Pharos RAG provides 6 well-named, read-only retrieval tools with detailed Chinese descriptions and comprehensive input schemas. Tool names follow verb_noun convention (retrieve, list_documents, get_outline, get_document, expand, retrieve_grouped). Schemas include types, enums, and descriptions for most parameters. However, descriptions are lengthy (200-400+ chars) and written in Chinese, making them harder for English-speaking LLMs to parse optimally. Output schemas are partially documented in code comments but not formally exposed in tool definitions. Error handling is structured (status/retriable/hint pattern) but lacks recovery guidance in tool descriptions. No tool annotations (readOnlyHint, etc.) despite all being read-only. Missing parameter constraints (e.g., top_k bounds, context_size ranges).

• Translate descriptions to English and condense to 80-150 chars per tool. E.g., 'retrieve' → 'Search knowledge base with hybrid/dense/sparse strategies. Returns ranked hits with context status (full_section/section_window/omitted_budget) to guide follow-up expand() calls.'
• Add readOnlyHint: true to all 6 tool definitions to signal safety and enable agent retry logic.
• Formally document output schema in each tool definition: hits[] array with fields (doc_id, chunk_id, anchor, text, context_status, score, n_tokens, trust, content_raw, image_path); status enum; retriable boolean; hint string.
• Add min/max constraints to numeric parameters: top_k (1-100), context_size (500-9999), rerank_top_n (1-50). Include in parameter descriptions.
• Enhance tool descriptions with recovery hints: 'If status=empty, try broader query or list_documents() to verify coverage. If status=backend_unavailable, retry after 5s.'