Wauldo Guard
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
