Fleet AI 上下文
Fleet AI Context 是一个包含高质量嵌入式的顶级 1200 个最受欢迎且许可条件宽松的 Python 库及其文档的数据集。
The
Fleet AI团队致力于嵌入世界上最重要的数据。他们从嵌入 top 1200 个Python库开始，以便使用最新的知识进行代码生成。他们很乐意分享他们对LangChain文档和API参考的嵌入。
让我们来看看我们如何使用这些嵌入来为文档检索系统提供动力，最终实现一个简单的代码生成链！
%pip install --upgrade --quiet langchain fleet-context langchain-openai pandas faiss-cpu # faiss-gpu for CUDA supported GPU
from operator import itemgetter
from typing import Any, Optional, Type
import pandas as pd
from langchain.retrievers import MultiVectorRetriever
from langchain_community.vectorstores import FAISS
from langchain_core.documents import Document
from langchain_core.stores import BaseStore
from langchain_core.vectorstores import VectorStore
from langchain_openai import OpenAIEmbeddings
def load_fleet_retriever(
df: pd.DataFrame,
*,
vectorstore_cls: Type[VectorStore] = FAISS,
docstore: Optional[BaseStore] = None,
**kwargs: Any,
):
vectorstore = _populate_vectorstore(df, vectorstore_cls)
if docstore is None:
return vectorstore.as_retriever(**kwargs)
else:
_populate_docstore(df, docstore)
return MultiVectorRetriever(
vectorstore=vectorstore, docstore=docstore, id_key="parent", **kwargs
)
def _populate_vectorstore(
df: pd.DataFrame,
vectorstore_cls: Type[VectorStore],
) -> VectorStore:
if not hasattr(vectorstore_cls, "from_embeddings"):
raise ValueError(
f"Incompatible vector store class {vectorstore_cls}."
"Must implement `from_embeddings` class method."
)
texts_embeddings = []
metadatas = []
for _, row in df.iterrows():
texts_embeddings.append((row.metadata["text"], row["dense_embeddings"]))
metadatas.append(row.metadata)
return vectorstore_cls.from_embeddings(
texts_embeddings,
OpenAIEmbeddings(model="text-embedding-ada-002"),
metadatas=metadatas,
)
def _populate_docstore(df: pd.DataFrame, docstore: BaseStore) -> None:
parent_docs = []
df = df.copy()
df["parent"] = df.metadata.apply(itemgetter("parent"))
for parent_id, group in df.groupby("parent"):
sorted_group = group.iloc[
group.metadata.apply(itemgetter("section_index")).argsort()
]
text = "".join(sorted_group.metadata.apply(itemgetter("text")))
metadata = {
k: sorted_group.iloc[0].metadata[k] for k in ("title", "type", "url")
}
text = metadata["title"] + "\n" + text
metadata["id"] = parent_id
parent_docs.append(Document(page_content=text, metadata=metadata))
docstore.mset(((d.metadata["id"], d) for d in parent_docs))
检索片段
在嵌入处理过程中，Fleet AI 团队首先将长文档分块后再进行嵌入。这意味着这些向量对应于 LangChain 文档页面的部分内容，而不是整个页面。默认情况下，当我们从这些嵌入创建检索器时，我们将检索这些嵌入的片段。
我们将使用 Fleet Context 的 download_embeddings() 来获取 Langchain 的文档嵌入。您可以在 https://fleet.so/context 查看所有受支持库的文档。
from context import download_embeddings
df = download_embeddings("langchain")
vecstore_retriever = load_fleet_retriever(df)
vecstore_retriever.invoke("How does the multi vector retriever work")
其他包
您可以从这个Dropbox链接下载并使用其他嵌入。
Retrieve parent docs
由Fleet AI提供的嵌入包含元数据，可以指示哪些嵌入片段对应于同一原始文档页。如果我们愿意的话，我们可以利用这些信息检索整个父文档，而不仅仅是嵌入片段。幕后我们将使用一个MultiVectorRetriever和一个BaseStore对象来搜索相关的片段，然后将它们映射到其父文档。
from langchain.storage import InMemoryStore
parent_retriever = load_fleet_retriever(
"https://www.dropbox.com/scl/fi/4rescpkrg9970s3huz47l/libraries_langchain_release.parquet?rlkey=283knw4wamezfwiidgpgptkep&dl=1",
docstore=InMemoryStore(),
)
API 参考:InMemoryStore
parent_retriever.invoke("How does the multi vector retriever work")
将其串联起来
让我们尝试在简单的工作流中使用我们的检索系统！
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough
from langchain_openai import ChatOpenAI
prompt = ChatPromptTemplate.from_messages(
[
(
"system",
"""You are a great software engineer who is very familiar \
with Python. Given a user question or request about a new Python library called LangChain and \
parts of the LangChain documentation, answer the question or generate the requested code. \
Your answers must be accurate, should include code whenever possible, and should assume anything \
about LangChain which is note explicitly stated in the LangChain documentation. If the required \
information is not available, just say so.
LangChain Documentation
------------------
{context}""",
),
("human", "{question}"),
]
)
model = ChatOpenAI(model="gpt-3.5-turbo-16k")
chain = (
{
"question": RunnablePassthrough(),
"context": parent_retriever
| (lambda docs: "\n\n".join(d.page_content for d in docs)),
}
| prompt
| model
| StrOutputParser()
)
for chunk in chain.invoke(
"How do I create a FAISS vector store retriever that returns 10 documents per search query"
):
print(chunk, end="", flush=True)