Blog
Support Chatbot Using Custom Knowledge Base with Langchain

Support Chatbot using Custom Knowledge Base with LangChain and Open LLM

Can you build a chatbot that can answer questions about your product or service? What if you could use your existing knowledge base to train the chatbot?

Join the AI BootCamp!

Ready to dive deep into the world of AI and Machine Learning? Join our BootCamp to transform your career with the latest skills and real-world project experience. LLMs, ML best practices, and more!

In this tutorial, we'll build a chatbot that can answer questions about Skyscanner's services . We'll use the company's FAQ section to extract the questions and answers. Then, we'll use the LangChain library to train a chatbot capable of answering questions in real-time on a single GPU (T4). We'll use free language model and embeddings.

In this part, we will be using Jupyter Notebook to run the code. If you prefer to follow along, you can find the notebook on GitHub: GitHub Repository (opens in a new tab)

Setup

Let's start by installing the required dependencies:

!pip install -Uqqq pip --progress-bar off
!pip install -qqq langchain==0.0.228 --progress-bar off
!pip install -qqq chromadb==0.3.26 --progress-bar off
!pip install -qqq sentence-transformers==2.2.2 --progress-bar off
!pip install -qqq auto-gptq==0.2.2 --progress-bar off
!pip install -qqq einops==0.6.1 --progress-bar off
!pip install -qqq unstructured==0.8.0 --progress-bar off
!pip install -qqq transformers==4.30.2 --progress-bar off
!pip install -qqq torch==2.0.1 --progress-bar off

Here's a list of the imports we'll be using:

from pathlib import Path
 
import torch
from auto_gptq import AutoGPTQForCausalLM
from langchain.chains import ConversationalRetrievalChain
from langchain.chains.question_answering import load_qa_chain
from langchain.document_loaders import DirectoryLoader
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.llms import HuggingFacePipeline
from langchain.memory import ConversationBufferMemory
from langchain.prompts import PromptTemplate
from langchain.text_splitter import CharacterTextSplitter
from langchain.vectorstores import Chroma
from transformers import AutoTokenizer, GenerationConfig, TextStreamer, pipeline

Data

The chatbot's custom knowledge is sourced from Skyscanner's FAQ1 section in the help center. We'll extract 12 questions and answers, creating separate text files for each. Let's create a helper function to write the files:

questions_dir = Path("skyscanner")
questions_dir.mkdir(exist_ok=True, parents=True)
 
 
def write_file(question, answer, file_path):
    text = f"""
Q: {question}
A: {answer}
""".strip()
    with Path(questions_dir / file_path).open("w") as text_file:
        text_file.write(text)

The format we're going to use for each question and answer is the following:

Q: Sample question?
A: Sample answer.

Time to write the files:

write_file(
    question="How do I search for flights on Skyscanner?",
    answer="""
Skyscanner helps you find the best options for flights on a specific date, or on any day in a given month or even year. For tips on how best to search, please head over to our search tips page.
 
If you're looking for inspiration for your next trip, why not try our everywhere feature. Or, if you want to hang out and ensure the best price, you can set up price alerts to let you know when the price changes.
""".strip(),
    file_path="question_1.txt",
)
 
write_file(
    question="What are mash-ups?",
    answer="""
These are routes where you fly with different airlines, because it`s cheaper than booking with just one. For example:
 
If you wanted to fly London to New York, we might find it`s cheaper to fly out with British Airways and back with Virgin Atlantic, rather than buy a round-trip ticket with one airline. This is called a "sum-of-one-way" mash-up. Just in case you're interested.
 
Another kind of mash-up is what we call a "self-transfer" or a "non-protected transfer". For example:
 
If you wanted to fly London to Sydney, we might find it`s cheaper to fly London to Dubai with Emirates, and then Dubai to Sydney with Qantas, rather than booking the whole route with one airline.
 
Pretty simple, right?
 
However, what`s really important to bear in mind is that mash-ups are NOT codeshares. A codeshare is when the airlines have an alliance. If anything goes wrong with the route - a delay, say, or a strike - those airlines will help you out at no extra cost. But mash-ups DO NOT involve an airline alliance. So if something goes wrong with a mash-up, it could cost you more money.
""".strip(),
    file_path="question_2.txt",
)
 
write_file(
    question="Why have I been blocked from accessing the Skyscanner website?",
    answer="""
Skyscanner's websites are scraped by bots many millions of times a day which has a detrimental effect on the service we're able to provide. To prevent this, we use a bot blocking solution which checks to ensure you're using the website in a normal manner.
 
Occasionally, this may mean that a genuine user may be wrongly flagged as a bot. This can be for a number of potential reasons, including, but not limited to:
 
You're using a VPN which we have had to block due to excessive bot traffic in the past
You're using our website at super speed which manages to beat our rate limits
You have a plug-in on your browser which could be interfering with how our website interacts with you as a user
You're using an automated browser
If you've been blocked during normal use, please send us your IP address (this website may help: http://www.whatismyip.com/), the website you're accessing (e.g. www.skyscanner.net) and the date/time this happened, via the Contact Us button below and we'll look into it as quickly as possible.
""".strip(),
    file_path="question_3.txt",
)
 
write_file(
    question="Where is my booking confirmation?",
    answer="""
You should get a booking confirmation by email from the company you bought your travel from. This can sometimes go into your spam/junk mail folder, so it's always worth checking there.
 
If you still can't find it, try getting in touch with the company you bought from to find out what's going on.
 
To find out who you need to contact, check the company name next to the charge on your bank account.
""".strip(),
    file_path="question_4.txt",
)
 
write_file(
    question="How do I change or cancel my booking?",
    answer="""
For all questions about changes, cancellations, and refunds - as well as all other questions about bookings - you'll need to contact the company you bought travel from. They'll have all the info about your booking and can advise you.
 
You'll find 1000s of travel agencies, airlines, hotels and car rental companies that you can buy from through our site and app. When you buy from one of these travel partners, they will take your payment (you'll see their name on your bank or credit card statement), contact you to confirm your booking, and provide any help or support you might need.
 
If you bought from one of these partners, you'll need to contact them as they have all the info about your booking. We unfortunately don't have any access to bookings you made with them.
""".strip(),
    file_path="question_5.txt",
)
 
write_file(
    question="I booked the wrong dates / times",
    answer="""
If you have found that you have booked the wrong dates or times, please contact the airline or travel agent that you booked your flight with as they will be able to help you change your flights to the intended dates or times.
 
The search box below can help you find the contact details for the travel provider you booked with.
 
You can search flexible or specific dates on Skyscanner to find your preferred flight, and when you select a flight on Skyscanner you are transferred to the website where you will make and pay for your booking. Once you are redirected to the airline or travel agent website, you might be required to select dates again, depending on the website. In all cases, you will be shown the flight details of your selection and you are required before confirming payment to state that you have checked all details and agreed to the terms and conditions. We strongly recommend that you always check this information carefully, as travel information can be subject to change.
""".strip(),
    file_path="question_6.txt",
)
 
write_file(
    question="I entered the wrong email address",
    answer="""
Please contact the airline or travel agent you booked with as Skyscanner does not have access to bookings made with airlines or travel agents.
 
If you can't remember who you booked with, you can check your credit card statement for a company name.
 
The search box below can help you find the contact details for the travel provider you booked with.
""".strip(),
    file_path="question_7.txt",
)
 
write_file(
    question="Luggage",
    answer="""
Depending on the flight provider, the rules, conditions and prices for luggage (including sports equipment) do vary.
It's always a good idea to check with the airline or travel agent directly (and you should be shown the options when you make your booking).
""".strip(),
    file_path="question_8.txt",
)
 
write_file(
    question="Changes, cancellation and refunds",
    answer="""
For changes, cancellations or refunds, we recommend that you contact the travel provider (airline or travel agent) agent that you completed your booking with.
 
As a travel search engine, Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through to your chosen airline or travel agent where you make your booking directly. We therefore don't have access or visibility to any of your booking information. Depending on the type of ticket you've booked, there may be different options for changes, cancellations and refunds, and the travel provider will be best placed to advise on these.
""".strip(),
    file_path="question_9.txt",
)
 
write_file(
    question="Why does the price sometimes change when I am redirected to a flight provider?",
    answer="""
Flight prices and availability change constantly, so we make sure the data is updated regularly to reflect this. When you redirect to a travel provider's site, the price is updated again so you can be sure that you will always see the best price available from the airline or travel agent at time of booking.
 
We make every effort to ensure the information you see on Skyscanner is accurate and up to date, but very occasionally there can be reasons why a price change has not updated accurately on the site. If you see a price difference between Skyscanner and a travel provider, please contact us with all the flight details (from, to, dates, departure times, airline and travel agent if applicable) and we will investigate further.
""".strip(),
    file_path="question_10.txt",
)
 
write_file(
    question="Why is Skyscanner free?",
    answer="""
Does Skyscanner charge commission?
 
Nope. Skyscanner is always free to search, and we never charge you any hidden fees.
 
Want to know how do we do it?
 
Well, we search through thousands of sites to find you the best deals for flights, hotels and car hire. That includes everything from fancy hotels to low cost airlines, so no matter what your budget is, we'll help you get there.
 
See a price you like? We'll connect you to that airline or travel company so you can book with them directly. And for this referral, the airline or travel company pays us a small fee.
 
And that's all there is to it!
""".strip(),
    file_path="question_11.txt",
)
 
write_file(
    question="Are my details safe?",
    answer="""
We take your privacy and safety online very seriously. We'll never sell, share or pass on your IP details, cookies, personal info and location data to others unless it's required by law, or it's necessary for one of the reasons set out in our Privacy Policy.
""".strip(),
    file_path="question_12.txt",
)

Model

I selected the model for our project based on the rankings provided by LMSYS Org2. My criteria included strong performance, availability on the HuggingFace Hub, and the ability to perform real-time inference on a single T4 GPU. Our chosen model is Nous-Hermes-13b3 by Nous Research, trained on GPT-4 synthetic data. To improve the inference speed, we'll utilize a quantized model using the AutoGPTQ library.

Luckily, the model has a quantized version available on the HuggingFace Hub, so we can use the AutoGPTQ library to load it. The quantized model4 we'll use is 4-bit, and it is provided by TheBloke. Let's load it:

DEVICE = "cuda:0" if torch.cuda.is_available() else "cpu"
 
model_name_or_path = "TheBloke/Nous-Hermes-13B-GPTQ"
model_basename = "nous-hermes-13b-GPTQ-4bit-128g.no-act.order"
 
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True)
 
model = AutoGPTQForCausalLM.from_quantized(
    model_name_or_path,
    model_basename=model_basename,
    use_safetensors=True,
    trust_remote_code=True,
    device=DEVICE,
)
 
generation_config = GenerationConfig.from_pretrained(model_name_or_path)

Note that we point to the exact file we want to load, and we also specify the model_basename parameter. We also load the tokenizer and the generation config.

To test the model, we need to format the prompt according to the standard format used by the model, which is similar to LLaMa models:

question = (
    "Which programming language is more suitable for a beginner: Python or JavaScript?"
)
prompt = f"""
### Instruction: {question}
### Response:
""".strip()

Let's run the prompt throught the tokenizer and the model:

%%time
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(DEVICE)
with torch.inference_mode():
    output = model.generate(inputs=input_ids, temperature=0.7, max_new_tokens=512)
CPU times: user 3.59 s, sys: 1.1 s, total: 4.69 s
Wall time: 12.4 s

This took about 12 seconds on a single T4 GPU. Let's see what the model generated:

print(tokenizer.decode(output[0]))
<s> ### Instruction: Which programming language is more suitable for a
beginner: Python or JavaScript?
### Response:Python is generally considered more suitable for beginners due to
its readability and simplicity compared to JavaScript.</s>

Note that the model generated a response that is relevant to the question. It also generated a <s> token at the beginning of the response, and a </s> token at the end. These tokens are used by the model to identify the beginning and the end of the response. We'll need to remove them before returning the response to the user.

Finally, let's have a look at the model's generation config to see what parameters it uses:

generation_config
GenerationConfig {
  "_from_model_config": true,
  "bos_token_id": 1,
  "eos_token_id": 2,
  "pad_token_id": 0,
  "transformers_version": "4.30.2"
}

Build a Pipeline

LangChain simplifies the usage of HuggingFace LLMs by providing the HuggingFacePipeline class. In this case, we will create a pipeline for text generation using our specific model and tokenizer. Additionally, we will apply a text streamer to stream the responses back to the user in real-time. It will also remove the <s> and </s> tokens, and the prompt from the response:

streamer = TextStreamer(
    tokenizer, skip_prompt=True, skip_special_tokens=True, use_multiprocessing=False
)
 
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_length=2048,
    temperature=0,
    top_p=0.95,
    repetition_penalty=1.15,
    generation_config=generation_config,
    streamer=streamer,
    batch_size=1,
)
 
llm = HuggingFacePipeline(pipeline=pipe)

Let's pass our prompt through the pipeline to test it out:

response = llm(prompt)
Python is generally considered to be more suitable for beginners due to its
readability and simplicity compared to JavaScript.

Great! The response is the same as the one we got when we ran the model directly, but it is cleaned up and ready to be returned to the user.

Embed Documents

To align with our preference for free and open-source models, we will utilize the E5-base-v25 embeddings, which were initially presented in the paper titled "Text Embeddings by Weakly-Supervised Contrastive Pre-training"6. The selection of this embedding model was based on its performance on the Massive Text Embedding Benchmark (MTEB) Leaderboard7. The model is available on the HuggingFace Hub, so we can load it using the HuggingFaceEmbeddings class:

embeddings = HuggingFaceEmbeddings(
    model_name="embaas/sentence-transformers-multilingual-e5-base",
    model_kwargs={"device": DEVICE},
)

We can utilize the DirectoryLoader in LangChain to load the text files as LangChain documents. This will allow us to conveniently work with the content of the files:

loader = DirectoryLoader("./skyscanner/", glob="**/*txt")
documents = loader.load()
len(documents)
12

Due to the context limit of our model (2048 tokens) and the length of the documents, we will need to split the documents into smaller chunks. We can use the CharacterTextSplitter to split the documents into chunks of 512 characters with no overlap:

text_splitter = CharacterTextSplitter(chunk_size=512, chunk_overlap=0)
texts = text_splitter.split_documents(documents)
texts[4]
Document(
    page_content='Q: Changes, cancellation and refunds A: For changes, cancellations or refunds, we recommend that you contact the travel provider (airline or travel agent) agent that you completed your booking with.',
    metadata={'source': 'skyscanner/question_9.txt'}
)

To create (and store) the embeddings we'll use the Chroma. We can create a database using the Chroma.from_documents method:

db = Chroma.from_documents(texts, embeddings)
db.similarity_search("flight search")
[
    Document(
        page_content="Q: How do I search for flights on Skyscanner? A: Skyscanner helps you find the best options for flights on a specific date, or on any day in a given month or even year. For tips on how best to search, please head over to our search tips page.\n\nIf you're looking for inspiration for your next trip, why not try our everywhere feature. Or, if you want to hang out and ensure the best price, you can set up price alerts to let you know when the price changes.",
        metadata={'source': 'skyscanner/question_1.txt'}
    ),
    Document(
        page_content="You're using a VPN which we have had to block due to excessive bot traffic in the past You're using our website at super speed which manages to beat our rate limits You have a plug-in on your browser which could be interfering with how our website interacts with you as a user You're using an automated browser If you've been blocked during normal use, please send us your IP address (this website may help: http://www.whatismyip.com/), the website you're accessing (e.g. www.skyscanner.net) and the date/time this happened, via the Contact Us button below and we'll look into it as quickly as possible.",
        metadata={'source': 'skyscanner/question_3.txt'}
    ),
    Document(
        page_content='Q: I booked the wrong dates / times A: If you have found that you have booked the wrong dates or times, please contact the airline or travel agent that you booked your flight with as they will be able to help you change your flights to the intended dates or times.\n\nThe search box below can help you find the contact details for the travel provider you booked with.',
        metadata={'source': 'skyscanner/question_6.txt'}
    ),
    Document(
        page_content='You can search flexible or specific dates on Skyscanner to find your preferred flight, and when you select a flight on Skyscanner you are transferred to the website where you will make and pay for your booking. Once you are redirected to the airline or travel agent website, you might be required to select dates again, depending on the website. In all cases, you will be shown the flight details of your selection and you are required before confirming payment to state that you have checked all details and agreed to the terms and conditions. We strongly recommend that you always check this information carefully, as travel information can be subject to change.',
        metadata={'source': 'skyscanner/question_6.txt'}
    )
]

With the smart agent (LLM) and the ability to retrieve relevant information (similarity search from the database), you have the necessary prerequisites to build your chatbot.

Conversational Chain

To bring everything together, we'll use a chain from LangChain. Let's begin by defining the prompt:

template = """
### Instruction: You're a travelling support agent that is talking to a customer.
 
Use only the chat history and the following information
{context}
to answer in a helpful manner to the question. If you don't know the answer -
say that you don't know. Keep your replies short, compassionate and informative.
 
{chat_history}
### Input: {question}
### Response:
""".strip()
 
prompt = PromptTemplate(
    input_variables=["context", "question", "chat_history"], template=template
)

Our prompt has 3 variables:

  • context - the relevant information from the retrieved documents (filled by the chain)
  • chat_history - the chat history thus far (filled by the memory)
  • question - the question asked by the user

As for the memory - we'll use the ConversationBufferMemory class from LangChain. It will store the chat history:

memory = ConversationBufferMemory(
    memory_key="chat_history",
    human_prefix="### Input",
    ai_prefix="### Response",
    output_key="answer",
    return_messages=True,
)

Note that we've also defined the human_prefix and ai_prefix - these will be used to format the chat history in the prompt.

Let's create the chain using our prompt:

chain = ConversationalRetrievalChain.from_llm(
    llm,
    chain_type="stuff",
    retriever=db.as_retriever(),
    memory=memory,
    combine_docs_chain_kwargs={"prompt": prompt},
    return_source_documents=True,
    verbose=True,
)

The chain combines the LLM, the retriever and the memory. It also uses the prompt to format the output. Finally, we've set return_source_documents=True to return the source documents from the database.

Let's try it out:

question = "How flight search works?"
answer = chain(question)
Chain output

Entering new chain...

Entering new chain... Prompt after formatting:

### Instruction: You're a travelling support agent that is talking to a customer. Use only the chat history and the following information

Q: How do I search for flights on Skyscanner? A: Skyscanner helps you find the best options for flights on a specific date, or on any day in a given month or even year. For tips on how best to search, please head over to our search tips page.

If you're looking for inspiration for your next trip, why not try our everywhere feature. Or, if you want to hang out and ensure the best price, you can set up price alerts to let you know when the price changes.

As a travel search engine, Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through to your chosen airline or travel agent where you make your booking directly. We therefore don't have access or visibility to any of your booking information. Depending on the type of ticket you've booked, there may be different options for changes, cancellations and refunds, and the travel provider will be best placed to advise on these.

You can search flexible or specific dates on Skyscanner to find your preferred flight, and when you select a flight on Skyscanner you are transferred to the website where you will make and pay for your booking. Once you are redirected to the airline or travel agent website, you might be required to select dates again, depending on the website. In all cases, you will be shown the flight details of your selection and you are required before confirming payment to state that you have checked all details and agreed to the terms and conditions. We strongly recommend that you always check this information carefully, as travel information can be subject to change.

Q: Why does the price sometimes change when I am redirected to a flight provider? A: Flight prices and availability change constantly, so we make sure the data is updated regularly to reflect this. When you redirect to a travel provider's site, the price is updated again so you can be sure that you will always see the best price available from the airline or travel agent at time of booking. to answer in a helpful manner to the question. If you don't know the answer - say that you don't know. Keep your replies short, compassionate and informative.

### Input: How flight search works?

### Response:

Skyscanner helps you find the best options for flights on a specific date, or on any day in a given month or even year. Our search algorithm scans hundreds of airlines and travel agents to bring you the most comprehensive range of flight options, including direct and connecting flights, stopovers, and layovers. For tips on how best to search, please visit our Search Tips page. As a travel search engine, Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through to your chosen airline or travel agent where you make your booking directly. We therefore don't have access or visibility to any of your booking information.

Finished chain.

Finished chain.

Skyscanner helps you find the best options for flights on a specific date, or
on any day in a given month or even year. Our search algorithm scans hundreds of
airlines and travel agents to bring you the most comprehensive range of flight options,
including direct and connecting flights, stopovers, and layovers.  For tips on how best
to search, please visit our Search Tips page. As a travel search engine,
Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through
to your chosen airline or travel agent where you make your booking directly. We
therefore don't have access or visibility to any of your booking information.

Great! Our agent has answered the question using the custom knowledge. Let's have a look at the answer structure:

answer.keys()
dict_keys(['question', 'chat_history', 'answer', 'source_documents'])

The answer contains the question, the chat history, the answer and the source documents. Let's try to ask another question:

question = "I bought flight tickets, but I can't find any confirmation. Where is it?"
response = chain(question)
Chain output

Entering new chain...

Prompt after formatting:

Given the following conversation and a follow up question, rephrase the follow up question to be a standalone question, in its original language.

Chat History:

Human: How flight search works?

Assistant: Skyscanner helps you find the best options for flights on a specific date, or on any day in a given month or even year. Our search algorithm scans hundreds of airlines and travel agents to bring you the most comprehensive range of flight options, including direct and connecting flights, stopovers, and layovers. For tips on how best to search, please visit our Search Tips page. As a travel search engine, Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through to your chosen airline or travel agent where you make your booking directly. We therefore don't have access or visibility to any of your booking information. Follow Up Input: I bought flight tickets, but I can't find any confirmation. Where is it? Standalone question: Can you help me locate my flight ticket confirmation?

Finished chain.

Entering new chain...

Entering new chain...

Prompt after formatting:

### Instruction: You're a travelling support agent that is talking to a customer. Use only the chat history and the following information Q: Where is my booking confirmation? A: You should get a booking confirmation by email from the company you bought your travel from. This can sometimes go into your spam/junk mail folder, so it's always worth checking there.

If you still can't find it, try getting in touch with the company you bought from to find out what's going on.

To find out who you need to contact, check the company name next to the charge on your bank account.

Q: I booked the wrong dates / times A: If you have found that you have booked the wrong dates or times, please contact the airline or travel agent that you booked your flight with as they will be able to help you change your flights to the intended dates or times.

The search box below can help you find the contact details for the travel provider you booked with.

Q: I entered the wrong email address A: Please contact the airline or travel agent you booked with as Skyscanner does not have access to bookings made with airlines or travel agents.

If you can't remember who you booked with, you can check your credit card statement for a company name.

The search box below can help you find the contact details for the travel provider you booked with.

As a travel search engine, Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through to your chosen airline or travel agent where you make your booking directly. We therefore don't have access or visibility to any of your booking information. Depending on the type of ticket you've booked, there may be different options for changes, cancellations and refunds, and the travel provider will be best placed to advise on these. to answer in a helpful manner to the question. If you don't know the answer - say that you don't know. Keep your replies short, compassionate and informative.

Human: How flight search works?

Assistant: Skyscanner helps you find the best options for flights on a specific date, or on any day in a given month or even year. Our search algorithm scans hundreds of airlines and travel agents to bring you the most comprehensive range of flight options, including direct and connecting flights, stopovers, and layovers. For tips on how best to search, please visit our Search Tips page. As a travel search engine, Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through to your chosen airline or travel agent where you make your booking directly. We therefore don't have access or visibility to any of your booking information. ### Input: Can you help me locate my flight ticket confirmation? ### Response: You should receive an email confirmation from the airline or travel agent you booked with. Sometimes this email might end up in your junk/spam folder. Try searching your inbox and spam folder first before reaching out to the airline or travel agent for assistance.

Finished chain.

Finished chain.

You should receive an email confirmation from the airline or travel agent you
booked with. Sometimes this email might end up in your junk/spam folder. Try
searching your inbox and spam folder first before reaching out to the airline or travel
agent for assistance.

The answer looks good! But, the chain is doing something quite strange. It rephrases the question (see the detailed output) and then answers it (haven't found a way to disable this in the current version of LangChain). This is done to reduce the amount of information that the agent needs to remember, but makes our agent slower and it's not always desirable. Next, we'll remove this rephrasing step.

QA Chain with Memory

We'll recreate our chain using the load_qa_chain:

memory = ConversationBufferMemory(
    memory_key="chat_history",
    human_prefix="### Input",
    ai_prefix="### Response",
    input_key="question",
    output_key="output_text",
    return_messages=False,
)
 
chain = load_qa_chain(
    llm, chain_type="stuff", prompt=prompt, memory=memory, verbose=True
)

This type of chain doesn't rephrase the question and doesn't have a retriever to search for the documents. We'll have to do it ourselves:

question = "How flight search works?"
docs = db.similarity_search(question)
answer = chain.run({"input_documents": docs, "question": question})
Skyscanner helps you find the best options for flights on a specific date, or
on any day in a given month or even year. Our search algorithm scans hundreds of
airlines and travel agents to bring you the most comprehensive range of flight options,
including direct and connecting flights, stopovers, and layovers.

For tips on how best to search, please visit our Search Tips page. As a travel search engine,
Skyscanner doesn't take your booking or payment ourselves. Instead, we pass you through
to your chosen airline or travel agent where you make your booking directly. We
therefore don't have access or visibility to any of your booking information.

Good, we got the same response! While it might be a bit more work, it's much more flexible and allows us to use the same chain as we want. Let's try another question:

question = "I entered wrong email address during my flight booking. What should I do?"
docs = db.similarity_search(question)
answer = chain.run({"input_documents": docs, "question": question})
Please contact the airline or travel agent you booked with as Skyscanner does
not have access to bookings made with airlines or travel agents. If you can't
remember who you booked with, you can check your credit card statement for a company
name. The search box below can help you find the contact details for the travel
provider you booked with.

No reprhasing and the answer is still good! Let's wrap everything into an easy to use package.

Support Chatbot

Let's create a class that makes it easy to use our chatbot:

DEFAULT_TEMPLATE = """
### Instruction: You're a travelling support agent that is talking to a customer.
 
Use only the chat history and the following information
{context}
to answer in a helpful manner to the question. If you don't know the answer -
say that you don't know. Keep your replies short, compassionate and informative.
 
{chat_history}
### Input: {question}
### Response:
""".strip()
 
 
class Chatbot:
    def __init__(
        self,
        text_pipeline: HuggingFacePipeline,
        embeddings: HuggingFaceEmbeddings,
        documents_dir: Path,
        prompt_template: str = DEFAULT_TEMPLATE,
        verbose: bool = False,
    ):
        prompt = PromptTemplate(
            input_variables=["context", "question", "chat_history"],
            template=prompt_template,
        )
        self.chain = self._create_chain(text_pipeline, prompt, verbose)
        self.db = self._embed_data(documents_dir, embeddings)
 
    def _create_chain(
        self,
        text_pipeline: HuggingFacePipeline,
        prompt: PromptTemplate,
        verbose: bool = False,
    ):
        memory = ConversationBufferMemory(
            memory_key="chat_history",
            human_prefix="### Input",
            ai_prefix="### Response",
            input_key="question",
            output_key="output_text",
            return_messages=False,
        )
 
        return load_qa_chain(
            text_pipeline,
            chain_type="stuff",
            prompt=prompt,
            memory=memory,
            verbose=verbose,
        )
 
    def _embed_data(
        self, documents_dir: Path, embeddings: HuggingFaceEmbeddings
    ) -> Chroma:
        loader = DirectoryLoader(documents_dir, glob="**/*txt")
        documents = loader.load()
        text_splitter = CharacterTextSplitter(chunk_size=512, chunk_overlap=0)
        texts = text_splitter.split_documents(documents)
        return Chroma.from_documents(texts, embeddings)
 
    def __call__(self, user_input: str) -> str:
        docs = self.db.similarity_search(user_input)
        return self.chain.run({"input_documents": docs, "question": user_input})

Our chatbot class takes a text pipeline, embeddings, and a directory with documents. The two main tasks are to create a chain and a database. We also override the __call__ method to make calling our chatbot as easy as calling a function.

Let's create an instance of our class:

chatbot = Chatbot(llm, embeddings, "./skyscanner/")

To create the chat, we'll use the while loop and ask the user for input. If the user enters bye or goodbye, we'll break the loop:

import warnings
 
warnings.filterwarnings("ignore", category=UserWarning)
 
while True:
    user_input = input("You: ")
    if user_input.lower() in ["bye", "goodbye"]:
        break
    answer = chatbot(user_input)
    print()

You: How flight search works?

Assistant: Skyscanner helps you find the best options for flights on a specific date, or on any day in a given month or even year. Our search algorithm scans hundreds of airlines and travel agents to bring you the most comprehensive range of flight options, including direct and connecting flights, stopovers, and layovers.

You: I bought flight tickets, but I can't find any confirmation. Where is it?

Assistant: It sounds like you purchased your flight tickets from another company or airline, rather than through Skyscanner. To locate your booking confirmation, please contact the company you purchased your ticket from directly. They will be able to provide you with more information regarding your booking and confirmation details.

You: I entered wrong email address during my flight booking. What should I do?

Assistant: Unfortunately, since Skyscanner isn't involved in the actual booking process, we are unable to assist with changing email addresses. However, you may reach out to the airline or travel agent you booked with to see if they can update your email address for you. Alternatively, you could try searching for your booking using the search bar provided on our website to see if you can locate your booking yourself.

You: What are the advantages of mash-ups?

Assistant: Mash-ups offer several benefits over traditional round-trip flights. Here are some key advantages:

  1. Cost savings: By combining two separate one-way flights, you can often save money compared to buying a single round-trip ticket.
  2. Flexibility: With a mash-up, you have more control over your travel plans. If you need to change your return date or time, you can simply adjust one leg of your trip without having to rebook everything else.
  3. More options: By flying with multiple airlines, you have access to a wider range of destinations and departure times. This means you can choose the flight that best fits your schedule and preferences.
3,000+ people already joined

Join the The State of AI Newsletter

Every week, receive a curated collection of cutting-edge AI developments, practical tutorials, and analysis, empowering you to stay ahead in the rapidly evolving field of AI.

I won't send you any spam, ever!

References

Footnotes

  1. Skyscanner help center (opens in a new tab) ↩

  2. Chatbot Arena Leaderboard by LMSYS Org (opens in a new tab) ↩

  3. Nous Hermes 13B (Original model) (opens in a new tab) ↩

  4. Nous-Hermes-13B GPTQ (opens in a new tab) ↩

  5. E5-base-v2 embeddings (opens in a new tab) ↩

  6. Text Embeddings by Weakly-Supervised Contrastive Pre-training (opens in a new tab) ↩

  7. Embeddings Leaderboard (opens in a new tab) ↩