Rule-based hallucination detection

[adaamko]

Motivation

LettuceDetect currently relies on encoder, generative, LLM-judge, and RAGFactChecker paths. A low-compute lexical detector could handle a narrow set of obvious unsupported details before a learned detector runs, while also providing a useful baseline.

Candidate checks include:

• numbers and quantities absent from the evidence;
• dates and durations absent from the evidence;
• named entities with no exact or normalized match;
• deliberately conservative relation checks where precision can be measured.

High precision at low coverage is a useful outcome. The objective is not to replace semantic verification.

Relation to RuleChef

RuleChef can synthesize regex, Python, and spaCy rules from labeled examples. A second-stage experiment could use RAGTruth gold annotations—or observations from a LettuceDetect model—to propose candidate rules.

Synthesized rules inherit errors and biases from their teacher and examples. They must therefore be frozen and evaluated against independently held-out gold labels; generated rules are not automatically correct or fully interpretable merely because their runtime is deterministic.

Proposed first cut

1. Add a rule branch to make_detector returning a RuleBasedDetector(BaseDetector) that implements predict, predict_prompt, and predict_prompt_batch for both spans and tokens output.
2. Begin with hand-written number, date, duration, and entity checks. Keep semantic embeddings out of the first implementation.
3. Evaluate the tier on the RAGTruth test split and report span precision, span recall, example-level F1, and routed coverage alongside the transformer detector.
4. Optionally freeze RuleChef-synthesized candidates and compare them with the hand-written tier under the same protocol.

Acceptance

• HallucinationDetector(method="rule") uses the documented prediction API and requires neither torch nor an API key at runtime.
• Network-free tests cover output formats, offsets, normalization, batch length validation, and representative false-positive cases; python -m pytest passes.
• A reproducible RAGTruth results table reports precision, recall, example-level F1, and coverage for both hand-written and synthesized variants that are attempted.
• Negative or partial results are documented and are a valid contribution.

Non-goals

• Replacing the transformer or generative detectors.
• Claiming that lexical absence proves hallucination.
• Embedding or LLM calls in the first implementation.

Start here

git clone https://github.com/KRLabsOrg/LettuceDetect.git
cd LettuceDetect
pip install -e ".[dev]"
python -m pytest

Relevant interfaces: lettucedetect/detectors/base.py, lettucedetect/detectors/factory.py, and lettucedetect/detectors/prompt_utils.py.