Introduction to LangChain and MongoDB Atlas Vector Search

1
!pip install langchain pypdf pymongo openai python-dotenv tiktoken

1
import os
2
from dotenv import load_dotenv
3
from pymongo import MongoClient
4
5
6
load_dotenv(override=True)
7
8
9
# Add an environment file to the notebook root directory called .env with MONGO_URI="xxx" to load these environment variables
10
11
12
OPENAI_API_KEY = os.environ["OPENAI_API_KEY"]
13
MONGO_URI = os.environ["MONGO_URI"]
14
DB_NAME = "langchain-test-2"
15
COLLECTION_NAME = "test"
16
ATLAS_VECTOR_SEARCH_INDEX_NAME = "default"
17
18
19
EMBEDDING_FIELD_NAME = "embedding"
20
client = MongoClient(MONGO_URI)
21
db = client[DB_NAME]
22
collection = db[COLLECTION_NAME]

1
from langchain.document_loaders import PyPDFLoader
2
from langchain.text_splitter import RecursiveCharacterTextSplitter
3
from langchain.embeddings import OpenAIEmbeddings
4
from langchain.vectorstores import MongoDBAtlasVectorSearch
5
6
7
8
9
loader = PyPDFLoader("https://arxiv.org/pdf/2303.08774.pdf")
10
data = loader.load()
11
12
13
text_splitter = RecursiveCharacterTextSplitter(chunk_size = 500, chunk_overlap = 50)
14
docs = text_splitter.split_documents(data)
15
16
17
# insert the documents in MongoDB Atlas Vector Search
18
x = MongoDBAtlasVectorSearch.from_documents(
19
documents=docs, embedding=OpenAIEmbeddings(disallowed_special=()), collection=MONGODB_COLLECTION, index_name=ATLAS_VECTOR_SEARCH_INDEX_NAME
20
)

1
docs[0]

1
{
2
  "fields": [
3
    {
4
      "type": "vector",
5
      "path": "embedding",  
6
      "numDimensions": 1536, 
7
      "similarity": "cosine"  
8
    },
9
    {
10
      "type": "filter",
11
      "path": "source"  
12
    }
13
   
14
  ]
15
}

1
from langchain.embeddings import OpenAIEmbeddings
2
from langchain.vectorstores import MongoDBAtlasVectorSearch
3
4
5
vector_search = MongoDBAtlasVectorSearch.from_connection_string(
6
   MONGO_URI,
7
   DB_NAME + "." + COLLECTION_NAME,
8
   OpenAIEmbeddings(disallowed_special=()),
9
   index_name=ATLAS_VECTOR_SEARCH_INDEX_NAME
10
)
11
query = "gpt-4"
12
results = vector_search.similarity_search(
13
   query=query,
14
   k=20,
15
)
16
17
18
for result in results:
19
   print( result)

1
qa_retriever = vector_search.as_retriever(
2
   search_type="similarity",
3
   search_kwargs={
4
       "k": 200,
5
       "post_filter_pipeline": [{"$limit": 25}]
6
   }
7
)
8
from langchain.prompts import PromptTemplate
9
prompt_template = """Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer.
10
11
12
{context}
13
14
15
Question: {question}
16
"""
17
PROMPT = PromptTemplate(
18
   template=prompt_template, input_variables=["context", "question"]
19
)
20
from langchain.chains import RetrievalQA
21
from langchain.chat_models import ChatOpenAI
22
from langchain.llms import OpenAI
23
24
25
qa = RetrievalQA.from_chain_type(llm=OpenAI(),chain_type="stuff", retriever=qa_retriever, return_source_documents=True, chain_type_kwargs={"prompt": PROMPT})
26
27
28
docs = qa({"query": "gpt-4 compute requirements"})
29
30
31
print(docs["result"])
32
print(docs['source_documents'])

1
GPT-4 requires a large amount of compute for training, it took 45 petaflops-days of compute to train the model. [Document(page_content='gpt3.5Figure 4. GPT performance on academic and professional exams. In each case, we simulate