化信息的示例查询。
nodes = retriever.retrieve(
"Tell me about some books/movies after the year 2000"
)Using query str:
Using filters: [('year', '>', 2000)]for node in nodes:
print(node.text)
print(node.metadata)Inception
{'director': 'Christopher Nolan', 'theme': 'Fiction', 'year': 2010}nodes = retriever.retrieve("Tell me about some books that are Fiction")Using query str: Fiction
Using filters: [('theme', '==', 'Fiction')]for node in nodes:
print(node.text)
print(node.metadata)Inception
{'director': 'Christopher Nolan', 'theme': 'Fiction', 'year': 2010}
To Kill a Mockingbird
{'author': 'Harper Lee', 'theme': 'Fiction', 'year': 1960}传入额外的元数据过滤器
如果你有希望传入但未被自动推断的额外元数据过滤器，请执行以下操作。
from llama_index.core.vector_stores import MetadataFilters
filter_dicts = [{"key": "year", "operator": "==", "value": 1997}]
filters = MetadataFilters.from_dicts(filter_dicts)
retriever2 = VectorIndexAutoRetriever(
index,
vector_store_info=vector_store_info,
empty_query_top_k=10,
# this is a hack to allow for blank queries in pinecone
default_empty_query_vector=[0] * 1536,
extra_filters=filters,
)nodes = retriever2.retrieve("Tell me about some books that are Fiction")
for node in nodes:
print(node.text)
print(node.metadata)Harry Potter and the Sorcerer's Stone
{'author': 'J.K. Rowling', 'theme': 'Fiction', 'year': 1997}失败查询示例
请注意，没有检索到任何结果！我们稍后会修复这个问题。
nodes = retriever.retrieve("Tell me about some books that are mafia-themed")Using query str: books
Using filters: [('theme', '==', 'mafia')]for node in nodes:
print(node.text)
print(node.metadata)可视化跟踪
让我们打开 Phoenix 查看跟踪！
让我们看看自动检索提示词。我们看到自动检索提示词使用了两个少样本示例。
第 2 部分：扩展自动检索（使用动态元数据检索）
我们现在通过自定义提示词来扩展自动检索。在第一部分中，我们显式添加一些规则。
在第二部分中，我们实现动态元数据检索，它将执行第一阶段检索，从向量数据库中获取相关元数据，并将其作为少样本示例插入自动检索提示词。（当然，第二阶段检索会从向量数据库中检索实际条目）。
2.a 改进自动检索提示词
我们的自动检索提示词可以工作，但可以在多个方面改进。例如，它包含 2 个硬编码的少样本示例（你如何包含自己的示例？），以及自动检索并不“总是”能推断出正确的元数据过滤器。
例如，所有 theme 字段都是大写的。我们如何告知 LLM 这一点，以免它错误地推断出小写的 “theme”？
让我们尝试修改提示词！
from llama_index.core.prompts import display_prompt_dict
from llama_index.core import PromptTemplateprompts_dict = retriever.get_prompts()display_prompt_dict(prompts_dict)# look at required template variables.
prompts_dict["prompt"].template_vars['schema_str', 'info_str', 'query_str']自定义提示词
让我们稍微自定义提示词。我们执行以下操作：
- 移除第一个少样本示例以节省 token
- 添加一条消息，在推断 “theme” 时始终将首字母大写。
请注意，提示词模板要求定义 schema_str、info_str 和 query_str。
# write prompt template, and modify it.
prompt_tmpl_str = """\
Your goal is to structure the user's query to match the request schema provided below.
<< Structured Request Schema >>
When responding use a markdown code snippet with a JSON object formatted in the following schema:
{schema_str}
The query string should contain only text that is expected to match the contents of documents. Any conditions in the filter should not be mentioned in the query as well.
Make sure that filters only refer to attributes that exist in the data source.
Make sure that filters take into account the descriptions of attributes.
Make sure that filters are only used as needed. If there are no filters that should be applied return [] for the filter value.
If the user's query explicitly mentions number of documents to retrieve, set top_k to that number, otherwise do not set top_k.
Do NOT EVER infer a null value for a filter. This will break the downstream program. Instead, don't include the filter.
<< Example 1. >>
Data Source:
```json
{{
"metadata_info": [
{{
"name": "author",
"type": "str",
"description": "Author name"
}},
{{
"name": "book_title",
"type": "str",
"description": "Book title"
}},
{{
"name": "year",
"type": "int",
"description": "Year Published"
}},
{{
"name": "pages",
"type": "int",
"description": "Number of pages"
}},
{{
"name": "summary",
"type": "str",
"description": "A short summary of the book"
}}
],
"content_info": "Classic literature"
}}用户查询：Jane Austen 在 1813 年之后出版、探讨为社会地位而婚姻这一主题的书有哪些？
附加说明：无
结构化请求：
{{"query": "Books related to theme of marriage for social standing", "filters": [{{"key": "year", "value": "1813", "operator": ">"}}, {{"key": "author", "value": "Jane Austen", "operator": "=="}}], "top_k": null}}<< 示例 2. >> 数据源：
{info_str}用户查询：{query_str}
附加说明：{additional_instructions}
结构化请求："""
```python
prompt_tmpl = PromptTemplate(prompt_tmpl_str)你会注意到我们添加了一个 additional_instructions 模板变量。这允许我们插入特定于向量集合的说明。
我们将使用 partial_format 添加该说明。
add_instrs = """\
If one of the filters is 'theme', please make sure that the first letter of the inferred value is capitalized. Only words that are capitalized are valid values for "theme". \
"""
prompt_tmpl = prompt_tmpl.partial_format(additional_instructions=add_instrs)retriever.update_prompts({"prompt": prompt_tmpl})重新运行一些查询
现在让我们尝试重新运行一些查询，我们将看到该值被自动推断出来。
nodes = retriever.retrieve(
"Tell me about some books that are friendship-themed"
)for node in nodes:
print(node.text)
print(node.metadata)2.b 实现动态元数据检索
除了在提示词中硬编码规则外，另一种选择是检索相关的元数据少样本示例，以帮助 LLM 更好地推断正确的元数据过滤器。
这将更好地防止 LLM 在推断 “where” 子句时出错，尤其是在值的拼写/正确格式等方面。
我们可以通过向量检索来实现这一点。现有的向量数据库集合存储原始文本 + 元数据；我们可以直接查询此集合，或者单独仅索引元数据并从那里检索。在本节中，我们选择前者，但在实践中你可能希望选择后者。
# define retriever that fetches the top 2 examples.
metadata_retriever = index.as_retriever(similarity_top_k=2)我们使用上一节中定义的同一个 prompt_tmpl_str。
from typing import List, Any
def format_additional_instrs(**kwargs: Any) -> str:
"""Format examples into a string."""
nodes = metadata_retriever.retrieve(kwargs["query_str"])
context_str = (
"Here is the metadata of relevant entries from the database collection. "
"This should help you infer the right filters: \n"
)
for node in nodes:
context_str += str(node.node.metadata) + "\n"
return context_str
ext_prompt_tmpl = PromptTemplate(
prompt_tmpl_str,
function_mappings={"additional_instructions": format_additional_instrs},
)retriever.update_prompts({"prompt": ext_prompt_tmpl})重新运行一些查询
现在让我们尝试重新运行一些查询，我们将看到该值被自动推断出来。
nodes = retriever.retrieve("Tell me about some books that are mafia-themed")
for node in nodes:
print(node.text)
print(node.metadata)Using query str: books
Using filters: [('theme', '==', 'Mafia')]
The Godfather
{'director': 'Francis Ford Coppola', 'theme': 'Mafia', 'year': 1972}nodes = retriever.retrieve("Tell me some books authored by HARPER LEE")
for node in nodes:
print(node.text)
print(node.metadata)Using query str: Books authored by Harper Lee
Using filters: [('author', '==', 'Harper Lee')]
To Kill a Mockingbird
{'author': 'Harper Lee', 'theme': 'Fiction', 'year': 1960}