File size: 12,021 Bytes
48a66db |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 |
import os
import re
import logging
import shutil
import string
import pinecone
import chromadb
import json, jsonlines
from tqdm import tqdm
from langchain_community.vectorstores import Pinecone
from langchain_community.vectorstores import Chroma
from langchain.text_splitter import RecursiveCharacterTextSplitter, CharacterTextSplitter
from langchain_openai import OpenAIEmbeddings
from langchain_community.embeddings import VoyageEmbeddings
from langchain_community.document_loaders import PyPDFLoader
from langchain_core.documents import Document as lancghain_Document
from ragatouille import RAGPretrainedModel
from dotenv import load_dotenv,find_dotenv
load_dotenv(find_dotenv(),override=True)
# Set secrets from environment file
OPENAI_API_KEY=os.getenv('OPENAI_API_KEY')
VOYAGE_API_KEY=os.getenv('VOYAGE_API_KEY')
PINECONE_API_KEY=os.getenv('PINECONE_API_KEY')
HUGGINGFACEHUB_API_TOKEN=os.getenv('HUGGINGFACEHUB_API_TOKEN')
def chunk_docs(docs,
chunk_method='tiktoken_recursive',
file=None,
chunk_size=500,
chunk_overlap=0,
use_json=False):
docs_out=[]
if file:
logging.info('Jsonl file to be used: '+file)
if use_json and os.path.exists(file):
logging.info('Jsonl file found, using this instead of parsing docs.')
with open(file, "r") as file_in:
file_data = [json.loads(line) for line in file_in]
# Process the file data and put it into the same format as docs_out
for line in file_data:
doc_temp = lancghain_Document(page_content=line['page_content'],
source=line['metadata']['source'],
page=line['metadata']['page'],
metadata=line['metadata'])
if has_meaningful_content(doc_temp):
docs_out.append(doc_temp)
logging.info('Parsed: '+file)
logging.info('Number of entries: '+str(len(docs_out)))
logging.info('Sample entries:')
logging.info(str(docs_out[0]))
logging.info(str(docs_out[-1]))
else:
logging.info('No jsonl found. Reading and parsing docs.')
logging.info('Chunk size (tokens): '+str(chunk_size))
logging.info('Chunk overlap (tokens): '+str(chunk_overlap))
for doc in tqdm(docs,desc='Reading and parsing docs'):
logging.info('Parsing: '+doc)
loader = PyPDFLoader(doc)
data = loader.load_and_split()
if chunk_method=='tiktoken_recursive':
text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
else:
raise NotImplementedError
pages = text_splitter.split_documents(data)
# Tidy up text by removing unnecessary characters
for page in pages:
page.metadata['source']=os.path.basename(page.metadata['source']) # Strip path
page.metadata['page']=int(page.metadata['page'])+1 # Pages are 0 based, update
page.page_content=re.sub(r"(\w+)-\n(\w+)", r"\1\2", page.page_content) # Merge hyphenated words
page.page_content = re.sub(r"(?<!\n\s)\n(?!\s\n)", " ", page.page_content.strip()) # Fix newlines in the middle of sentences
page.page_content = re.sub(r"\n\s*\n", "\n\n", page.page_content) # Remove multiple newlines
# Add metadata to the end of the page content, some RAG models don't have metadata.
page.page_content += str(page.metadata)
doc_temp=lancghain_Document(page_content=page.page_content,
source=page.metadata['source'],
page=page.metadata['page'],
metadata=page.metadata)
if has_meaningful_content(page):
docs_out.append(doc_temp)
logging.info('Parsed: '+doc)
logging.info('Sample entries:')
logging.info(str(docs_out[0]))
logging.info(str(docs_out[-1]))
if file:
# Write to a jsonl file, save it.
logging.info('Writing to jsonl file: '+file)
with jsonlines.open(file, mode='w') as writer:
for doc in docs_out:
writer.write(doc.dict())
logging.info('Written: '+file)
return docs_out
def load_docs(index_type,
docs,
query_model,
index_name=None,
chunk_method='tiktoken_recursive',
chunk_size=500,
chunk_overlap=0,
clear=False,
use_json=False,
file=None,
batch_size=50):
"""
Loads PDF documents. If index_name is blank, it will return a list of the data (texts). If it is a name of a pinecone storage, it will return the vector_store.
"""
# Chunk docs
docs_out=chunk_docs(docs,
chunk_method=chunk_method,
file=file,
chunk_size=chunk_size,
chunk_overlap=chunk_overlap,
use_json=use_json)
# Initialize client
db_path='../db/'
if index_name:
if index_type=="Pinecone":
# Import and initialize Pinecone client
pinecone.init(
api_key=PINECONE_API_KEY
)
# Find the existing index, clear for new start
if clear:
try:
pinecone.describe_index(index_name)
except:
raise Exception(f"Cannot clear index {index_name} because it does not exist.")
index=pinecone.Index(index_name)
index.delete(delete_all=True) # Clear the index first, then upload
logging.info('Cleared database '+index_name)
# Upsert docs
try:
pinecone.describe_index(index_name)
except:
logging.info(f"Index {index_name} does not exist. Creating new index.")
logging.info('Size of embedding used: '+str(embedding_size(query_model))) # TODO: set this to be backed out of the embedding size
pinecone.create_index(index_name,dimension=embedding_size(query_model))
logging.info(f"Index {index_name} created. Adding {len(docs_out)} entries to index.")
pass
else:
logging.info(f"Index {index_name} exists. Adding {len(docs_out)} entries to index.")
index = pinecone.Index(index_name)
vectorstore = Pinecone(index, query_model, "page_content") # Set the vector store to calculate embeddings on page_content
vectorstore = batch_upsert(index_type,
vectorstore,
docs_out,
batch_size=batch_size)
elif index_type=="ChromaDB":
# Upsert docs. Defaults to putting this in the ../db directory
logging.info(f"Creating new index {index_name}.")
persistent_client = chromadb.PersistentClient(path=db_path+'/chromadb')
vectorstore = Chroma(client=persistent_client,
collection_name=index_name,
embedding_function=query_model)
logging.info(f"Index {index_name} created. Adding {len(docs_out)} entries to index.")
vectorstore = batch_upsert(index_type,
vectorstore,
docs_out,
batch_size=batch_size)
logging.info("Documents upserted to f{index_name}.")
# Test query
test_query = vectorstore.similarity_search('What are examples of aerosapce adhesives to avoid?')
logging.info('Test query: '+str(test_query))
if not test_query:
raise ValueError("Chroma vector database is not configured properly. Test query failed.")
elif index_type=="RAGatouille":
logging.info(f'Setting up RAGatouille model {query_model}')
vectorstore = RAGPretrainedModel.from_pretrained(query_model)
logging.info('RAGatouille model set: '+str(vectorstore))
# Create an index from the vectorstore.
docs_out_colbert = [doc.page_content for doc in docs_out]
if chunk_size>500:
raise ValueError("RAGatouille cannot handle chunks larger than 500 tokens. Reduce token count.")
vectorstore.index(
collection=docs_out_colbert,
index_name=index_name,
max_document_length=chunk_size,
overwrite_index=True,
split_documents=True,
)
logging.info(f"Index created: {vectorstore}")
# Move the directory to the db folder
logging.info(f"Moving RAGatouille index to {db_path}")
ragatouille_path = os.path.join(db_path, '.ragatouille')
if os.path.exists(ragatouille_path):
shutil.rmtree(ragatouille_path)
logging.info(f"RAGatouille index deleted from {ragatouille_path}")
shutil.move('./.ragatouille', db_path)
logging.info(f"RAGatouille index created in {db_path}:"+str(vectorstore))
# Return vectorstore or docs
if index_name:
return vectorstore
else:
return docs_out
def delete_index(index_type,index_name):
"""
Deletes an existing Pinecone index with the given index_name.
"""
if index_type=="Pinecone":
# Import and initialize Pinecone client
pinecone.init(
api_key=PINECONE_API_KEY
)
try:
pinecone.describe_index(index_name)
logging.info(f"Index {index_name} exists.")
except:
raise Exception(f"Index {index_name} does not exist, cannot delete.")
else:
pinecone.delete_index(index_name)
logging.info(f"Index {index_name} deleted.")
elif index_type=="ChromaDB":
# Delete existing collection
logging.info(f"Deleting index {index_name}.")
persistent_client = chromadb.PersistentClient(path='../db/chromadb')
persistent_client.delete_collection(name=index_name)
logging.info("Index deleted.")
elif index_type=="RAGatouille":
raise NotImplementedError
def batch_upsert(index_type,vectorstore,docs_out,batch_size=50):
# Batch insert the chunks into the vector store
for i in range(0, len(docs_out), batch_size):
chunk_batch = docs_out[i:i + batch_size]
if index_type=="Pinecone":
vectorstore.add_documents(chunk_batch)
elif index_type=="ChromaDB":
vectorstore.add_documents(chunk_batch) # Happens to be same for chroma/pinecone, leaving if statement just in case
return vectorstore
def has_meaningful_content(page):
"""
Test whether the page has more than 30% words and is more than 5 words.
"""
text=page.page_content
num_words = len(text.split())
alphanumeric_pct = sum(c.isalnum() for c in text) / len(text)
if num_words < 5 or alphanumeric_pct < 0.3:
return False
else:
return True
def embedding_size(embedding_model):
"""
Returns the embedding size of the model.
"""
if isinstance(embedding_model,OpenAIEmbeddings):
return 1536 # https://platform.openai.com/docs/models/embeddings, test-embedding-ada-002
elif isinstance(embedding_model,VoyageEmbeddings):
return 1024 # https://docs.voyageai.com/embeddings/, voyage-02
else:
raise NotImplementedError |