human-in-the-loop

app.py

@@ -6,6 +6,7 @@ from knowledge import graph
 from pii import derisk
 from classify import judge
 from entity import resolve
+from human import email, feedback
 
 # Define the Google Analytics script
 head = """
@@ -478,4 +479,44 @@ For example, Comcast reduced repeat service calls by 17% after deploying entity
 
 The result? Less agent time lost, higher customer satisfaction, and data pipelines that actually speak human.
     """)
+  with gr.Tab("Human Feedback Content"):
+    gr.Markdown("""
+    Objective: Leveraging Human Feedback to Deliver Personalized Content that Proactively Solves Customer Pain Points
+    ================================================
+    """)
+
+    in_verbatim = gr.Textbox(label="Persona")
+    in_campaign = gr.Textbox(label="campaign")
+    out_product = gr.Textbox(label="Your Personalized Email Campaign Bot")
+
+    gr.Examples(
+      [
+        ["""My mortgage rate is 9%, I cannot afford it anymore, I need to refinance and I'm unemploy right now.""", "MORT"],
+        ["my credit card limit is too low, I need a card with bigger limit and low fee", "CARD"]
+      ],
+      [in_verbatim, in_campaign]
+    )
+    btn_recommend = gr.Button("Resolve")
+    btn_recommend.click(fn=email, inputs=[in_verbatim, in_campaign, out_product], outputs=out_product)
+
+    h_feedback = gr.Radio(['approved', 'rejected'], label="Human Feedback", info="Which campaign you want to approve?")
+    h_campaign = gr.Textbox(label="campaign")
+    btn_deliver = gr.Button("Deliver?")
+    btn_deliver.click(fn=feedback, inputs=[h_feedback, h_campaign, out_product], outputs=out_product)
+
+    gr.Markdown("""
+    Human Feedback for Personalized Content enables brands like Mr. Cooper to analyze customer preferences and pain points, then deliver tailored solutions. By embedding real-time feedback loops, they created personalized mortgage-refinancing videos showcasing individual home equity data and financial goals, resulting in 18% higher engagement and 12% lower churn.
+
+    #### Outcome:
+    
+    - Dynamic content adaptation based on behavioral data (e.g., Hilton Honors’ app reduced booking friction by 40% via predictive analytics)
+    
+    - Proactive problem-solving (e.g., Orangetheory Fitness used workout metrics to boost class attendance to 97%)
+    
+    - 52% faster ROI through AI-driven personalization scaling
+    
+    #### Ask: Ready to turn customer frustrations into loyalty drivers with content that feels personally crafted?"
+    
+    This approach aligns with best-in-class use cases where feedback-driven personalization drives measurable business growth
+    """)
 demo.launch(allowed_paths=["."])

human.py

@@ -0,0 +1,166 @@
+from typing import Annotated, Any, Literal
+
+from langchain_community.tools.tavily_search import TavilySearchResults
+from langchain_core.prompts import ChatPromptTemplate
+from langchain_core.tools import tool
+from langgraph.checkpoint.memory import MemorySaver
+from langgraph.graph import StateGraph, START, END
+from langgraph.graph.message import add_messages
+from langgraph.types import interrupt, Command
+from typing_extensions import TypedDict
+
+"""
+from langchain_anthropic import ChatAnthropic
+from langchain_ollama.llms import OllamaLLM
+from langchain_experimental.llms.ollama_functions import OllamaFunctions
+llm = OllamaFunctions(model="qwen2.5", format="json")  
+llm_with_tools = llm #.bind_tools(tools)
+"""
+
+from langchain_groq import ChatGroq
+
+llm = ChatGroq(
+    model="llama-3.1-8b-instant",
+    temperature=0,
+    max_tokens=None,
+    timeout=None,
+    max_retries=2,
+    # other params...
+)
+
+template = """Question: {question}
+Answer: Let's think step by step."""
+
+prompt = ChatPromptTemplate.from_template(template)
+
+# model = OllamaLLM(model="deepseek-r1")
+
+chain = prompt | llm
+
+
+# print(chain.invoke({"question": "Explain like I'm 5 for capacity planning?"}))
+
+@tool
+def human_assistance(query: str) -> str:
+  """Request assistance from a human."""
+  human_response = interrupt({"query": query})
+  return human_response["data"]
+
+
+tool = TavilySearchResults(max_results=2)
+tools = [tool, human_assistance]
+llm_with_tools=llm.bind_tools(tools)
+
+# llm = OllamaLLM(model="deepseek-r1") #ChatAnthropic(model="claude-3-5-sonnet-20240620")
+
+class State(TypedDict):
+  messages: Annotated[list, add_messages]
+  persona: str
+  email: str
+  release: Literal['approve', 'reject']
+
+
+graph_builder = StateGraph(State)
+
+
+def write_email(state: State):
+  prompt = f"""Write an promotional personalized email for this persona:
+    {state["persona"]}
+    """
+  email = llm_with_tools.invoke(prompt)
+  # Because we will be interrupting during tool execution,
+  # we disable parallel tool calling to avoid repeating any
+  # tool invocations when we resume.
+  # assert len(email.tool_calls) <= 1
+  return Command(update={"email": email.content})
+
+
+graph_builder.add_node("write_email", write_email)
+
+
+def delivery(state: State):
+  print(f"""Delivering: {state['email']}""")
+
+  return Command(update={"messages": ["Email delivered to customer"]})
+
+
+graph_builder.add_node("delivery", delivery)
+
+
+def human_approval(state: State) -> Command[Literal["delivery", END]]:
+  is_approved = interrupt(
+    "Approval for release the promotional email to customer? (type: approved or rejected):"
+  )
+
+  if is_approved == "approved":
+    return Command(goto="delivery", update={"release": "approved"})
+  else:
+    return Command(goto=END, update={"release": "rejected"})
+
+
+# Add the node to the graph in an appropriate location
+# and connect it to the relevant nodes.
+graph_builder.add_node("human_approval", human_approval)
+
+graph_builder.add_edge(START, "write_email")
+graph_builder.add_edge("write_email", "human_approval")
+
+graph_builder.add_edge("delivery", END)
+
+checkpointer = MemorySaver()
+graph = graph_builder.compile(checkpointer=checkpointer)
+
+
+def email(persona, campaign, history):
+  thread_config = {"configurable": {"thread_id": campaign}}
+  for event in graph.stream({"persona": persona}, config=thread_config):
+    for value in event.values():
+      return r"\nAssistant: ", value, r"\nValue: ", graph.get_state(thread_config).values
+
+
+def feedback(deliver, campaign, history):
+  thread_config = {"configurable": {"thread_id": campaign}}
+  for event in graph.stream(Command(resume=deliver), config=thread_config):
+    for value in event.values():
+      return r"\nAssistant: ", value, r"\nValue: ", graph.get_state(thread_config).values
+
+
+'''
+from IPython.display import Image, display
+
+try:
+  display(Image(graph.get_graph().draw_mermaid_png()))
+except Exception:
+  # This requires some extra dependencies and is optional
+  pass
+'''
+
+def campaign(user_input: Any, id: str):
+  thread_config = {"configurable": {"thread_id": id}}
+  for event in graph.stream(user_input, config=thread_config):
+    for value in event.values():
+      print("Assistant:", value, "Value: ", graph.get_state(thread_config).values)
+
+"""
+campaign({"persona": "My mortgage rate is 9%, I cannot afford it anymore, I need to refinance and I'm unemploy right now."}, "MOR")
+
+campaign({"persona": "my credit card limit is too low, I need a card with bigger limit and low fee"}, "CARD")
+
+campaign(Command(resume="approved"), "MOR")
+"""
+
+while False:
+  try:
+    user_input = input("User: ")
+    if user_input.lower() in ["quit", "exit", "q"]:
+      print("Goodbye!")
+      break
+    campaign(user_input, "MORT")
+    # stream_graph_updates(user_input)
+  except Exception as e:
+    # fallback if input() is not available
+    user_input = "What do you know about LangGraph?"
+    print("User: " + user_input)
+    campaign(user_input, "MORT")
+    # stream_graph_updates(user_input)
+    break