Loading ToolBench
MCP server for RAG search and project item management with dense/lexical/hybrid search, document chunking, and project memory/doc/bug/todo item storage
Contextarium has well-structured tool definitions with complete JSON schemas and clear descriptions for most tools. All 11 tools have input schemas with type definitions and required fields. Descriptions are generally adequate (100-200 chars), explaining what each tool does. However, parameter descriptions are largely absent, most parameters lack inline documentation explaining their purpose, constraints, or expected format. Output schemas are defined for search/chunk tools and items, but parameter-level guidance is minimal. Error handling and recovery guidance are not evident in the tool definitions. The hybrid_search tool includes a helpful note about sequential execution, but this pattern is not applied consistently across other tools.
Retrieve every chunk extracted from a given documentation URL.
Delete an item from a project.
Dense vector search using cosine similarity over the embedding index.
Retrieve a single item by ID.
Hybrid dense+lexical search with score fusion, MMR and optional reranking. Recommended top_k: 6 (increase for recall). Do not call in parallel; run sequentially.
Lexical BM25 search using DuckDB FTS.
List items in a project, optionally filtered by type and/or status.
Parameter descriptions missing across all tools. Parameters like 'query', 'top_k', 'url', 'project', 'type', 'title', 'body_md', 'tags', 'status', 'meta', 'typed', 'edits' lack inline documentation explaining their purpose, format, constraints, or expected values. LLMs cannot infer parameter semantics from names alone.
Destructive operations (delete_item) lack confirmation or dry-run support. No guidance on recovery or undo. Agents cannot safely preview consequences before executing irreversible deletes.
No error handling or recovery guidance in tool definitions. Tools do not document what errors are retryable, what user-fixable errors look like, or what the agent should do on failure. No examples of error messages or recovery paths.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2025-06-18+ | v2 |
Apply a series of text edits to a doc item's body_md field.
Full-text search over item titles and bodies within a project.
Create a memory/doc/bug/todo item inside a project (uses project or project_id).
Update item fields (title, tags, status, meta, typed). MCP body diff edits for docs go through patch_doc.
Ambiguous parameter naming: 'project' vs 'project_id' both accepted in item tools without clear guidance on which to use or how they differ. LLMs will guess and pass both, causing ambiguous or failing calls.
Search tools (dense_search, lexical_search, hybrid_search) do not document result limits or pagination. No indication of max results returned, whether results are paginated, or how to fetch additional results. Large result sets could exhaust context.