AI RAG Eval & Hybrid Search Tester
Test Faithfulness (hallucination detection), Context Precision and Vector + tsvector FTS hybrid RRF ranking in real time.
RAG Eval & Hybrid RRF Search Guide
Relying solely on dense vector similarity in RAG systems often misses exact product codes, acronyms and domain terms. Production-grade assistants need both hybrid RRF ranking and sentence-level Faithfulness evaluation.
Why Reciprocal Rank Fusion (RRF) is the standard
RRF merges dense vector cosine rankings and PostgreSQL tsvector FTS rankings using 1 / (k + rank) without brittle score normalization. Set up the underlying schema with Drizzle Schema & Zod Builder and AI SQL & Postgres Optimizer.
Faithfulness and hallucination detection
LLMs sometimes inject prices, SLA numbers or dates from pretraining that do not exist in the retrieved chunks. Sentence-level grounding highlights unsupported claims so you can catch hallucinations before shipping.
End-to-end RAG workflow
Pair this evaluator with AI RAG Chunking Visualizer for chunking, AI Prompt Injection Tester for security and AI Model Arena for side-by-side LLM benchmarking.