Vidilearn has 7 tools with basic descriptions and partial schemas. Most tools lack comprehensive parameter documentation and output schema definitions. Naming is reasonable (ingest, ask, search, transcript, extract, audit, doctor) but descriptions are terse (avg ~80 chars, below the 194-char baseline). Critical gaps: no output schemas documented, missing parameter type constraints (e.g., 'limit' lacks min/max bounds), no error handling guidance, and no tool annotations. The 'ingest' tool accepts multiple input types (file, directory, RSS, Git, YouTube, URL) but the 'target' parameter lacks format constraints or examples of valid inputs. The 'ask' tool's 'ollama' boolean parameter could default to false but lacks guidance on when synthesis is appropriate. Search tool has debug/hybrid flags but no documentation of what these flags return or how they affect output structure.

• Document output schema for each tool. E.g., 'search' should return [{document_id, chunk_id, text, relevance_score, source_url}]. Include field types and whether results are paginated.
• Add min/max bounds to numeric parameters: limit (1 - 100), lang as enum (en, es, fr, ...), and provider as enum (transformers, ollama).
• Expand tool descriptions to 150 - 250 chars. Include WHEN to use (e.g., 'Use ingest to add new sources to the knowledge base; use ask to retrieve and synthesize answers from existing sources').
• Add error handling guidance: 'If ingest fails with "unsupported format", try extract() instead. If ask returns no results, try search() with a simpler query.'
• Annotate tools with destructiveHint (ingest), readOnlyHint (ask, search, transcript, extract, audit, doctor), and idempotentHint (audit, doctor) so agents understand safety properties.