Using Langchain and OpenAI APIs in Python to Query Your Docs


Using the APIs from OpenAI and the langchain project, it is quite easy to implement a bot that is fed with your documentation and other product information. The answers of the bot are then specific to the trained knowledge domain. Background on the techniques used can be found here; this article describes the required Python code. In a nutshell:

To create a question-answering bot at a high level, we need to:

  • Prepare and upload a training dataset
  • Find the most similar document embeddings to the question embedding
  • Add the most relevant document sections to the query prompt
  • Answer the user's question based on additional context

The first step should be done only once (or once with every doc update), as calculating the embeddings has a price tag attached. The language of choice for such projects is Python, with a rich field of 3rd party tools. langchain is especially useful as it offers almost everything the DocBot heart desires.

Embedding documents

Langchain offers different loaders, e.g., for PDF, TXT, or sitemaps. The web pages are then automatically scraped and de-HTMLized. This is what the embedding code looks like for our use case at combit.

from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.llms import OpenAI
from langchain.chains import RetrievalQA
from langchain.chat_models import ChatOpenAI
from langchain.document_loaders.sitemap import SitemapLoader
from langchain.document_loaders import UnstructuredPDFLoader
from langchain.document_loaders import TextLoader

# This adds documents from a langchain loader to the database. The customized splitters serve to be able to break at sentence level if required.
def add_documents(loader, instance):
    documents = loader.load()
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100, separators= ["\n\n", "\n", ".", ";", ",", " ", ""])
    texts = text_splitter.split_documents(documents)

# Create embeddings instance
embeddings = OpenAIEmbeddings(openai_api_key="...")

# Create Chroma instance
instance = Chroma(embedding_function=embeddings, persist_directory="C:\\DocBot")

# add Knowledgebase Dump (CSV file)
loader = TextLoader("C:\\DocBot\\Input\\en-kb@forum-combit-net-2023-04-25.dcqresult.csv")
add_documents(loader, instance)

# add EN sitemap
loader = SitemapLoader(web_path='')
add_documents(loader, instance)

# add EN Blog sitemap, only use English blog posts
loader = SitemapLoader(web_path='', filter_urls=[""])
add_documents(loader, instance)

# add documentation PDFs
pdf_files = ["C:\\DocBot\\Input\\Ad-hoc Designer-Manual.pdf",

for file_name in pdf_files:
    loader = UnstructuredPDFLoader(file_name)
    add_documents(loader, instance)

instance = None

The costs are within tolerable limits - you'd rarely exceed a few Dollars, even for extensive documentation. Each query also costs another mini-amount, using two calls (embedding the question and then completing the answer, see below).

Q&A query with GPT 3.5 turbo

For this purpose, Flask can provide a WebAPI that works via langchain with the persistent vector database created in the first step. For the actual completion, gpt-3.5-turbo is used so the bot is personalized with a custom prompt template. This enables us to set the tone for the answers and make sure the model is not hallucinating answers. To make CORS calls work (for testing on a local web page), an access-control-allow-origin header is added here.

from flask import Flask, request,make_response
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.text_splitter import CharacterTextSplitter
from langchain.llms import OpenAI
from langchain.chains import RetrievalQA
from langchain.chat_models import ChatOpenAI
from langchain.prompts import PromptTemplate

app = Flask(__name__)

embeddings = OpenAIEmbeddings(openai_api_key="...")
instance = Chroma(persist_directory="C:\\DocBot", embedding_function=embeddings)

tech_template = """As a combit support bot, your goal is to provide accurate
and helpful technical information about List & Label, a powerful reporting tool used for
building various applications. You should answer user inquiries based on the
context provided and avoid making up answers. If you don't know the answer,
simply state that you don't know. Provide concrete examples like code snippets
or function prototypes wherever possible. Remember to provide relevant information
about List & Label's features, benefits, and API to assist the user in
understanding how to best use it for application development.


Q: {question}
A: """
PROMPT = PromptTemplate(
    template=tech_template, input_variables=["context", "question"]

qa = RetrievalQA.from_chain_type(llm=ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0,
                                                chain_type_kwargs={"prompt": PROMPT})    

def my_api():
    query = request.args.get('query')
    # process the input string here
    output_string =

    response = make_response(output_string, 200)
    response.mimetype = "text/plain"
    response.headers.add('Access-Control-Allow-Origin', '*')
    return response

if __name__ == '__main__':

Usage in a Web Page

The API can then be used, e.g., via fetch from JavaScript. Like this, here with localhost as a source if you're testing the bot locally. The HTML code was written by ChatGPT, of course :).

    <title>combit DocBot</title>
    <link rel="stylesheet" href="">
      #spinner {
        display: none;
      #result {
        margin-top: 20px;
    <div class="container my-5">
      <div class="row justify-content-center">
        <div class="col-md-6">
          <div class="card">
            <div class="card-header bg-primary text-white">
              <h4 class="mb-0">combit DocBot</h4>
            <div class="card-body">
                <div class="mb-3">
                  <label for="prompt" class="form-label">Enter a prompt:</label>
                  <input type="text" id="prompt" name="prompt" class="form-control">
                <div class="d-grid gap-2">
                  <button type="submit" class="btn btn-primary">
                    <span class="spinner-border spinner-border-sm ms-2" id="spinner" role="status" aria-hidden="true"></span>
            <div class="card-footer">
              <div class="result" id="result"></div>
    <script src=""></script>    
      const form = document.querySelector('form');
      const resultDiv = document.querySelector('#result'); = 'none';

      form.addEventListener('submit', async (event) => {
        event.preventDefault(); = 'inline-block';
        resultDiv.textContent = '';
        const prompt = document.querySelector('#prompt').value;

          .then((response) => response.text())
          .then((text) => {
   = 'none'; resultDiv.innerHTML = marked.parse(text);});

The result might then look like this.

Wrap up

Using OpenAI, Python, and langchain, it's easy to write a bot that can be customized to answer questions based on domain knowledge. The code presented here is a great starting point for your custom encounters and might save a few hours of searching the web for best practices.

Similar Articles