Conceptual RAG with LanceDB, WordNet, and multi-signal hybrid search for MCP
Concept RAG is a well-structured conceptual RAG system with 9 read-only tools featuring detailed descriptions and clear use-case guidance. All tools have non-empty descriptions (194-400+ chars) with explicit 'USE THIS TOOL WHEN' and 'DO NOT USE for' sections, which is excellent for LLM tool selection. All tools have documented input schemas with proper JSON Schema types. However, output schemas are NOT explicitly documented in the visible source code, they are only described in natural language within the tool descriptions. Input parameters mostly have type declarations but descriptions lack specific constraints (min/max, formats, allowed ranges). Tool names follow verb_noun convention (catalog_search, chunks_search, broad_chunks_search, etc.), though there is some naming redundancy: catalog_search, chunks_search, broad_chunks_search, and concept_search are four separate search tools that could benefit from consolidation guidance or clearer differentiation. Error handling and recovery guidance are not visible in the provided source. Parameter descriptions are good but sometimes lack specificity around formatting (e.g., what exactly is a 'source path' format?). All tools are read-only (no state modification), which is positive for safety but means no destructive tool patterns to validate.
Search across all documents for relevant passages. Find where topics are discussed across your entire library. USE THIS TOOL WHEN: User asks 'find passages about [topic]' without specifying a document; Searching for a concept or topic across all documents; Comparing how different documents discuss the same topic; Doing cross-document research; Looking for all occurrences of a theme or concept. DO NOT USE for: Finding documents about a topic (use catalog_search to get document titles first); Searching within a single known document (use chunks_search); Finding which documents discuss a concept (use concept_search); Discovering categories or metadata (use list_categories or category_search). RETURNS: Array of relevant text passages from across all documents with source document references, relevance scores, concept tags, and surrounding context. Shows which documents each passage comes from.
Search documents by title, keywords, content, or concept terms. Find relevant sources in your library. USE THIS TOOL WHEN: User asks 'what documents are about [topic]?' and wants to find source documents; Searching for documents by content or keywords; Need to discover relevant sources before diving into specific passages; Getting an overview of what documents discuss a topic. DO NOT USE for: Searching within a specific known document (use chunks_search); Finding where a concept is discussed across documents (use broad_chunks_search or concept_search); Discovering what categories exist (use list_categories). RETURNS: Array of documents with title, source path, preview text, primary concepts, category assignments, and relevance scores. Includes total count of matching documents.
Find documents by category. Browse documents in a specific domain or subject area. Returns all documents in the category. USE THIS TOOL WHEN: User asks 'what documents do I have about [domain]?' where domain is a category; Browsing documents by subject area (e.g., 'software engineering', 'philosophy'); Need all documents in a specific category, not just search matches; Exploring a domain after discovering categories via list_categories; Getting a complete inventory of documents in a subject area. DO NOT USE for: Searching for specific content within documents (use broad_chunks_search); Finding documents by title or author (use catalog_search); Discovering what categories exist (use list_categories first); Finding where a concept is discussed (use concept_search or source_concepts). RETURNS: Category metadata (name, description, hierarchy, aliases, related categories), statistics (document count, chunk count, concept count), and all documents in the category with titles, previews, and primary concepts.
Output schemas not explicitly documented in source code. Tool descriptions reference return values in natural language, but no formal schema definitions (JSON Schema) for response objects are visible. LLMs cannot reliably parse unstructured output descriptions.
Four search tools (catalog_search, chunks_search, broad_chunks_search, concept_search) may cause LLM confusion due to overlapping functionality. Tool names do not clearly distinguish when to use catalog_search (document-level) vs chunks_search (within a document) vs broad_chunks_search (across all) vs concept_search (by concept name). Descriptions help, but names themselves are not self-documenting.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 69 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 51 | - | v1 |
Search within a specific document's content. Find relevant passages in a known source. USE THIS TOOL WHEN: User has a known document and wants to search within it; Need specific passages from a document about a topic; Drilling down after catalog_search found a relevant document; Extracting relevant sections from a particular source. DO NOT USE for: Searching across all documents (use broad_chunks_search or catalog_search); Finding documents about a topic (use catalog_search first); Discovering concepts (use concept_search or extract_concepts); Finding where a concept appears across documents (use broad_chunks_search). RETURNS: Array of text chunks from the document with source reference, position info (page/section), concept tags, relevance scores, and surrounding context. COMMON WORKFLOW: Use catalog_search to find relevant documents, then use chunks_search with the source path to find specific passages within that document.
Find which documents discuss a specific concept. Locate concept occurrences with hierarchical source breakdown. USE THIS TOOL WHEN: User asks 'which books discuss [concept]?' or 'find [concept]'; Need to locate specific concepts and see all documents mentioning them; Exploring concept usage across documents; Building a concept-centric view of the library; Finding related concepts and their sources. DO NOT USE for: Searching for text content (use catalog_search or broad_chunks_search); Finding concepts in a category (use list_concepts_in_category); Searching within a document (use chunks_search); Extracting concepts from a new document (use extract_concepts). RETURNS: Concept metadata (name, weight/importance, related concepts with similarity scores), hierarchical source breakdown (organized by document → chunks), document count, chunk count, and detailed source list with concept relevance scores. COMMON WORKFLOW: Use list_categories or catalog_search to discover topics, then use concept_search with specific concept names to locate all occurrences.
Extract and analyze concepts from a document. Identify key ideas, themes, and topics in a source. USE THIS TOOL WHEN: User asks 'what are the main concepts in [document]?' or 'extract concepts from [title]'; Need to identify and categorize key topics in a document; Analyzing document themes and ideas; Building concept metadata for a new or existing document; Understanding document structure and primary topics. DO NOT USE for: Finding documents about a concept (use concept_search); Finding where concepts appear (use broad_chunks_search or concept_search); Discovering categories (use list_categories); Searching within documents (use chunks_search). RETURNS: Primary concepts (ranked by importance), related concepts (conceptually connected), concept categories, concept descriptions, document summary, and structured metadata for further analysis.
List all available categories with statistics. Discover what subject areas are in your library. USE THIS TOOL WHEN: User asks 'what categories exist?' or 'what subject areas do you have?'; Starting category-based exploration of the library; Need to discover available domains before drilling into specific categories; Building an overview of the library's organizational structure; Finding the right category name/ID for use with category_search. DO NOT USE for: Finding documents about a specific topic (use catalog_search instead); Searching within document content (use broad_chunks_search instead); Finding where a concept appears (use concept_search or source_concepts); Getting documents in a known category (use category_search with the category name). RETURNS: Array of categories with names, descriptions, aliases, hierarchy paths, parent/child relationships, and statistics (document count, chunk count, concept count). Includes summary with total categories and root category count.
Find all unique concepts appearing in documents of a specific category. Analyze the conceptual landscape of a domain. USE THIS TOOL WHEN: User asks 'what concepts are covered in [domain]?' or 'what topics appear in [category]?'; Analyzing the conceptual landscape of a subject area; Discovering what ideas are discussed within a specific domain; Building a concept inventory for a category (e.g., 'what concepts appear in software engineering books?'); Finding related concepts that span documents in the same domain. DO NOT USE for: Finding where a specific concept appears (use concept_search or source_concepts); Getting concepts from a single document (use extract_concepts with the document); Searching for content about a concept (use broad_chunks_search); Discovering categories (use list_categories first). RETURNS: Category metadata, statistics (total documents, chunks, unique concepts), and list of concepts sorted by document count (how many documents mention each concept) or alphabetically. Includes concept weight/importance scores.
Extract all concepts from a specific document. Understand the conceptual content of a source. USE THIS TOOL WHEN: User asks 'what concepts are in [document]?' or 'what does [book] discuss?'; Need a conceptual overview of a document; Analyzing a specific source's concept coverage; Building a concept inventory for a document; Understanding a source's primary topics and themes. DO NOT USE for: Finding documents about a concept (use concept_search); Finding where concepts appear across documents (use broad_chunks_search); Discovering categories (use list_categories); Searching document content (use catalog_search or chunks_search). RETURNS: Source metadata (title, path, document count info), concept list sorted by importance/weight with relevance scores, concept statistics (total count, primary vs related), and document category assignments.
Parameter descriptions lack specific constraints. 'source' parameter in chunks_search is documented as 'Document source path (e.g. sample-docs/Philosophy/Sun Tzu - Art Of War.pdf)' but does not specify format rules, required structure, or what happens if path is malformed. 'text' parameters in search tools lack guidance on minimum/maximum length or character restrictions.
Enum-typed parameters (sortBy, format, includeChildren) are correctly declared in input schemas, but descriptions do not explicitly repeat the allowed values inline. LLMs may not reliably read JSON Schema enums, repeating valid options in descriptions ('Sort by: name, popularity, or documentCount') ensures clarity.
No error handling or recovery guidance visible in tool definitions. If a tool call fails (e.g., document not found, malformed query), the source code does not show error response schemas or guidance for LLM recovery actions.