genesis-rag-grounding-regression

genesis-rag-grounding-regression

Deterministic RAG factual grounding verifier and hallucination regression kit.

---

The Problem

Retrieval-Augmented Generation (RAG) applications frequently suffer from insidious subtle hallucinations:

• Models introduce unsupported numbers (e.g. inflating discounts, fabricating transaction amounts, hallucinating statistical claims).
• Answers weave external parametric training memory that contradicts retrieved enterprise context chunks.
• LLM-as-a-judge evaluators are non-deterministic, slow, expensive, and subject to their own hallucinations.

The Solution

genesis-rag-grounding-regression provides deterministic, fast, zero-dependency factual grounding auditing for RAG outputs. It computes sentence-level lexical alignment against source chunks and performs strict numeric claim validation to catch fabricated quantities, dates, and amounts before answers reach users or production databases.

---

Key Features

• Zero Third-Party Dependencies: Pure Python 3 standard library.
• Sentence-Level Grounding: Tokenizes outputs into discrete claims and audits lexical citation coverage.
• Strict Numeric Verification: Flags any numeric value (currencies, percentages, quantities, years) in the answer not explicitly present in source chunks.
• Normalized Grounding Score: Generates factual grounding metrics from 0.00 to 1.00 with customizable failure gates (--min-score).
• Instant CI Execution: Audits standard multi-paragraph RAG outputs in under 2ms.

---

Quickstart

# Clone the repository
git clone git@github.com:genesiscode2026/genesis-rag-grounding-regression.git
cd genesis-rag-grounding-regression

# Verify answer against retrieved context chunks
./bin/genesis-rag-grounding --contexts ./contexts.json --answer ./generated_answer.txt

# Enforce strict 0.90 grounding threshold in CI
./bin/genesis-rag-grounding --contexts contexts.json --answer answer.txt --min-score 0.90

# Output machine-readable JSON report
./bin/genesis-rag-grounding --contexts contexts.json --answer answer.txt --json

---

Commercial Distribution

• Product Name: genesis-rag-grounding-regression
• Catalog ID: P11
• Price: $79
• Seller: genesiscode2026
• Authorized Payout Rail: USDC on Base (0xC6F86e170411182114FcCdb28793dC76B5e8D144)
• License: MIT