Building Predictive Intelligence with Multi-Agent RAG in FMCG Retail

For the past few years, the industry standard for Retrieval-Augmented Generation (RAG) has been purely reactive: a user asks a question, the system retrieves documents, and the LLM generates an answer. While valuable for internal search, pure QA RAG is fundamentally limited. It waits for a human to realize something is wrong and ask about it. In enterprise retail and Fast-Moving Consumer Goods (FMCG), waiting for a human to ask, "Why are we stocking out of Product X?" is too late. The margin has already been lost. To drive real ROI, we must transition from Reactive QA RAG to Proactive Predictive Intelligence RAG. In a predictive workflow, the system doesn't wait for a prompt. It monitors structured data streams, detects anomalies, proactively retrieves unstructured context to understand the causes, forecasts the future trajectory, and triggers automated business actions.

Here is an end-to-end guide on how we built a multi-agent predictive RAG system using LangGraph for an FMCG retail client, complete with state management and persistent memory.

The Use Case: Proactive Stockout & Promotional Pivot

The Scenario: An FMCG company launches a new line of organic, cold-pressed beverages. These products have a short shelf life and are highly sensitive to external factors (weather, social media trends, local events).

The Problem: Traditional time-series forecasting fails to predict sudden demand spikes caused by unstructured events (e.g., a viral TikTok review, an unexpected heatwave, or a competitor's stockout). By the time the supply chain reacts, the shelves are empty.

The Predictive RAG Solution: Instead of a QA bot, we built an autonomous multi-agent workflow that continuously monitors sales velocity. When it detects an anomaly (e.g., a 30% spike in sales velocity for a specific SKU in the Texas region), it proactively wakes up.

It uses RAG to search unstructured data (local news, weather forecasts, social sentiment, supplier notes) to find the causal context.
It uses Predictive Analytics to forecast the exact stockout date based on the combined structured and unstructured data.
It uses Strategic Reasoning to recommend an immediate supply chain pivot or promotional adjustment.

The Multi-Agent Workflow

To achieve this, we utilized LangGraph to orchestrate three specialized agents. Because we are avoiding static architectural diagrams, let's walk through the dynamic flow of state and memory:

The Sentinel (Trigger & RAG Agent): Monitors structured sales data. When an anomaly is detected, it queries the vector database for unstructured causal context (weather, news, social trends) relevant to the specific geographic region and SKU.
The Oracle (Predictive Analyst): Takes the structured sales data and the unstructured RAG context. It calculates the revised demand forecast and predicts the exact time of stockout.
The Strategist (Action Agent): Evaluates the prediction against current inventory, warehouse capacity, and historical memory (past actions taken in similar scenarios). It formulates a concrete action plan (e.g., "Reroute 500 units from Oklahoma to Texas, pause digital ad spend in Oklahoma").

State and Memory: The entire workflow shares a unified State. Furthermore, we use LangGraph’s persistent MemorySaver. This allows the Strategist to remember that "Last time a heatwave hit Texas in July, we rerouted inventory, but it caused a stockout in Oklahoma. This time, we need to pull from the central Dallas hub instead."

Code Implementation

Below is the complete Python implementation using LangGraph, LangChain, and Pydantic for state management.

1. Setup and State Definition

First, we define the shared state. In LangGraph, the state is the single source of truth that passes between agents.

from typing import TypedDict, Annotated, List, Dict, Any
from langgraph.graph import StateGraph, END
from langgraph.checkpoint.memory import MemorySaver
from langchain_core.messages import HumanMessage, AIMessage
import datetime

# Define the shared state for the predictive workflow
class PredictiveState(TypedDict):
    thread_id: str
    sku_id: str
    region: str
    # Structured data inputs
    current_inventory: int
    sales_velocity_anomaly: float # e.g., 1.3 means 30% above normal
    # RAG outputs
    retrieved_unstructured_context: List[str] 
    # Predictive outputs
    predicted_stockout_date: str
    # Action outputs
    action_plan: str
    # Memory context
    historical_lessons: str
    messages: List[Any]

2. Mock Tools for RAG and Data Retrieval

In a real enterprise environment, these would connect to Pinecone/Weaviate (for RAG) and Snowflake/Databricks (for structured data).

def search_causal_context(sku: str, region: str) -> List[str]:
    """Mock RAG retrieval for unstructured causal data."""
    # In production: Vector DB search using semantic queries like 
    # "weather events, social trends, or local news in {region} affecting {sku}"
    return [
        f"NEWS: Unprecedented heatwave warning issued for {region} for the next 5 days.",
        f"SOCIAL: Viral TikTok trend in {region} promoting {sku} as the ultimate hydration hack (2M views).",
        f"COMPETITOR: Major rival brand recalled their citrus beverage in {region}."
    ]

def get_historical_memory(region: str, event_type: str) -> str:
    """Mock retrieval of past actions from long-term memory."""
    return f"MEMORY: During the last heatwave in {region}, we rerouted stock from neighboring states, which caused secondary stockouts. Lesson: Pre-position inventory from the central regional hub instead."

3. Defining the Agent Nodes

Now we define the nodes. Each node is a function that reads from the PredictiveState, performs its specific task, and returns an updated state.

def sentinel_rag_agent(state: PredictiveState) -> Dict[str, Any]:
    """Agent 1: Detects anomaly and retrieves causal context via RAG."""
    sku = state["sku_id"]
    region = state["region"]
    
    print(f"[Sentinel] Anomaly detected for {sku} in {region}. Fetching causal context...")
    context = search_causal_context(sku, region)
    
    return {
        "retrieved_unstructured_context": context,
        "messages": state["messages"] + [AIMessage(content="Context retrieved. Passing to Oracle.")]
    }

def oracle_predictive_agent(state: PredictiveState) -> Dict[str, Any]:
    """Agent 2: Combines structured velocity with unstructured RAG context to forecast."""
    context = "\n".join(state["retrieved_unstructured_context"])
    velocity = state["sales_velocity_anomaly"]
    inventory = state["current_inventory"]
    
    print(f"[Oracle] Analyzing velocity ({velocity}x) and context to predict stockout...")
    
    # Simulated LLM call for predictive math
    # Prompt would be: "Given normal velocity is 100 units/day, current velocity is {velocity*100}, 
    # and context is {context}, when will {inventory} units run out?"
    predicted_days = inventory / (100 * velocity)
    stockout_date = (datetime.date.today() + datetime.timedelta(days=int(predicted_days))).isoformat()
    
    return {
        "predicted_stockout_date": stockout_date,
        "messages": state["messages"] + [AIMessage(content=f"Stockout predicted on {stockout_date}. Passing to Strategist.")]
    }

def strategist_action_agent(state: PredictiveState) -> Dict[str, Any]:
    """Agent 3: Formulates action plan using predictions and historical memory."""
    stockout_date = state["predicted_stockout_date"]
    region = state["region"]
    
    # Retrieve historical lessons using the memory saver context
    historical_lessons = get_historical_memory(region, "heatwave")
    
    print(f"[Strategist] Formulating action plan considering past lessons...")
    
    # Simulated LLM call for strategy
    # Prompt: "Stockout is imminent on {stockout_date}. Based on {historical_lessons}, 
    # what is the optimal supply chain and marketing action?"
    action = (
        f"1. IMMEDIATE: Reroute 2,000 units from the Central Dallas Hub to {region} "
        f"(Avoid pulling from neighboring states to prevent secondary stockouts). "
        f"2. MARKETING: Increase digital ad spend for {state['sku_id']} in {region} by 15% "
        f"to capitalize on the viral trend, but cap it to match the rerouted inventory."
    )
    
    return {
        "action_plan": action,
        "historical_lessons": historical_lessons,
        "messages": state["messages"] + [AIMessage(content="Action plan finalized. Workflow complete.")]
    }

4. Graph Construction and Execution

Finally, we wire the nodes together into a LangGraph StateGraph, add conditional routing (if necessary, though here it's a linear predictive flow), and attach the MemorySaver.

def build_predictive_graph():
    # Initialize the graph with the defined state
    workflow = StateGraph(PredictiveState)

    # Add nodes
    workflow.add_node("sentinel_rag", sentinel_rag_agent)
    workflow.add_node("oracle_predict", oracle_predictive_agent)
    workflow.add_node("strategist_action", strategist_action_agent)

    # Define the entry point
    workflow.set_entry_point("sentinel_rag")

    # Define edges (the flow of the predictive workflow)
    workflow.add_edge("sentinel_rag", "oracle_predict")
    workflow.add_edge("oracle_predict", "strategist_action")
    workflow.add_edge("strategist_action", END)

    # Initialize Persistent Memory
    # In production, use SqliteSaver or PostgresSaver for enterprise durability
    memory = MemorySaver()

    # Compile the graph
    app = workflow.compile(checkpointer=memory)
    return app

# --- Execution ---
if __name__ == "__main__":
    app = build_predictive_graph()
    
    # Configuration for the thread (Memory isolation)
    config = {"configurable": {"thread_id": "fmcg-tx-organic-bev-001"}}
    
    # Initial trigger state (simulating an automated alert from the data warehouse)
    initial_state = {
        "thread_id": "fmcg-tx-organic-bev-001",
        "sku_id": "ORG-BEV-CITRUS-500ML",
        "region": "Texas",
        "current_inventory": 4500,
        "sales_velocity_anomaly": 2.5, # 2.5x normal sales velocity
        "retrieved_unstructured_context": [],
        "predicted_stockout_date": "",
        "action_plan": "",
        "historical_lessons": "",
        "messages": [HumanMessage(content="Anomaly detected. Initiating predictive workflow.")]
    }

    print("--- Starting Predictive Intelligence Workflow ---")
    
    # Invoke the graph
    final_state = app.invoke(initial_state, config)
    
    print("\n--- Workflow Complete ---")
    print(f"Predicted Stockout Date: {final_state['predicted_stockout_date']}")
    print(f"Action Plan:\n{final_state['action_plan']}")

Why This Shift Matters for Enterprise FMCG

By moving from a pure QA RAG workflow to a predictive multi-agent workflow, we fundamentally change the value proposition of AI in the supply chain:

Causal Inference over Correlation: Pure QA RAG just finds documents. Predictive RAG uses RAG to find causes. It doesn't just know sales are up; it knows sales are up because of a viral trend and a heatwave, allowing for highly accurate forecasting.
Institutional Memory: By utilizing LangGraph's persistent memory, the Strategist agent doesn't make the same mistake twice. It remembers past supply chain pivots and adjusts its recommendations, effectively capturing the tribal knowledge of veteran supply chain managers.
Zero-Latency Response: The system doesn't wait for a category manager to notice a dashboard anomaly and ask a chatbot about it. The AI detects the anomaly, retrieves the context, predicts the outcome, and drafts the solution autonomously, requiring only human-in-the-loop approval for the final execution.

In the high-margin, low-tolerance world of FMCG retail, the difference between reacting to a stockout and predicting it days in advance is the difference between lost revenue and optimized profitability. Predictive RAG is the bridge to that future.