In Day 14, we built a content-based recommedation system using Amazon Reviews' 23, ntenceTransformer embeddings, and FAISS. Our system could retrieve products based on a user’s written preferences, but it still required the user to describe what they wanted. Today, we will take the next step by using user review and rating history to build a personalized recommendation workflow with LangGraph. Instead of starting with a manually written query, we will identify products a user previously liked, use them to build a user profile, and retrieve similar products that the user has not interacted with. By connecting these steps through LangGraph’s shared state and nodes, we will turn our Day 14 recommender into a workflow that can support more personalized recommendations.

Today’s goal: Given a user ID, generate Top-K product recommendations using the user’s interaction history, product embeddings, FAISS, and LangGraph.

In Day 15, we will replace that manually written preference with a profile derived from the user's previous reviews and ratings.

1. Workflow:

             START
               |
               v
      Input Amazon User ID
               |
               v
      Load User History
               |
               v
       Build User Profile
               |
               v
     Generate User Embedding
               |
               v
    Retrieve Candidates (FAISS)
               |
               v
     Filter Previously Seen
               |
               v
     Top-K Recommendations
               |
               v
              END
   

2. The LangGraph Structure

        Node	                  Responsibility
   load_user_history           Find the selected user's reviews 
   build_user_profile           Identify highly rated products
   create_user_embedding        Generate a vector representing
   retrieve_candidates           Search the FAISS product index
   filter_seen_items          Remove previously reviewed products
   

Step 1. Define the shared state

from typing import TypedDict


class RecommendationState(TypedDict):
    user_id: str
    user_history: list[dict]
    liked_products: list[dict]
    user_embedding: list[float]
    candidates: list[dict]
    recommendations: list[dict]

For Day 15, I would keep the experiment focused on generating a user embedding from review history and using LangGraph to organize the recommendation workflow. That gives readers a complete implementation while leaving conditional routing, cold-start handling, and eventually Qwen-based reranking for the following days.

To be continue......

Reference:
1.LangGraph Graph API
2.LangGraph StateGraph API
3.Amazon Reviews'23 Dataset
4.Amazon Reviews'23 Data Loading Guide
5.Sentence Transformers Documentation
6.FAISS: Metric Types and Distances