Detect Hallucination Against Retrieved Context
The problem, from first principles
A RAG answer is supposed to be grounded in the context it retrieved — every factual claim it makes should trace back to something the context actually said. The cheapest possible check for this doesn't need a model call at all: split the answer into individual claims, and check whether each one appears, verbatim, somewhere in the retrieved context. A claim that doesn't is flagged as unsupported — a naive but genuinely useful first-pass faithfulness check, the same kind of exact-match philosophy behind Track Citations So Every Claim Maps to a Source.
From theory to code
You're given claim_sentences (individual sentences the answer asserted) and context_sentences (the retrieved context). Implement detect_hallucination(claim_sentences, context_sentences). A claim is flagged as hallucinated (True) iff it does not appear verbatim (as an exact substring) within any context sentence. A claim appearing verbatim in at least one context sentence is grounded (False).
Return one bool per claim, in the same order as claim_sentences.
Constraints
- 1 to 200 claims; 0 to 200 context sentences; matching is exact-substring, case-sensitive.
Hints
Hint 1
For each claim, any(claim in context for context in context_sentences) checks all context sentences in one line — a claim is grounded iff that check is true for at least one of them.
Hint 2
A claim doesn't need to match a context sentence exactly — it just needs to appear somewhere inside it. claim in context (substring containment), not claim == context (exact equality), is the right check.
Keywords
No results yet
Run your code to check it against the examples, or submit to run every test.