Spaces:
Sleeping
Sleeping
Merge pull request #1 from leftbyte/dphung/build_pinecone_index
Browse files- requirements.txt +8 -0
- scripts/docsToEmbeddings.py +209 -0
requirements.txt
CHANGED
|
@@ -1,4 +1,12 @@
|
|
| 1 |
black
|
|
|
|
|
|
|
| 2 |
gradio
|
|
|
|
| 3 |
openai
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
typer
|
|
|
|
|
|
| 1 |
black
|
| 2 |
+
bs4
|
| 3 |
+
fastparquet
|
| 4 |
gradio
|
| 5 |
+
langchain
|
| 6 |
openai
|
| 7 |
+
pandas
|
| 8 |
+
pinecone-client
|
| 9 |
+
pyarrow
|
| 10 |
+
tiktoken
|
| 11 |
typer
|
| 12 |
+
unstructured
|
scripts/docsToEmbeddings.py
ADDED
|
@@ -0,0 +1,209 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
"""
|
| 3 |
+
Builds a Pinecone index by converting a given set of documents to its embeddings.
|
| 4 |
+
|
| 5 |
+
To run, you must first create an account with Pinecone and create an
|
| 6 |
+
index in the UI with the appropriate embedding dimension (1536 if you
|
| 7 |
+
are using text-embedding-ada-002 like this script). You also need an
|
| 8 |
+
OpenAI API key. This implementation relies on the fact that the
|
| 9 |
+
documentation is hosted with Gitbook. If your documentation does not use
|
| 10 |
+
Gitbook, you should use a different document loader.
|
| 11 |
+
"""
|
| 12 |
+
|
| 13 |
+
import argparse
|
| 14 |
+
import logging
|
| 15 |
+
import sys
|
| 16 |
+
from functools import partial
|
| 17 |
+
from typing import Dict, List, Optional
|
| 18 |
+
|
| 19 |
+
import numpy as np
|
| 20 |
+
import openai
|
| 21 |
+
import pandas as pd
|
| 22 |
+
import pinecone # type: ignore
|
| 23 |
+
import tiktoken
|
| 24 |
+
from langchain.docstore.document import Document
|
| 25 |
+
from langchain.document_loaders import GitbookLoader
|
| 26 |
+
from langchain.document_loaders import DirectoryLoader
|
| 27 |
+
from langchain.embeddings.base import Embeddings
|
| 28 |
+
from langchain.embeddings.openai import OpenAIEmbeddings
|
| 29 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
| 30 |
+
from langchain.vectorstores import Pinecone
|
| 31 |
+
from tiktoken import Encoding
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def load_local_docs(docs_path: str) -> List[Document]:
|
| 35 |
+
# By default the DirectoryLoader will use the 'Unstructured'
|
| 36 |
+
# Document. Use mode="elements" if we need more metadat in the
|
| 37 |
+
# index.
|
| 38 |
+
loader = DirectoryLoader(docs_path, glob="**/*.pdf")
|
| 39 |
+
return loader.load()
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def load_gitbook_docs(docs_url: str) -> List[Document]:
|
| 43 |
+
"""
|
| 44 |
+
Loads documentation from a Gitbook URL.
|
| 45 |
+
"""
|
| 46 |
+
|
| 47 |
+
loader = GitbookLoader(
|
| 48 |
+
docs_url,
|
| 49 |
+
load_all_paths=True,
|
| 50 |
+
)
|
| 51 |
+
return loader.load()
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
def tiktoken_len(text: str, tokenizer: Encoding) -> int:
|
| 55 |
+
"""
|
| 56 |
+
Returns the number of tokens in a text.
|
| 57 |
+
"""
|
| 58 |
+
|
| 59 |
+
tokens = tokenizer.encode(text, disallowed_special=())
|
| 60 |
+
return len(tokens)
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
def chunk_docs(
|
| 64 |
+
documents: List[Document], embedding_model_name: str
|
| 65 |
+
) -> List[Document]:
|
| 66 |
+
"""
|
| 67 |
+
Chunks the documents.
|
| 68 |
+
|
| 69 |
+
The chunking strategy used in this function is from the following notebook and accompanying
|
| 70 |
+
video:
|
| 71 |
+
|
| 72 |
+
- https://github.com/pinecone-io/examples/blob/master/generation/langchain/handbook/
|
| 73 |
+
xx-langchain-chunking.ipynb
|
| 74 |
+
- https://www.youtube.com/watch?v=eqOfr4AGLk8
|
| 75 |
+
"""
|
| 76 |
+
|
| 77 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
| 78 |
+
chunk_size=400,
|
| 79 |
+
chunk_overlap=20,
|
| 80 |
+
length_function=partial(
|
| 81 |
+
tiktoken_len,
|
| 82 |
+
tokenizer=tiktoken.encoding_for_model(embedding_model_name),
|
| 83 |
+
),
|
| 84 |
+
separators=["\n\n", "\n", " ", ""],
|
| 85 |
+
)
|
| 86 |
+
return text_splitter.split_documents(documents)
|
| 87 |
+
|
| 88 |
+
|
| 89 |
+
def build_pinecone_index(
|
| 90 |
+
documents: List[Document], embeddings: Embeddings, index_name: str
|
| 91 |
+
) -> None:
|
| 92 |
+
"""
|
| 93 |
+
Builds a Pinecone index from a list of documents.
|
| 94 |
+
"""
|
| 95 |
+
|
| 96 |
+
Pinecone.from_documents(
|
| 97 |
+
documents, embeddings, index_name=pinecone_index_name
|
| 98 |
+
)
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
# , engine="pyarrow":
|
| 102 |
+
def save_dataframe_to_parquet(dataframe: pd.DataFrame, save_path: str) -> None:
|
| 103 |
+
"""
|
| 104 |
+
Saves a dataframe to parquet.
|
| 105 |
+
"""
|
| 106 |
+
|
| 107 |
+
dataframe.to_parquet(save_path)
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
class OpenAIEmbeddingsWrapper(OpenAIEmbeddings):
|
| 111 |
+
"""
|
| 112 |
+
Wrapper around OpenAIEmbeddings that stores the query and document embeddings in memory.
|
| 113 |
+
"""
|
| 114 |
+
|
| 115 |
+
query_text_to_embedding: Dict[str, List[float]] = {}
|
| 116 |
+
document_text_to_embedding: Dict[str, List[float]] = {}
|
| 117 |
+
|
| 118 |
+
def embed_query(self, text: str) -> List[float]:
|
| 119 |
+
embedding = super().embed_query(text)
|
| 120 |
+
self.query_text_to_embedding[text] = embedding
|
| 121 |
+
return embedding
|
| 122 |
+
|
| 123 |
+
def embed_documents(
|
| 124 |
+
self, texts: List[str], chunk_size: Optional[int] = 0
|
| 125 |
+
) -> List[List[float]]:
|
| 126 |
+
embeddings = super().embed_documents(texts, chunk_size)
|
| 127 |
+
for text, embedding in zip(texts, embeddings):
|
| 128 |
+
self.document_text_to_embedding[text] = embedding
|
| 129 |
+
return embeddings
|
| 130 |
+
|
| 131 |
+
@property
|
| 132 |
+
def query_embedding_dataframe(self) -> pd.DataFrame:
|
| 133 |
+
return self._convert_text_to_embedding_map_to_dataframe(
|
| 134 |
+
self.query_text_to_embedding
|
| 135 |
+
)
|
| 136 |
+
|
| 137 |
+
@property
|
| 138 |
+
def document_embedding_dataframe(self) -> pd.DataFrame:
|
| 139 |
+
return self._convert_text_to_embedding_map_to_dataframe(
|
| 140 |
+
self.document_text_to_embedding
|
| 141 |
+
)
|
| 142 |
+
|
| 143 |
+
@staticmethod
|
| 144 |
+
def _convert_text_to_embedding_map_to_dataframe(
|
| 145 |
+
text_to_embedding: Dict[str, List[float]]
|
| 146 |
+
) -> pd.DataFrame:
|
| 147 |
+
texts, embeddings = map(list, zip(*text_to_embedding.items()))
|
| 148 |
+
embedding_arrays = [np.array(embedding) for embedding in embeddings]
|
| 149 |
+
return pd.DataFrame.from_dict(
|
| 150 |
+
{
|
| 151 |
+
"text": texts,
|
| 152 |
+
"text_vector": embedding_arrays,
|
| 153 |
+
}
|
| 154 |
+
)
|
| 155 |
+
|
| 156 |
+
|
| 157 |
+
if __name__ == "__main__":
|
| 158 |
+
logging.basicConfig(level=logging.INFO, stream=sys.stdout)
|
| 159 |
+
|
| 160 |
+
parser = argparse.ArgumentParser()
|
| 161 |
+
parser.add_argument("--pinecone-api-key", type=str, help="Pinecone API key")
|
| 162 |
+
parser.add_argument(
|
| 163 |
+
"--pinecone-index-name", type=str, help="Pinecone index name"
|
| 164 |
+
)
|
| 165 |
+
parser.add_argument(
|
| 166 |
+
"--pinecone-environment", type=str, help="Pinecone environment"
|
| 167 |
+
)
|
| 168 |
+
parser.add_argument("--openai-api-key", type=str, help="OpenAI API key")
|
| 169 |
+
parser.add_argument(
|
| 170 |
+
"--output-parquet-path",
|
| 171 |
+
type=str,
|
| 172 |
+
help="Path to output parquet file to save the index",
|
| 173 |
+
)
|
| 174 |
+
parser.add_argument(
|
| 175 |
+
"--docs-path", type=str, help="path containing documentation"
|
| 176 |
+
)
|
| 177 |
+
parser.add_argument(
|
| 178 |
+
"--docs-url",
|
| 179 |
+
type=str,
|
| 180 |
+
help="URL containing documentation"
|
| 181 |
+
# default="https://docs.arize.com/arize/",
|
| 182 |
+
)
|
| 183 |
+
|
| 184 |
+
args = parser.parse_args()
|
| 185 |
+
|
| 186 |
+
pinecone_api_key = args.pinecone_api_key
|
| 187 |
+
pinecone_index_name = args.pinecone_index_name
|
| 188 |
+
pinecone_environment = args.pinecone_environment
|
| 189 |
+
openai_api_key = args.openai_api_key
|
| 190 |
+
output_parquet_path = args.output_parquet_path
|
| 191 |
+
|
| 192 |
+
openai.api_key = openai_api_key
|
| 193 |
+
pinecone.init(api_key=pinecone_api_key, environment=pinecone_environment)
|
| 194 |
+
embedding_model_name = "text-embedding-ada-002"
|
| 195 |
+
|
| 196 |
+
if args.docs_url is not None:
|
| 197 |
+
print(f"docs_url {args.docs_url}")
|
| 198 |
+
documents = load_gitbook_docs(args.docs_url)
|
| 199 |
+
elif args.docs_path is not None:
|
| 200 |
+
documents = load_local_docs(args.docs_path)
|
| 201 |
+
else:
|
| 202 |
+
assert False, "Either the docs_url or docs_path must be set"
|
| 203 |
+
|
| 204 |
+
documents = chunk_docs(documents, embedding_model_name=embedding_model_name)
|
| 205 |
+
embeddings = OpenAIEmbeddingsWrapper(model=embedding_model_name) # type: ignore
|
| 206 |
+
build_pinecone_index(documents, embeddings, pinecone_index_name)
|
| 207 |
+
save_dataframe_to_parquet(
|
| 208 |
+
embeddings.document_embedding_dataframe, output_parquet_path
|
| 209 |
+
)
|