Welcome to the Prem AI cookbook section. In this recipe, we will build a simple URL Summarizer tool step by step using LangChain Prem AI. The Prem SDK can be used with LangChain for extended development. This example shows one of the ways we can extend development with Prem AI. To give a nice visualization, we use Streamlit, and here is how the final app would look:


So without further ado, let’s get started. You can find the full code here.


This recipe aims to show developers and users how to get started with Prem’s Generative AI Platform and build different use cases around it. Through this tutorial, we also want to show how you can easily integrate Prem with existing Open Source LLM orchestration tools like LangChain. The overall tutorial can be completed in four simple steps:

  1. Setting up all the required packages.
  2. We will introduce LangChain Prem AI and how to get started with it.
  3. Understand the MapReduce summarization technique and implement the summarization pipeline.
  4. Writing out a simple frontend using Streamlit.

Setting up the project

Let’s start by creating a virtual environment and installing dependencies.

Before getting started, make sure you have an account at Prem AI Platform and a valid project ID and an API Key. As a last requirement, you also need to have at least one repository-id.

python3 -m venv .venv
source .venv/bin/activate

Next, we need to install some dependencies. You can check all the dependencies in this requirements file and install them from here.

Now, create one folder named .streamlit. Inside that folder, create a file called secrets.toml. Inside secrets.toml add your PREMAI_API_KEY and PROJECT_ID as shown here:

premai_api_key = "xxxx-xxxx-xxxx"
premai_project_id = xxxx

So now our folder structure looks something like this:

├── .streamlit
│   ├── config.toml
│   └── secrets.toml.template
├── app.py

Inside app.py we import our required libraries as shown below:

import os 
import json 
import streamlit as st 
from urllib.parse import urlparse
from langchain_community.chat_models.premai import ChatPremAI

# langchain related imports

from langchain_community.document_loaders import WebBaseLoader
from langchain_community.chat_models.premai import ChatPremAI

from langchain.chains.llm import LLMChain
from langchain_core.prompts import PromptTemplate
from langchain_text_splitters import CharacterTextSplitter
from langchain.chains.combine_documents.stuff import StuffDocumentsChain
from langchain.chains import MapReduceDocumentsChain, ReduceDocumentsChain

# Set all the settings here 
premai_api_key = st.secrets.premai_api_key
premai_project_id = st.secrets.premai_project_id
os.environ["PREMAI_API_KEY"] = premai_api_key
prem_client = ChatPremAI(project_id=premai_project_id)

Please make sure you have a valid project_id. You can learn more about how to find your project_id here.

Understanding Map Reduce Summarization

We are going to implement the Map Reduce Summarization chain from LangChain. First, it is essential to understand how this chain operates on a high level. After this, we will implement it using Prem AI and LangChain.

The actual Map Reduce method

The original concept comes from the famous Map-Reduce programming model, which is part of the Hadoop framework from Apache. Primarily, Map-Reduce is used to simplify the processing of vast datasets across many machines in a cluster. The model consists of two main steps:

Map: This step converts an input dataset into a set of key-value pairs. Each input element is processed independently, producing intermediate key-value pairs.

Reduce: This step takes the intermediate key-value pairs produced by the Map step, groups them by key, and processes each group to make the final output.

You can deeply dive into the original workings here.

Map Reduce method in Summarization


The MapReduce summarization method is used for lengthy documents that exceed the token limit of the language model. So, this method tries to reduce the token length by summarizing it using the following steps:

Map step: The map step divides the documents into smaller chunks and then applies the summarization pipeline (LLMChain) to each chunk.

Reduce Step: Once all the chunks (parts of documents) are summarized, we combine all those summarized chunks and then call the LLM once again to summarize them, which gives us the final summarization.

Implement Map Reduce Summarization

Now let’s dive into coding our summarization chain.

Define Prompt Templates

We start by defining our prompt templates. We need to define two templates. The first one, map_template, instructs the LLM to take all the documents (in this case, the chunks of the documents extracted from the URL), identify the central theme, extract useful information, etc. In the second template, reduce_template, we ask it to summarize the document (a combined summary of the chunks) into a single summary that provides valuable insights to the reader.

map_template = """
The following is a set of documents
Based on this list of docs, please identify the main themes and extract 
informations from them which are highly valuable. So the main point should
be the theme/topic and subpoints should be the information extracts of the
documents which are very invaluable. 
Helpful Answer:

reduce_template = """
The following is set of summaries:
Take this and then make a very good summary with a very human like way so that
people can use this summary to decide whether to use that resource to finally read
or not. Do not give them the decision, give them enough insights or summary
that will help them. Include the following sub headings

1. What is this is about
2. Main key takeaways 
3. Things to note additionally
Helpful Answer:

Writing the summarization pipeline

Now, let’s write our summarization pipeline. This can be done with the following steps.

  1. First, load the langchain WebBaseLoader and the document from the given URL. We also load the CharacterTextSplitter to split the documents into chunks.
  2. Initialize the map_template and reduce_template from the above templates.
  3. Initialize the map_chain and reduce_chain from langchain LLMChain, which will use an LLM (in our case, a ChatPremAI client) and the corresponding template.
  4. We define a combined_document_chain which combines all the summarized chunks (stacking each of them one after the other) using the StuffDocumentsChain.
  5. Then, a reduce_document_chain takes those combined documents and runs the reduce (i.e. another summarization) on those stuffed documents.
  6. Finally, a map_reduce_chain combines all the above chains into a single chain.
  7. Run the chain and return the results
def summarize_url(url: str, llm: ChatPremAI) -> dict:
    # Step 1: Load the WebBasedLoader and CharacterTextSplitter to load and split documents 
    loader = WebBaseLoader(url)
    docs = loader.load()
    text_splitter = CharacterTextSplitter.from_tiktoken_encoder(
        chunk_size=1000, chunk_overlap=0
    split_docs = text_splitter.split_documents(docs)

    # Step 2: Get the prompts from prompt templates
    map_prompt = PromptTemplate.from_template(template=map_template)
    reduce_prompt = PromptTemplate.from_template(template=reduce_template)

    # Step 3: Make the map and reduce LLM chains
    map_chain = LLMChain(llm=llm, prompt=map_prompt)
    reduce_chain = LLMChain(llm=llm, prompt=reduce_prompt)

    # Step 4: Define `combined_document_chain` which combines all the summarized chunks 
    combined_documents_chain = StuffDocumentsChain(
        llm_chain=reduce_chain, document_variable_name="docs"
    # Step 5: Define `reduce_document_chain`
    reduce_documents_chain = ReduceDocumentsChain(

    # Step 6: Define the final map-reduce-chain which combines all of the above and run in one single pipeline
    map_reduce_chain = MapReduceDocumentsChain(

    # Step 7: Run the chain and get the results back
    result = map_reduce_chain.invoke(split_docs)
    return result

Phew, that was a lot. Congratulations if you made it this far. Now we will code our frontend interface using Streamlit.

Writing the frontend using Streamlit

We are only going to focus on the main summarizing component of the application. In the actual code, some additional styling is applied to enhance the appearance of the app.

Define the summarization component

Let’s write a simple helper function that uses the summarize_url function and creates an expandable card-like container which contains the summary of the URL.

# Additional function to check if the url is valid or not 
def is_valid_url(url: str):
        result = urlparse(url)
        return all([result.scheme, result.netloc])
    except ValueError:
        return False 

def summarize_component(urls: list[str], client: ChatPremAI):
    # Track which URLs got summarised and which are not
    passed, failed = [], []

    # Use a loader which will load untill all provided links are summarized. 
    with st.spinner("Please wait while we summarize the content"):
        for url in urls:
            if not is_valid_url(url):
                # Use the function 
                results = summarize_url(url, llm=client)
                if results:
                        "url": url,
                        "document": results["input_documents"],
                        "intermediate": results["intermediate_steps"],
                        "summary": results["output_text"]

                    # Create an expander which will contain the summarized content of the URL
                    with st.expander(label=f"URL: {url}"):
        if len(failed) > 0:
    return passed, failed 

Finally, let’s use the function with the main running code. We create a form with a text area where we dump all the URLs. Our code first checks the URLs. For all successful runs, it makes a card that, when expanded, contains the summary. It also returns a JSON object that indicates which links are not summarized or invalid (if any).

with st.form("Dump all URLs here"):
    input_urls = st.text_area(
        placeholder="Should be comma seperated valid urls"
    button = st.form_submit_button("Submit")
    if button:
        urls = [url.strip() for url in input_urls.split(",")]
        _, _ = summarize_component(urls=urls, client=prem_client)

We intentionally used _ because there is no use for this returned list of summaries. However, we can extend this by uploading all the summaries and documents (collected from passed) to Prem Repositories. We will show that in our following tutorial.

Congratulations you have created your first application using Prem AI. To run this application you just need to run the following command:

streamlit run app.py

You can check out more tutorials in our cookbook, as well as their full source code.