How swagath18/personalized-recommender works
Project type and purpose
• This is a Python-based, full-stack recommender prototype combining Retrieval-Augmented Generation (RAG) with GPT-4, FAISS vector search, hybrid reranking, a Gradio UI, and feedback sentiment analysis (VADER).
• Purpose: accept a user query + context, retrieve semantically relevant products, ask GPT-4 to craft recommendations, display results, collect feedback, analyze sentiment, and log everything.
File structure
• Top-level scripts:
• llmpoweredrag.py – main end-to-end application (UI + retrieval + RAG)
• llmragsenti.py – likely includes feedback-sentiment components
• Data/config:
• amazon_electronics_products.csv – product catalog
• .env – API keys
• recommendation_log.txt, feedback_log.txt – audit logs
• Supporting: .gradio folder holds UI session data
README for additional insights
• Confirms use of: SentenceTransformers → FAISS → GPT-4 → hybrid reranking → Gradio UI → VADER
• Describes multi-tab UI (Recommend + Feedback)
• Shows expected query/context inputs and output examples
• Mentions logging for auditability
Main components & their interactions
a. User Interface (Gradio)
• Input “Query” + optional “Context” fields
• Two tabs: Recommendations, Feedback
b. Retrieval Engine
• SentenceTransformers embeds query/context
• FAISS vector search on product embeddings → top-k items
c. RAG / GPT-4 Service
• Constructs prompt: retrieved items + user context
• Sends to OpenAI GPT-4 API → returns natural-language recommendations
d. Reranking Module
• Computes cosine similarity between user context and each retrieved product
• Reorders items (hybrid reranking)
e. Feedback Sentiment Analyzer
• Captures user free-text feedback
• Runs VADER → sentiment score
f. Logging / Audit Trail
• recommendation_log.txt logs queries, prompts, GPT responses
• feedback_log.txt logs feedback + timestamp + sentiment
g. Data Store
• CSV of products (static)
• In-memory or ephemeral storage for embeddings
Proper flow from input to output
1) User enters “Query” + “Context” in Gradio UI → hits Recommend
2) Application loads embeddings of products (precomputed or on-the-fly)
3) Sentence Transformers encodes user input → FAISS returns top N candidates
4) Hybrid Reranking computes context-cosine scores → sorts candidate list
5) Build a RAG prompt with top candidates + context → call GPT-4 API
6) Receive GPT-4 response → parse recommended items + descriptions
7) In Recommendations tab
8) User provides feedback in Feedback tab → app logs feedback text
9) VADER analyzes sentiment → logs sentiment to feedback_log.txt
10) End-to-end logs saved for audit
Architectural patterns & design principles
• Modular pipelines (retrieval, reranking, generation, sentiment)
• Separation of concerns: UI, business logic, external services
• RAG pattern: augment LLM with external knowledge store (FAISS)
• Hybrid ranking: mix vector-search relevance + semantic context scoring
• Auditability: append-only logs for compliance