In Day 14, we built a simple content-based recommendation system using Amazon Reviews’23, SentenceTransformer embeddings, and FAISS. That system could retrieve products that were semantically similar to a user’s preferences, but the overall workflow was still relatively simple. Today, we will take the next step by introducing LangGraph to organize the recommendation process into multiple stages. Inspired by the amine-akrout/llm-based-recommender project, which separates recommendation into retrieval, ranking, and final response generation, we will build our own lightweight workflow using LangGraph, FAISS, and Amazon Reviews’23. Instead of treating recommendation as a single vector-search operation, we will begin to model it as a sequence of connected decisions that can later be extended with reranking, LLM reasoning, and agent-like behavior.

    User Preference
          ↓
    Check Request
          ↓
    Retrieve Candidates
          ↓
    Rank Candidates
          ↓
    Filter Results
          ↓
    Generate Recommendations

Reference:

1. amine-akrout/llm-based-recommender