Semantic hybrid search (sqlite-vec + FTS5 with optional cross-encoder reranking) over a local knowledge base of Markdown, plain-text, PDF, Office documents and source code, served over MCP or a web UI.
grooveseek has 6 tools with mostly complete schemas and descriptions. The search tool is exceptionally detailed with 13 parameters covering semantic search, filtering, reranking, and diversity control, well above the baseline of 4 params per tool. Descriptions are present and substantive (194-char baseline met). However, several tools lack output schema documentation, parameter descriptions are sometimes generic, and error handling guidance is absent. The rebuild_index tool accepts a WRITE operation but provides no confirmation or dry-run safety guardrails beyond the dry_run flag itself. No tool annotations (readOnlyHint/destructiveHint) are visible in the source.
Retrieve a best-practice document by target name. Resolves the target against configured path templates and returns the matching document.
Retrieve the full text of a document by its resource URI. Returns the complete document content up to a size limit.
Retrieve the topic relationship graph, optionally filtered by seed topics and traversal depth.
List all topics (tags) in the knowledge base with their document counts and relationships.
Rebuild the search index from the knowledge base directory. Walks the filesystem, parses documents, chunks them, embeds chunks, and upserts into the database. Supports progress reporting and dry-run mode.
Search the knowledge base with semantic and full-text hybrid search, with optional reranking and structured filtering by path, tags, and date ranges.
Output schemas not documented for any tool. LLMs cannot plan downstream calls or extract required fields without knowing what get_document, list_topics, or get_topic_graph return.
rebuild_index is a destructive WRITE operation but lacks tool annotations (destructiveHint) and no confirmation/approval flow. Agents could accidentally rebuild the entire index without safeguards.
No error handling guidance in any tool description. If search fails, get_document returns 404, or rebuild_index encounters a parse error, the LLM has no recovery path.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
search parameter descriptions use legacy language ('Prefer `path_globs` / `tags_any` / `tags_all` for new clients') suggesting API evolution without clear deprecation timeline. Ambiguous for LLM parameter selection.
get_topic_graph strategy parameter lacks enum constraint. Description says 'e.g., breadth-first, depth-first' but does not enforce valid values, inviting hallucinated strategies.