Retrieval-Augmented AI Agents for Entity Matching at Scale
summary
Retrieval plus multi-stage LLM reasoning powers a scalable pipeline that builds a unified product graph in batch and matches new offers in real time via a four-layer stack (recall, L1, L2, L3 multimodal). The system reports about 98.7% precision and ~70–74% recall for offer matching and ~98% precision / 96% completeness for product synthesis, uses DeBERTa v3–style transformers and vision LLMs, runs pre-flight checks, adaptive thresholds, rollbacks, and continuous feedback to self-audit and improve.
AI-written summary · about
chapters (15)
- 0:00Introduction to retrieval-augmented AI agents
- 1:05Fragmented product data challenge
- 3:15Unified product graph solution
- 8:16Why rule-based systems fail
- 10:24Scalable pipeline design
- 11:50Multi-layer measurement framework
- 14:41Self-auditing with RAG
- 16:08Product synthesis pipeline
- 19:01Four-layer offer matching
- 21:34Mismatches with multimodal reasoning
- 22:35Learning from hard negatives
- 24:41High precision and recall
- 25:23Pre-flight checks and guardrails
- 26:05Impact on user trust
- 27:30Future roadmap
related talks
24mConversational Search with LLMs
47mModern Approaches to Product Matching Using LLMs
16mBeyond LLMs - The Next Frontier of Machine Reasoning
36mLLMs and AI agents demystified
01h 30mOptimizing Retrieval for Agentic Systems
44mProduction-Grade LLMs & AI Agents
21marchive – Clara Zang
33mDeep research at scale