VigilantRAG - Self-Correcting Multi-Stage RAG Engine

Pipeline Trace Map

Input Query

Hybrid Search

Re-ranking

Query Expansion

LLM Generate

NLI Audit

Answer

Verified Factual

Answer will appear here...

Response Time: 0.0s Queries Expanded: 0

Stage Details

Click nodes above to inspect execution telemetry

Select any step in the Pipeline Trace Map above to inspect dense/sparse data hits, Cross-Encoder weights, Query Expansion synonyms, NLI logic, or generation history.

Welcome to VigilantRAG

This is a live playground to test a Self-Correcting Retrieval-Augmented Generation Engine. It guarantees factuality and fixes ambiguous queries on the fly using a multi-stage pipeline:

Hybrid Search

Combines FAISS vector embeddings (all-MiniLM-L6-v2) for meaning, and BM25 keywords for precision.

Cross-Encoder Re-rank

Scores the top 50 matches dynamically, filtering down to the 5 most contextually relevant chunks.

Synonym Expansion

Rewrites queries using thesaurus/LLM synonym trees when initial relevance scores fall below 0.4.

NLI Hallucination Guard

Uses Natural Language Inference (DeBERTa-v3) to audit answers. Discards and regenerates replies if the AI hallucinates details.

Get started by loading some documents in the Ingest Document sidebar, then ask your question above!