response-block-refunds.tengo

// Response Filter: Block Refund Promises
// This filter prevents LLMs from promising refunds to customers

text := import("text")

// Forbidden refund-related patterns
forbidden_patterns := [
    "will refund",
    "issue a refund",
    "provide a refund",
    "get a refund",
    "refund you",
    "send you a refund",
    "process your refund"
]

// For streaming, check the accumulated buffer (contains all text so far)
// For non-streaming, check the full response
check_text := ""
if input.is_chunk {
    check_text = input.current_buffer
} else {
    check_text = input.raw_input
}

// Initialize output at top level (required by Tengo)
output := {
    block: false,
    message: ""
}

// For streaming: wait for at least 100 characters before checking
// This prevents false negatives when patterns span chunks
min_buffer_length := 100

// Only check if we have enough context
if !input.is_chunk || len(check_text) >= min_buffer_length {
    // We have enough text - check for forbidden patterns
    should_block := false
    matched_pattern := ""

    for pattern in forbidden_patterns {
        if text.contains(text.to_lower(check_text), pattern) {
            should_block = true
            matched_pattern = pattern
            break
        }
    }

    // Build compliance events for audit trail
    compliance_events_list := []
    if should_block {
        compliance_events_list = [
            {
                event_type: "refund_promise_detected",
                severity: "critical",
                description: "LLM attempted to promise a refund: '" + matched_pattern + "'",
                metadata: { "matched_pattern": matched_pattern }
            }
        ]
    }

    // Update output (use = not :=)
    output = {
        block: should_block,
        message: should_block ? "Response blocked: Contains forbidden phrase '" + matched_pattern + "'" : "",
        compliance_events: compliance_events_list
    }
}