Construct an Clever FAQ Chatbot Utilizing Agentic RAG

AI brokers at the moment are part of enterprises large and small. From filling types at hospitals and checking authorized paperwork to analyzing video footage and dealing with buyer assist – we now have AI brokers for every kind of duties. Firms typically spend lots of of hundreds of {dollars} on hiring buyer assist workers who can perceive the wants of a buyer and resolve them based mostly on the corporate’s tips. Right this moment, having an clever chatbot to reply FAQs can effectively enhance customer support. On this article, we are going to discover ways to construct an FAQ chatbot that may resolve buyer queries in seconds, utilizing agentic RAG (Retrieval Augmented Technology), LangGraph and ChromaDB.

Temporary on Agentic RAG

RAG is a scorching subject these days. Everyone seems to be speaking about RAG and constructing purposes on prime of it. RAG helps LLMs to get entry to the real-time information, which makes LLMs extra correct than ever earlier than.  Nevertheless, conventional RAG programs are likely to fail with regards to selecting the most effective retrieval technique, altering the retrieval workflow, or offering multi-step reasoning. That is the place agentic RAG is available in.

Agentic RAG enhances conventional RAG by incorporating the capabilities of AI brokers into it. With this superpower, RAGs can dynamically change the workflow based mostly on the character of the question, do multi-step reasoning, and multi-step retrieval as effectively. We are able to even combine instruments into the agentic RAG system, and it will probably dynamically resolve which device to make use of when. General, it ends in improved accuracy and makes the system extra environment friendly and scalable.

Right here’s an instance of an agentic RAG workflow.

The picture above denotes the structure of an agentic RAG framework. It exhibits how AI brokers, when mixed with RAG, could make selections beneath sure circumstances. The picture clearly exhibits that if a conditional node is there, the agent will resolve which edge to decide on based mostly on the context offered.

Additionally Learn: 10 Enterprise Functions of LLM Brokers

Structure of the Clever FAQ Chatbot

Now we’re going to dive into the structure of the chatbot we’re going to construct. We’ll be exploring the way it works and what its essential parts are.

The next determine exhibits the general construction of our system. We will probably be implementing this utilizing LangGraph, which is an open-source AI brokers framework from LangChain.

The important thing parts of our system embrace:

1. LangGraph: A strong open-source AI agent framework that effectively creates complicated, multi-agent, cyclic graph-based brokers. These brokers can keep the states all through the workflow and might effectively deal with the complicated queries.
2. LLM: An environment friendly and highly effective Giant Language Mannequin that may comply with the directions of the consumer and reply accordingly with the most effective of its information. Right here we will probably be utilizing OpenAI’s o4-mini, which is a small reasoning mannequin that’s particularly designed for velocity, affordability, and power use.
3. Vector Database: A vector database is used to retailer, handle and retrieve vector embeddings that are normally the numeric illustration of knowledge. Right here we’re utilizing ChromaDB which is an open supply AI native vector database. It’s designed to empower the programs that depend upon similarity searches, semantic searches, and different duties involving vector information.

Additionally Learn: Construct a Buyer Help Voice Agent

Arms-on Implementation on Constructing the Clever FAQ Chatbot

Now, we will probably be implementing the end-to-end workflow of our chatbot based mostly on the structure that we now have mentioned above. We will probably be doing it step-by-step with detailed explanations, code, in addition to pattern outputs. So let’s start.

Step 1: Set up Dependencies

We’ll begin by putting in all of the required libraries into our Jupyter pocket book. This contains libraries equivalent to langchain, langgraph, langchain-openai, langchain-community, chromadb, openai, python-dotenv, pydantic, and pysqlite3.

!pip set up -q langchain langgraph langchain-openai langchain-community chromadb openai python-dotenv pydantic pysqlite3

Step 2: Import Required Libraries

Now we’re able to import all of the remaining libraries that we’ll want for this mission.

import os
import json
from typing import Checklist, TypedDict, Annotated, Dict
from dotenv import load_dotenv

# Langchain & LangGraph particular imports
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from pydantic import BaseModel, Subject
from langchain_core.messages import SystemMessage, HumanMessage, AIMessage
from langchain_core.paperwork import Doc
from langchain_community.vectorstores import Chroma
from langgraph.graph import StateGraph, END

Step 3: Set Up the OpenAI API Key

Enter your OpenAI key to set it as an atmosphere variable.

from getpass import getpass
OPENAI_API_KEY = getpass("OpenAI API Key:")
load_dotenv()
os.getenv("OPENAI_API_KEY")

Step 4: Obtain the Dataset

We’ve got made a pattern FAQ dataset in json format for various departments. We’ll must obtain it from the drive and unzip it.

!gdown 1j6pdIansfQzKOZSEUinnHd8w6GlkKE6w
!unzip -o /content material/blog_faq_files.zip

Output:

Step 5: Defining the Division Names for Mapping

Now, let’s outline the mapping of the departments in order that our agentic system can perceive which file belongs to which division.

# Outline Division Names (guarantee these match metadata used throughout ingestion)
DEPARTMENTS = [
   "Customer Support",
   "Product Information",
   "Loyalty Program / Rewards"
]
UNKNOWN_DEPARTMENT = "Unknown/Different"

FAQ_FILES = {
   "Buyer Help": "customer_support_faq.json",
   "Product Data": "product_information_faq.json",
   "Loyalty Program / Rewards": "loyalty_program_faq.json",
}

Step 6: Outline the Helper Capabilities

We’ll outline some helper features which will probably be chargeable for loading FAQs from the json information and in addition storing them in ChromaDB.

1. load_faqs(…): It’s a helper perform which hundreds the FAQ from the json information and retailer them in an inventory known as all_faqs.

def load_faqs(file_paths: Dict[str, str]) -> Dict[str, List[Dict[str, str]]]:
   """Masses QA pairs from JSON information for every division."""
   all_faqs = {}
   print("Loading FAQs...")
   for dept, file_path in file_paths.gadgets():
       attempt:
           with open(file_path, 'r', encoding='utf-8') as f:
               all_faqs[dept] = json.load(f)
               print(f"  - Loaded {len(all_faqs[dept])} FAQs for {dept}")
       besides FileNotFoundError:
           print(f"  - WARNING: FAQ file not discovered for {dept}: {file_path}. Skipping.")
       besides json.JSONDecodeError:
           print(f"  - ERROR: Couldn't decode JSON for {dept} from {file_path}. Skipping.")
   return all_faqs

2. setup_chroma_vector_store(…): This perform units up the ChromaDB to retailer the vector embeddings. For this, we are going to first outline the Chroma configuration i.e., the listing which is able to include the chroma database information. Then we are going to convert the FAQs to LangChain’s Paperwork. It is going to include metadata and web page content material which is the predefined format for an correct RAG. We are able to mix query and solutions for higher contextual retrieval or simply embed the reply. We’re holding the query as effectively division title within the metadata.

# ChromaDB Configuration
CHROMA_PERSIST_DIRECTORY = "./chroma_db_store"
CHROMA_COLLECTION_NAME = "Chatbot_faqs"

def setup_chroma_vector_store(
   all_faqs: Dict[str, List[Dict[str, str]]],
   persist_directory: str,
   collection_name: str,
   embedding_model: OpenAIEmbeddings,
) -> Chroma:
   """Creates or hundreds a Chroma vector retailer with FAQ information and metadata."""
   paperwork = []
   print("nPreparing paperwork for vector retailer...")
   for division, faqs in all_faqs.gadgets():
       for faq in faqs:
           # Mix Q&A for higher contextual embedding, or simply embed solutions
           # content material = f"Query: {faq['question']}nAnswer: {faq['answer']}"
           content material = faq['answer'] # Typically embedding simply the reply is efficient for FAQ retrieval
           doc = Doc(
               page_content=content material,
               metadata={
                   "division": division,
                   "query": faq['question'] # Maintain query in metadata for potential show
                   }
           )
           paperwork.append(doc)

   print(f"Complete paperwork ready: {len(paperwork)}")

   if not paperwork:
       increase ValueError("No paperwork discovered so as to add to the vector retailer. Test FAQ loading.")

   print(f"Initializing ChromaDB vector retailer (Persistence: {persist_directory})...")
   vector_store = Chroma(
       collection_name=collection_name,
       embedding_function=embedding_model,
       persist_directory=persist_directory,
   )
   attempt:
     vector_store = Chroma.from_documents(
             paperwork=paperwork,
             embedding=embedding_model,
             persist_directory=persist_directory,
             collection_name=collection_name
             )
     print(f"Created and populated ChromaDB with {len(paperwork)} paperwork.")
     vector_store.persist() # Guarantee persistence after creation
     print("Vector retailer persevered.")
   besides Exception as create_e:
           print(f"FATAL ERROR: Couldn't create Chroma vector retailer: {create_e}")
           increase create_e

   print("ChromaDB setup full.")
   return vector_store

Step 7: Outline the LangGraph Agent Elements

Let’s now outline our AI agent part which is the principle part of our work stream.

1. State definition: It’s a python class containing the present state of the agent whereas working. It accommodates variables equivalent to question, sentiment, division.

class AgentState(TypedDict):
   question: str
   sentiment: str
   division: str
   context: str # Retrieved context for RAG
   response: str # Closing response to the consumer
   error: str | None # To seize potential errors

2. Pydantic mannequin: We’ve got outlined a pydantic mannequin right here which is able to guarantee a structured LLM output. It accommodates a sentiment which can have three values, “optimistic”, “damaging” and “impartial” and a division title which will probably be predicted by the LLM.

class ClassificationResult(BaseModel):
   """Structured output for question classification."""
   sentiment: str = Subject(description="Sentiment of the question (optimistic, impartial, damaging)")
   division: str = Subject(description=f"Most related division from the listing: {DEPARTMENTS + [UNKNOWN_DEPARTMENT]}. Use '{UNKNOWN_DEPARTMENT}' if not sure or not relevant.")

3. Nodes: The next are the node features which is able to deal with every process one after the other.

• Classify_query_node: It classifies the incoming question into the sentiment in addition to the goal division title based mostly on the character of the question.
• retrieve_context_node:  It performs the RAG over the vector database and filter the outcomes on the idea of division title.
• generate_response_node: It generates the ultimate response based mostly on the question and retrieved context from the database.
• Human_escalation_node: If the sentiment is damaging or the goal division is unknown, it would escalate the question to the human consumer.
• route_query: It determines the subsequent step based mostly on the question and output of the classification node.

# 3. Nodes
def classify_query_node(state: AgentState) -> Dict[str, str]:
   """
   Classifies the consumer question for sentiment and goal division utilizing an LLM.
   """
   print("--- Classifying Question ---")
   question = state["query"]
   llm = ChatOpenAI(mannequin="o4-mini", api_key=OPENAI_API_KEY) # Use a dependable, cheaper mannequin

   # Put together immediate for classification
   prompt_template = ChatPromptTemplate.from_messages([
       SystemMessage(
           content=f"""You are an expert query classifier for ShopUNow, a retail company.
Analyze the user's query to determine its sentiment and the most relevant department.
The available departments are: {', '.join(DEPARTMENTS)}.
If the query doesn't clearly fit into one of these, or is ambiguous, classify the department as '{UNKNOWN_DEPARTMENT}'.
If the query expresses frustration, anger, dissatisfaction, or complains about a problem, classify sentiment as 'negative'.
If the query is asking a question, seeking information, or making a neutral statement, classify sentiment as 'neutral'.
If the query expresses satisfaction, praise, or positive feedback, classify sentiment as 'positive'.
Respond ONLY with the structured JSON output format."""
       ),
       HumanMessage(content=f"User Query: {query}")
   ])

   # LLM Chain with structured output
   classifier_chain = prompt_template | llm.with_structured_output(ClassificationResult)

   attempt:
       end result: ClassificationResult = classifier_chain.invoke({}) # Go empty dict as enter appears required now
       print(f"  Classification Consequence: Sentiment="{end result.sentiment}", Division="{end result.division}"")
       return {
           "sentiment": end result.sentiment.decrease(), # Normalize
           "division": end result.division
           }
   besides Exception as e:
       print(f"  Error throughout classification: {e}")
       return {
           "sentiment": "impartial", # Default on error
           "division": UNKNOWN_DEPARTMENT,
           "error": f"Classification failed: {e}"
           }

def retrieve_context_node(state: AgentState) -> Dict[str, str]:
   """
   Retrieves related context from the vector retailer based mostly on the question and division.
   """
   print("--- Retrieving Context ---")
   question = state["query"]
   division = state["department"]

   if not division or division == UNKNOWN_DEPARTMENT:
       print("  Skipping retrieval: Division unknown or not relevant.")
       return {"context": "", "error": "Can't retrieve context with no legitimate division."}

   # Initialize embedding mannequin and vector retailer entry
   embedding_model = OpenAIEmbeddings(api_key=OPENAI_API_KEY)
   vector_store = Chroma(
       collection_name=CHROMA_COLLECTION_NAME,
       embedding_function=embedding_model,
       persist_directory=CHROMA_PERSIST_DIRECTORY,
   )
   retriever = vector_store.as_retriever(
       search_type="similarity",
       search_kwargs={
           'ok': 3, # Retrieve prime 3 related docs
           'filter': {'division': division} # *** CRITICAL: Filter by division ***
           }
   )

   attempt:
       retrieved_docs = retriever.invoke(question)
       if retrieved_docs:
           context = "nn---nn".be part of([doc.page_content for doc in retrieved_docs])
           print(f"  Retrieved {len(retrieved_docs)} paperwork for division '{division}'.")
           # print(f"  Context Snippet: {context[:200]}...") # Elective: log snippet
           return {"context": context, "error": None}
       else:
           print("  No related paperwork present in vector retailer for this division.")
           return {"context": "", "error": "No related context discovered."}
   besides Exception as e:
       print(f"  Error throughout context retrieval: {e}")
       return {"context": "", "error": f"Retrieval failed: {e}"}

def generate_response_node(state: AgentState) -> Dict[str, str]:
   """
   Generates a response utilizing RAG based mostly on the question and retrieved context.
   """
   print("--- Producing Response (RAG) ---")
   question = state["query"]
   context = state["context"]
   llm = ChatOpenAI(mannequin="o4-mini", api_key=OPENAI_API_KEY) # Can use a extra succesful mannequin for technology

   if not context:
       print("  No context offered, producing generic response.")
       # Fallback if retrieval failed however routing determined RAG path anyway
       response_text = "I could not discover particular data associated to your question in our information base. Might you please rephrase or present extra particulars?"
       return {"response": response_text}

   # RAG Immediate
   prompt_template = ChatPromptTemplate.from_messages([
       SystemMessage(
           content=f"""You are a helpful AI Chatbot for ShopUNow. Answer the user's query based *only* on the provided context.
Be concise and directly address the query. If the context doesn't contain the answer, state that clearly.
Do not make up information.
Context:
---
{context}
---"""
       ),
       HumanMessage(content=f"User Query: {query}")
   ])

   RAG_chain = prompt_template | llm

   attempt:
       response = RAG_chain.invoke({})
       response_text = response.content material
       print(f"  Generated RAG Response: {response_text[:200]}...")
       return {"response": response_text}
   besides Exception as e:
       print(f"  Error throughout response technology: {e}")
       return {"response": "Sorry, I encountered an error whereas producing the response.", "error": f"Technology failed: {e}"}

def human_escalation_node(state: AgentState) -> Dict[str, str]:
   """
   Gives a message indicating the question will probably be escalated to a human.
   """
   print("--- Escalating to Human Help ---")
   purpose = ""
   if state.get("sentiment") == "damaging":
       purpose = "Because of the nature of your question,"
   elif state.get("division") == UNKNOWN_DEPARTMENT:
       purpose = "As your question requires particular consideration,"

   response_text = f"{purpose} I must escalate this to our human assist staff. They are going to overview your request and get again to you shortly. Thanks on your endurance."
   print(f"  Escalation Message: {response_text}")
   return {"response": response_text}

# 4. Conditional Routing Logic
def route_query(state: AgentState) -> str:
   """Determines the subsequent step based mostly on classification outcomes."""
   print("--- Routing Resolution ---")
   sentiment = state.get("sentiment", "impartial")
   division = state.get("division", UNKNOWN_DEPARTMENT)

   if sentiment == "damaging" or division == UNKNOWN_DEPARTMENT:
       print(f"  Routing to: human_escalation (Sentiment: {sentiment}, Division: {division})")
       return "human_escalation"
   else:
       print(f"  Routing to: retrieve_context (Sentiment: {sentiment}, Division: {division})")
       return "retrieve_context"

Step 8: Outline the Graph Perform

Let’s construct the perform for the graph and assign the nodes and edges to the graph.

# --- Graph Definition ---

def build_agent_graph(vector_store: Chroma) -> StateGraph:
   """Builds the LangGraph agent."""
   graph = StateGraph(AgentState)

   # Add nodes
   graph.add_node("classify_query", classify_query_node)
   graph.add_node("retrieve_context", retrieve_context_node)
   graph.add_node("generate_response", generate_response_node)
   graph.add_node("human_escalation", human_escalation_node)

   # Set entry level
   graph.set_entry_point("classify_query")

   # Add edges
   graph.add_conditional_edges(
       "classify_query", # Supply node
       route_query,      # Perform to find out the route
       {                 # Mapping: output of route_query -> vacation spot node
           "retrieve_context": "retrieve_context",
           "human_escalation": "human_escalation"
       }
   )
   graph.add_edge("retrieve_context", "generate_response")
   graph.add_edge("generate_response", END)
   graph.add_edge("human_escalation", END)

   # Compile the graph
   # reminiscence = SqliteSaver.from_conn_string(":reminiscence:") # Instance for in-memory persistence
   app = graph.compile() # checkpointer=reminiscence non-obligatory for stateful conversations
   print("nAgent graph compiled efficiently.")
   return app

Step 9: Provoke Agent Execution

Now, we will probably be initialising the agent and start executing the workflow.

1. Let’s begin by loading the FAQs.

# 1. Load FAQs
faqs_data = load_faqs(FAQ_FILES)

if not faqs_data:
   print("ERROR: No FAQ information loaded. Exiting.")
   exit()

Output:



2. Arrange the embedding fashions. Right here, we’ll be establishing OpenAI embedding fashions for a quicker retrieval.

# 2. Setup Vector Retailer
embedding_model = OpenAIEmbeddings(api_key=OPENAI_API_KEY)
vector_store = setup_chroma_vector_store(
   faqs_data,
   CHROMA_PERSIST_DIRECTORY,
   CHROMA_COLLECTION_NAME,
   embedding_model
)

Output:



Additionally Learn: Select the Proper Embedding for Your RAG Mannequin?

3. Now, construct the agent utilizing the predefined perform, visualizing the agent stream utilizing the mermaid diagram.

# 3. Construct the Agent Graph
agent_app = build_agent_graph(vector_store)

from IPython.show import show, Picture, Markdown

show(Picture(agent_app.get_graph().draw_mermaid_png()))

Output:



Step 10: Testing the Agent

We’ve got arrived on the final a part of our workflow. Thus far we now have constructed a number of nodes and features. Now’s the time to check our agent and see the output.

1. First let’s outline the check queries.

# Check the Agent
test_queries = [
   "How do I track my order?",
   "What is the return policy?",
   "Tell me about the 'Urban Explorer' jacket materials.",
]

2. Now let’s check the agent.

print("n--- Testing Agent ---")
for question in test_queries:
   print(f"nInput Question: {question}")
   # Outline the enter for the graph invocation
   inputs = {"question": question}
   # attempt:
   # Invoke the graph
   # The config argument is non-obligatory however helpful for stateful execution if wanted
   # config = {"configurable": {"thread_id": "user_123"}} # Instance config
   final_state = agent_app.invoke(inputs) #, config=config)

   print(f"Closing State Division: {final_state.get('division')}")
   print(f"Closing State Sentiment: {final_state.get('sentiment')}")
   print(f"Agent Response: {final_state.get('response')}")
   if final_state.get('error'):
         print(f"Error encountered: {final_state.get('error')}")

   # besides Exception as e:
   #     print(f"ERROR working agent graph for question '{question}': {e}")
   #     import traceback
   #     traceback.print_exc() # Print detailed traceback for debugging

print("n--- Agent Testing Full ---")

print(“n— Testing Agent —“)

Output:



We are able to see within the output that our agent is performing effectively. Firstly, it classifies the question after which routes the choice to the retrieval node or the human node. Then, the retrieval half comes it efficiently retrieves the context from the vector database. Within the final, producing the response as wanted. Therefore, we now have made our clever FAQ Chatbot.

You possibly can entry the Colab Pocket book with all of the code right here.

Conclusion

In case you have reached this far, it means you’ve got realized easy methods to construct an clever FAQ chatbot utilizing agentic RAG and LangGraph. Right here, we noticed that constructing an clever agent which might purpose and decide, shouldn’t be that tough. The agentic chatbot that we constructed is price environment friendly, quick, and is able to absolutely understanding the context of the questions or enter queries. The structure we’ve used right here is absolutely customizable which implies one can edit any node of the agent for his or her specific use case. With agentic RAG, LangGraph, and ChromaDB, making brokers has by no means been this straightforward. by no means really easy earlier than. I’m positive what we now have coated on this information has given you the foundational information to construct extra complicated system utilizing these instruments.




Harsh Mishra is an AI/ML Engineer who spends extra time speaking to Giant Language Fashions than precise people. Enthusiastic about GenAI, NLP, and making machines smarter (in order that they don’t substitute him simply but). When not optimizing fashions, he’s most likely optimizing his espresso consumption. 🚀☕