Hitchhiker’s Guide to RAG with ChatGPT API and LangChain

generate tons of phrases and responses based mostly on normal data, however what occurs after we want solutions requiring correct and particular data? Solely generative fashions continuously wrestle to offer solutions on area particular questions for a bunch of causes; possibly the information they have been educated on are actually outdated, possibly what we’re asking for is actually particular and specialised, possibly we wish responses that keep in mind private or company information that simply aren’t public… 🤷‍♀️ the record goes on.

So, how can we leverage generative AI whereas preserving our responses correct, related, and down-to-earth? A great reply to this query is the Retrieval-Augmented Generation (RAG) framework. RAG is a framework that consists of two key parts: retrieval and technology (duh!). Not like solely generative fashions which might be pre-trained on particular information, RAG incorporates an additional step of retrieval that enables us to push extra info into the mannequin from an exterior supply, corresponding to a database or a doc. To place it otherwise, a RAG pipeline permits for offering coherent and pure responses (offered by the technology step), that are additionally factually correct and grounded in a data base of our selection (offered by the retrieval step).

On this means, RAG might be an especially precious software for purposes the place extremely specialised information is required, as as an example buyer assist, authorized recommendation, or technical documentation. One typical instance of a RAG software is buyer assist chatbots, answering buyer points based mostly on an organization’s database of assist paperwork and FAQs. One other instance can be advanced software program or technical merchandise with in depth troubleshooting guides. Another instance can be authorized recommendation — a RAG mannequin would entry and retrieve customized information from legislation libraries, earlier circumstances, or agency pointers. The examples are actually countless; nevertheless, in all these circumstances, the entry to exterior, particular, and related to the context information allows the mannequin to supply extra exact and correct responses.

So, on this submit, I stroll you thru constructing a easy RAG pipeline in Python, using ChatGPT API, LangChain, and FAISS.

What about RAG?

From a extra technical perspective, RAG is a method used to reinforce an LLM’s responses by injecting it with extra, domain-specific info. In essence, RAG permits for a mannequin to additionally keep in mind extra exterior info — like a recipe ebook, a technical handbook, or an organization’s inside data base — whereas forming its responses.

This is essential as a result of it permits us to get rid of a bunch of issues inherent to LLMs, as as an example:

• Hallucinations — making issues up
• Outdated info — if the mannequin wasn’t educated on current information
• Transparency — not understanding the place responses are coming from

To make this work, the exterior paperwork are first processed into vector embeddings and saved in a vector database. Then, after we submit a immediate to the LLM, any related information is retrieved from the vector database and handed to the LLM together with our immediate. Because of this, the response of the LLM is fashioned by contemplating each our immediate and any related info current within the vector database within the background. Such a vector database might be hosted domestically or within the cloud, utilizing a service like Pinecone or Weaviate.

What about ChatGPT API, LangChain, and FAISS?

The primary part for constructing a RAG pipeline is the LLM mannequin that can generate the responses. This may be any LLM, like Gemini or Claude, however on this submit, I will likely be utilizing OpenAI’s ChatGPT fashions through their API platform. In an effort to use their API, we have to sign up and acquire an API key. We additionally want to verify the respective Python libraries are put in.

pip set up openai

The opposite main part of constructing a RAG is processing exterior information — producing embeddings from paperwork and storing them in a vector database. The most well-liked framework for performing such a activity is LangChain. Specifically, LangChain permits:

• Load and extract textual content from varied doc varieties (PDFs, DOCX, TXT, and many others.)
• Break up the textual content into chunks appropriate for producing the embeddings
• Generate vector embeddings (on this submit, with the help of OpenAI’s API)
• Retailer and search embeddings through vector databases like FAISS, Chroma, and Pinecone

We are able to simply set up the required LangChain libraries by:

pip set up langchain langchain-community langchain-openai

On this submit, I’ll be utilizing LangChain along with FAISS, an area vector database developed by Fb AI Analysis. FAISS is a really light-weight package deal, and is thus applicable for constructing a easy/small RAG pipeline. It may be simply put in with:

pip set up faiss-cpu

Placing every thing collectively

So, in abstract, I’ll use:

• ChatGPT fashions through OpenAI’s API because the LLM
• LangChain, together with OpenAI’s API, to load the exterior recordsdata, course of them, and generate the vector embeddings
• FAISS to generate an area vector database

The file that I will likely be feeding into the RAG pipeline for this submit is a textual content file with some info about me. This textual content file is positioned within the folder ‘RAG recordsdata’.

Now we’re all arrange, and we will begin by specifying our API key and initializing our mannequin:

from openai import OpenAI # Chat_GPT API key api_key = "your key" 

# initialize LLM 
llm = ChatOpenAI(openai_api_key=api_key, mannequin="gpt-4o-mini", temperature=0.3)

Then we will load the recordsdata we need to use for the RAG, generate the embeddings, and retailer them as a vector database as follows:

# loading paperwork for use for RAG 
text_folder = "rag_files"  

all_documents = []
for filename in os.listdir(text_folder):
    if filename.decrease().endswith(".txt"):
        file_path = os.path.be part of(text_folder, filename)
        loader = TextLoader(file_path)
        all_documents.prolong(loader.load())

# generate embeddings
embeddings = OpenAIEmbeddings(openai_api_key=api_key)

# create vector database w FAISS 
vector_store = FAISS.from_documents(paperwork, embeddings)
retriever = vector_store.as_retriever()

Lastly, we will wrap every thing in a easy executable Python file:

def principal():
    print("Welcome to the RAG Assistant. Kind 'exit' to stop.n")
    
    whereas True:
        user_input = enter("You: ").strip()
        if user_input.decrease() == "exit":
            print("Exiting…")
            break

        # get related paperwork
        relevant_docs = retriever.get_relevant_documents(user_input)
        retrieved_context = "nn".be part of([doc.page_content for doc in relevant_docs])

        # system immediate
        system_prompt = (
            "You're a useful assistant. "
            "Use ONLY the next data base context to reply the consumer. "
            "If the reply just isn't within the context, say you do not know.nn"
            f"Context:n{retrieved_context}"
        )

        # messages for LLM 
        messages = [
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_input}
        ]

        # generate response
        response = llm.invoke(messages)
        assistant_message = response.content material.strip()
        print(f"nAssistant: {assistant_message}n")

if __name__ == "__main__":
    principal()

Discover how the system immediate is outlined. Basically, a system immediate is an instruction given to the LLM that units the habits, tone, or constraints of the assistant earlier than the consumer interacts. For instance, we may set the system immediate to make the LLM present responses like speaking to a 4-year-old or a rocket scientist — right here we ask to offer responses solely based mostly on the exterior information we offered, the ‘Maria info’

So, let’s see what we’ve cooked! 🍳

Firstly, I ask a query that’s irrelevant to the offered exterior datasource, to ensure that the mannequin solely makes use of the offered datasource when forming the responses and never normal data.

… after which I requested some questions particularly from the file I offered…

✨✨✨✨

On my thoughts

Apparently, it is a very simplistic instance of a RAG setup — there’s rather more to think about when implementing it in an actual enterprise surroundings, corresponding to safety issues round how information is dealt with, or efficiency points when coping with a bigger, extra lifelike data corpus and elevated token utilization. Nonetheless, I consider OpenAI’s API is actually spectacular and gives immense, untapped potential for constructing customized, context-specific AI purposes.

Liked this submit? Let’s be mates! Be part of me on

📰Substack 💌 Medium 💼LinkedIn ☕Buy me a coffee!