Back to case studies
AI · RAG System

GiftxAI

Enterprise-grade RAG-powered gift recommender with triple-validation, FAISS vector store, Groq Llama 3.3 backbone, and zero-hallucination guardrails.

GiftxAI 1GiftxAI 2

Problem

Generic LLM gift recommenders hallucinate SKUs that don't exist. The system needed to recommend only from a verified catalog and explain its reasoning.

Approach

  1. Indexed product catalog with FAISS over E5-base embeddings.
  2. Wrote a triple-validation chain: retrieval relevance check → factual grounding check → catalog existence check.
  3. Used Groq Llama 3.3 70B for low-latency generation; fell back to a smaller model on rate-limit.
  4. Surfaced retrieval traces in the UI so users could see why each gift was chosen.

Results

Hallucinations
0 / 200 sample queries
p50 latency
740 ms
Relevance@5
0.82