Ask a URL using Llama 2 running locally

ollama-langchain.py

# Ask questions about a document using Ollama and Langchain.
# This leverages a local LLM to provide insights into any document.
# It's far from perfect, as it's power is limited by the LLM size.

# It leverages the langchain library to perform tasks like document loading,
# text splitting, embedding generation, and using the Ollama client for generating answers.

# You need to have Ollama installed (with model llama2:13b available). See https://github.com/jmorganca/ollama

# Usage:
# python ollama-langchain.py [URL]

# Dependencies:
# pip install langchain bs4 gpt4all chromadb

# Script based on https://github.com/jmorganca/ollama/blob/main/docs/tutorials/langchainpy.md

system_prompt = """
Answer based on the document provided only, unless otherwise stated. If you don't know the answer, type "I don't know".
"""

# Init the Ollama client
from langchain.llms import Ollama

ollama = Ollama(base_url="http://localhost:11434", model="llama2:13b", temperature=0)

#  Load the text from the argument. If its a local file, load from the path. If its a URL, load from the URL.
import sys
from langchain.document_loaders import WebBaseLoader

file_or_url = sys.argv[1]
print("Loading from", file_or_url)
if file_or_url.startswith("http"):
    loader = WebBaseLoader(web_path=[file_or_url])
else:
    raise Exception("Only URLs are supported right now")
data = loader.load()

# Split the text into chunks
# See https://github.com/langchain-ai/langchain/issues/2026
from langchain.text_splitter import RecursiveCharacterTextSplitter

text_splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=10)

all_splits = text_splitter.split_documents(data)

# Generate embeddings
from langchain.embeddings import GPT4AllEmbeddings
from langchain.vectorstores import Chroma

vectorstore = Chroma.from_documents(documents=all_splits, embedding=GPT4AllEmbeddings())

while True:
    # Ask for input
    question = input("> ")

    # Gather relevant parts of the document
    docs = vectorstore.similarity_search(question)

    # Send question and the relevant parts of the docs to the model
    from langchain.chains import RetrievalQA

    qachain = RetrievalQA.from_chain_type(ollama, retriever=vectorstore.as_retriever())
    response = qachain({"query": system_prompt + question})

    print(response["result"])