RAG corpus poisoning detector — embedding anomalies, backdoor triggers
Single tool 'scan' has a reasonable name and description, but critical gaps undermine quality. The input schema is visible and typed (corpus: string, dup_threshold: float), but parameter descriptions are minimal (under 20 chars for dup_threshold). No output schema is documented, the tool's return structure is unknown to LLMs. The description (67 chars) is adequate but lacks WHEN to use it, prerequisites, or what fields are returned. No error handling guidance. The tool appears to be a wrapper around an internal scan_fn, but the actual implementation logic is not visible in the provided code, only the MCP server registration stub. This makes it impossible to verify idempotency, side effects, or recovery paths. The server uses a custom 'cognis_core.mcp' framework (not the official MCP SDK), which raises concerns about spec compliance.
RAG corpus poisoning detector — embedding anomalies, backdoor triggers
No output schema documented. LLMs cannot infer what fields the scan tool returns, preventing downstream tool chaining and forcing agents to guess at response structure.
Parameter 'dup_threshold' description is too short (under 20 chars: 'near-duplicate Jaccard threshold (default: 0.9)'). Does not explain valid range, units, or when to adjust it. LLMs cannot reason about appropriate values.
Tool description lacks WHEN to use it and prerequisites. Does not state whether the tool modifies state, whether it is idempotent, or what errors it may return. Agents cannot plan safely.
Implementation hidden behind custom 'cognis_core.mcp' framework. Actual scan logic not visible in provided code, only a stub calling build_mcp_server(). Cannot verify error handling, input validation, or idempotency.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
No error handling guidance. Tool description does not explain what happens if corpus file is missing, malformed, or unreadable. No recovery hints for LLMs.