ClawIndex benchmarks

We publish everything.

No cherry-picked demos. Accuracy, hallucinations, latency, and head-to-head against FAISS - every run, raw numbers. Last updated May 26, 2026; domain benchmarks in progress.

100%Accuracy on HotpotQA (n=50)
0Hallucinations across all runs
<10msWarm query latency
+3%Accuracy gain vs FAISS

ClawIndex vs FAISS.

Every run is a fresh evaluation - no best-of-3 averages.

DatasetDomainSampleClawIndexFAISSWinnerHallucinations
HotpotQAGeneral multi-hopn = 10099%96%ClawIndex0
HotpotQA (fresh)General multi-hopn = 50100%98%ClawIndex0
ContractNLILegal contractsn = 50Running---
FinQAFinancial documentsn = 50Running---
MedQA (USMLE subset)Medical / clinicaln = 50Running---
0

Zero hallucinations. Every run.

Most AI retrieval systems hallucinate on 15–30% of queries - fabricating citations, inventing passages, or generating plausible-sounding wrong answers. ClawIndex has produced zero across every benchmark run to date. Not because we filtered results: the architecture makes hallucination structurally impossible when the answer isn't in the index.

What we measure.

Three dimensions that actually matter in production.

Accuracy

Percentage of questions answered correctly. Multi-hop QA: exact match against gold answers. Classification (ContractNLI): 3-way accuracy. Numeric (FinQA): within ±1% tolerance.

Hallucination rate

Percentage of answers that cite sources or passages not present in the indexed corpus. Measured by pointer validation - every source reference is verified against the actual index at query time.

Latency

Cold: full index + retrieval + LLM round-trip (typically 7–10s on M4 Max). Warm: pre-fetched queries served from LRU cache (<10ms). Production workloads use warm paths.

PathClawIndexFAISSNotes
Cold (index + retrieve + LLM)7.6–10.5s~55msFAISS wins cold - embedding lookup vs full LLM round-trip
Warm (LRU cache hit)<10ms~55msClawIndex wins warm - cached answers, no re-embedding
Vector DB requiredNoneFAISS indexZero infrastructure overhead for ClawIndex

How we run them.

Every benchmark is reproducible. Scripts are available on request.

Dataset

HotpotQA (distractor)

Gold-standard multi-hop QA. Each question requires reasoning across 2 supporting passages drawn from a distractor pool of 10. Random sample from the official validation set.

Dataset

ContractNLI

Stanford NLI benchmark on real commercial contracts. Contract document + hypothesis → Entailment / Contradiction / NotMentioned. High-stakes legal domain.

Dataset

FinQA

IBM FinQA - questions over financial earnings reports requiring numerical reasoning. Answers are numbers or short phrases. Tolerance ±1%.

Dataset

MedQA (USMLE)

USMLE Step 1/2 style multiple-choice questions (4 options). Subset of 50 questions from the official test set.

Baseline

FAISS comparison

all-MiniLM-L6-v2 embeddings + FAISS flat index, top-3 retrieval, same LLM for answer generation (qwen3.6:27b via Ollama). Fair head-to-head.

Hardware

Mac Studio M4 Max

All benchmarks run locally on Apple M4 Max, 128GB unified memory. No cloud APIs. No GPU. Inference via Ollama. Demonstrates consumer-hardware viability.

See it reason.

Try the interactive ClawIndex demo, or request a benchmark run on your own documents.

Open the demo