Loading ToolBench
MCP Server for assessing grounding ratio between generated text and source context to detect hallucinations
Single tool with basic schema and minimal descriptions. Tool name lacks action verb (should be 'check_hallucination' or 'verify_grounding'). Parameter descriptions are present but extremely terse (7-8 chars each). No output schema documented. No error handling guidance. No pagination, no result limits, no recovery paths. Tool definition is visible and properly registered, but quality gaps are substantial across naming, description depth, and output documentation.
Assess grounding ratio between generated text and source context.
Tool name 'verify_hallucination' lacks action verb prefix. Should be 'check_hallucination' or 'verify_grounding' to signal intent clearly to LLMs.
Parameter descriptions are critically short (7-8 characters). 'The generated text to verify for hallucinations' is 50 chars, but 'The source context to ground the answer against' is 48 chars, both below the 72-char baseline. No format constraints, no examples of valid input, no guidance on expected length or structure.
Output schema is not documented. Tool returns a complex nested object (grounded, overall_grounding_score, supported_sentences, total_sentences, sentence_details array with sentence, overlap_ratio, is_supported fields), but LLMs have no schema to parse it. This forces unstructured parsing and risks field extraction errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | 2026-07-28+ | v2 |
No error handling or recovery guidance. Tool will fail silently on empty input, malformed text, or edge cases (e.g., answer with no sentences). No error messages guide the LLM on what went wrong or how to retry.
Tool description (54 chars) is below the 10 - 1024 character baseline and lacks context on WHEN to use it, WHAT it returns, and HOW to interpret results. LLMs cannot determine if this tool is appropriate for their task.