Loading ToolBench
A service that re-ranks RAG (Retrieval-Augmented Generation) context documents based on query relevance using simple keyword overlap scoring
This server has a single tool 're_rank' with critical quality gaps. The tool has a name that does not clearly indicate action (should be verb_noun like 'rank_documents' or 'rerank_documents'), a description that is present but lacks actionable context, and an input schema that is visible but minimally documented. The implementation is trivial (keyword overlap via set intersection) and lacks error handling, input validation, pagination support, and guidance for LLM selection. The output schema is not documented. Overall, this reads as a proof-of-concept rather than a production-grade agent tool.
Re-ranks documents based on query relevance using keyword overlap scoring
Tool name 're_rank' lacks action verb prefix. Should be 'rank_documents' or 'rerank_documents' to clearly signal to LLM what action occurs. Current name is vague, LLMs cannot easily infer intent.
Description 'Re-ranks documents based on query relevance using keyword overlap scoring' is 77 characters but lacks WHEN to use this tool, prerequisites, or how it differs from alternatives. Should explain: Is this a post-retrieval ranker? Does it require pre-retrieved documents? What similarity threshold matters?
Input schema present but parameter descriptions are generic. 'query' and 'documents' lack guidance on format, length limits, expected structure, or constraints. LLMs cannot infer that query is free-form text or that documents array has no size limit.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 36 | 2026-07-28+ | v2 |
Output schema is not documented. Only visible output is a dict with 're_ranked_documents' key, but no formal schema, type definition, or documentation of the returned structure. LLMs cannot plan downstream operations without knowing output structure.
No error handling visible. Tool will fail silently or raise unhandled exceptions on invalid input (e.g. non-string query, non-string array items, empty documents). No validation, no recovery guidance.
Implementation uses naive keyword overlap (set intersection of split tokens). This is brittle, case-sensitive, ignores stemming/synonyms, treats punctuation as token delimiters. No documentation of limitations or when this ranking strategy is appropriate.
No pagination support. If documents array is large, response could balloon context window. Tool description does not state result limits or guidance on batch sizes.
Missing input validation and constraints. No stated limits on query length, array size, or individual document length. Unbounded inputs can cause performance degradation or memory exhaustion.