Lemon Cloud AI Tax Consulting Expert

Problem Addressed

Business tax consulting often gets stuck on fragmented regulations, conflicting old and new rules, and unclear applicability. The same question can depend on tax type, region, tax period, taxpayer entity, and invoice scenario; answering from model memory can produce unsupported conclusions. @org-422lyynw/tax-knowledge targets such practice-oriented tax Q&A by requiring answers to be grounded in local tax-knowledge retrieval results before any conclusion is drawn.

How It Works

• Question decomposition: identifies multiple independent questions in one user message while preserving conditions such as tax type, region, period, entity type, and business scenario; it avoids over-splitting one question during the first search.
• Retrieval gating: runs one primary query per question; keyword-level supplemental searches are allowed only after a successful query returns zero records or records that are clearly irrelevant. API Key, capability permission, timeout, or interface errors block policy conclusions.
• Validity checks: does not rely only on invalid_type; it compares issue dates, effective dates, and revision or repeal relationships to prefer the latest policy applicable to the user's period.
• Traceable answers: outputs key points, applicability, conclusions, evidence, recommendations, and risk notes; each evidence item must include Source link: or state that the knowledge base did not return one.

Boundaries and Caveats

The skill answers only from successfully retrieved knowledge-base records in the current turn, not from external web pages, model memory, or unretrieved user text. For large amounts, historical adjustments, audits, or disputes, users should still consult the competent tax authority or qualified professionals. If the API Key, capability permission, or input encoding is faulty, the workflow should pause rather than treating the error as a policy non-result.