Two tools with basic definitions and schemas. semantic_search has a clear description (127 chars) and three typed parameters with descriptions. list_sources has a minimal description (67 chars) and no parameters. Both tools return unstructured strings rather than structured objects, forcing LLMs to parse free-text output. No error handling guidance, no output schema documentation, and no parameter constraints (e.g., limit bounds). Naming is verb-forward (semantic_search, list_sources) but descriptions lack WHEN/WHY context and recovery guidance.

• Return structured JSON from both tools: semantic_search should return {results: [{source_id, title, similarity, chunk_index, total_chunks, text}], total_count} instead of markdown string. list_sources should return {sources: [{name, chunk_count, embedding_model}]}.
• Add bounds to semantic_search 'limit' parameter: 'Maximum number of results to return (1 - 100, default 10)'. Validate input and return error if out of range.
• Enhance semantic_search description: 'Search the user's personal knowledge base for relevant documents. Call this BEFORE answering questions about topics in the indexed sources. Requires sources to be synced via 'lake sync' first. Returns up to N results ranked by similarity.'
• Enhance list_sources description: 'List all available knowledge base sources that have been indexed. Call this first to discover what sources are available before searching. Returns source names, chunk counts, and embedding models used.'
• Add error recovery guidance: If search returns no results, suggest 'No results found. Try a broader query, check spelling, or call list_sources() to verify indexed data exists.'