Official-source capability guide · vendor disclosed

Choosing a hallucination detection and grounding tool

PrismShine, Vectara HHEM-2.1-Open, Ragas faithfulness, LettuceDetect, and Guardrails AI provenance validators differ in output, model cost, span detail, enforcement, and runtime evidence. Insight IT Solutions develops PrismShine; only the HHEM row has a published PrismShine head-to-head.

“Hallucination detector” is not one category. Some tools score factual consistency, some extract unsupported spans, some run LLM-judge evaluation, and some enforce a runtime verdict. Decide whether your use case is offline evaluation, online gating, debugging, or cause-side prevention.

Tool	Primary method	Output	LLM dependency	Best fit
PrismShine	Deterministic cause/copy/coverage tiers, optional ONNX spans and judge	Pass/block/review `ShineVerdict`, reason, evidence hash, resolution gate	Zero LLM on default path; optional gray-zone judge	Online enforcement plus optional runtime evidence for retrieval, tool, and cache failures
HHEM-2.1-Open	Open premise/hypothesis factual-consistency classifier	Support score from 0 to 1	No generative LLM required	Focused, efficient factual-consistency scoring and evaluation integrations
Ragas faithfulness	Extract claims, then check each claim against retrieved context	Faithfulness ratio from 0 to 1	Common path uses an LLM; official HHEM-backed option exists	Dataset-level RAG evaluation and diagnostic metric suites
LettuceDetect	ModernBERT/EuroBERT token classifier	Unsupported answer spans with confidence	No generative LLM required	Long-context, multilingual span localization and developer review
Guardrails AI provenance	Embedding similarity or LLM validation against provided sources	Validator pass/fail with configured on-fail action	Depends on selected provenance validator	Teams already composing Guardrails Hub validators and reask/fix workflows

Decision factors

Score or enforce

HHEM and Ragas naturally produce evaluation scores. LettuceDetect provides unsupported spans. Guardrails AI wraps validators with configurable failure actions. PrismShine’s core contract is a named verdict designed for application enforcement and audit records.

Content or cause

Most grounding tools compare an answer with supplied context. PrismShine can also consume explicit runtime evidence such as empty retrieval, tool error, stale cache, and trace completeness. That advantage only exists when the application actually supplies the evidence contract.

Model and cost path

HHEM and LettuceDetect are encoder-class models. Ragas faithfulness commonly uses an LLM to extract and verify claims, though official HHEM-backed faithfulness is available. PrismShine keeps its default path deterministic/ONNX and reserves a judge for gray zones.

Span detail and languages

LettuceDetect specializes in unsupported spans and documents multilingual model options. PrismShine’s current default is English-centric and optional Tier 3 provides spans. HHEM returns a consistency score rather than a runtime trace or policy reason.

What PrismShine has measured

The published 20 July 2026 Azure ACI receipt compares PrismShine with HHEM-2.1-Open on identical HaluEval data. B1 QA uses 200 samples: F1 0.831 versus 0.746 and p50 90 ms versus 216 ms. B2 numbers uses 50 samples: PrismShine F1 1.000 with zero false positives versus HHEM F1 0.926. The Bsum summarization track remains modest at F1 0.600 versus 0.474, and HHEM has higher Bsum AUROC.

There is no completed Ragas performance row in the canonical PrismShine report. LettuceDetect appears in a separate small runtime integration receipt, not the primary product scoreboard. PrismShine publishes no performance claim against Guardrails AI or NVIDIA NeMo Guardrails. See the full PrismShine vs HHEM methodology and limitations.

Honest limitations

All groundedness tools can approve an answer that faithfully repeats incorrect, poisoned, or stale source material.
PrismShine verifies buffered output; it does not retract tokens that have already streamed to a user.
PrismShine’s run4-parity Tier-3 ONNX artifact is optional and approximately 1 GB; it is not bundled in the bare wheel.
Model scores require threshold calibration and a representative false-positive holdout before production gating.
Offline evaluation quality does not automatically prove online reliability, recovery behavior, or audit completeness.
Runtime cause detection requires accurate instrumentation; incomplete traces can hide the very failure being checked.

Evaluation plan

Create a dataset that reflects your domains, languages, answer lengths, structured outputs, calculations, source quality, and retrieval failure modes. Measure F1 and AUROC, but also inspect false positives, unsupported-span usefulness, latency after warm-up, memory, model downloads, judge cost, and the action your system takes after detection. Run separate tests for wrong evidence, no evidence, tool failure, stale cache, and legitimate conversational memory.

For production, pair answer grounding with input security and source integrity. PrismGuard handles prompt injection; PrismShine handles cause/effect evidence; neither converts untrusted source content into world truth.

Official source registry

HHEM: official model card

Ragas: faithfulness documentation

LettuceDetect: repository · documentation

Guardrails AI: Provenance LLM · Provenance Embeddings

PrismShine: repository · PyPI · demo

Official documentation accessed 21 July 2026. Product names and trademarks belong to their owners.