Every RAG framework hallucinates. We catch them with a number.

We tested 6 popular RAG frameworks (LangChain, LlamaIndex, Haystack, CrewAI, vanilla LLM, Wauldo) on 70 adversarial prompts covering prompt injection, source contradiction,

out-of-scope refusal, and factual recall. Every framework hallucinated.

The actual numbers (live, daily refresh):

- LangChain — 40% hallucination (42/70 pass)

- LlamaIndex — 54% hallucination (32/70)

- Haystack — 40% hallucination (42/70)

- CrewAI — 29% hallucination (50/70)

- Vanilla LLM — 14% hallucination (60/70)

- Wauldo Guard — 4% hallucination (67/70)

## How it works

Wauldo Guard adds a verification layer that checks every LLM claim against your source data before anything reaches users.

- Numeric trust score — every response returns confidence ∈ [0, 1]. A real number, not "verified: true" hand-waving.

- Source-grounded — every claim mapped back to the exact passage. No source, no answer.

- Deterministic — Rust + regex + structural comparison. Not LLM-as-judge. 100% reproducible.

- Open leaderboard — 6 frameworks, 70 tests, GitHub Actions daily refresh, MIT license.

- 3-line install — pip install wauldo, then wauldo.guard(answer, sources).

## Try it live, no signup

Paste an LLM answer + sources at wauldo.com/tools/trust-score and get a trust score in <500ms. Same engine used in production.

## Pricing

Free $0 · Guard $9/mo · Pro $29/mo · Scale $99/mo · Enterprise custom

Leaderboard: https://wauldo.com/leaderboard

Trust Score Tool: https://wauldo.com/tools/trust-score

GitHub: https://github.com/wauldo/wauldo-leaderboard

Zin Benzin

Projects

Wauldo Guard

Zin Benzin

Projects

Wauldo Guard