Python: Agent Evaluation

from rail_score_sdk import RailScoreClient

client = RailScoreClient(api_key="YOUR_RAIL_API_KEY")

result = client.agent.detect_injection(
    text=user_input,
    context="user input",
    sensitivity="medium",
)

if result.injection_detected:
    return f"Input blocked: {result.attack_types}"

result = client.agent.evaluate_tool_call(
    tool_name="send_email",
    tool_input={"to": "admin@company.com", "body": "Click here: http://suspicious.com"},
    agent_context="Customer support chatbot",
    allowed_tools=["send_email", "search_kb"],
)

if result.recommendation == "block":
    return f"Tool call blocked: {result.explanation}"
elif result.recommendation == "warn":
    log.warning(f"Risky tool call: {result.flags}")

# Safe है, आगे बढ़ें
execute_tool(tool_name, tool_input)

result = client.agent.scan_tool_result(
    tool_name="search_database",
    tool_result=raw_tool_output,
    redact_pii=True,
)

if result.pii_detected:
    # Redacted version use करें
    safe_output = result.redacted_result
    log.info(f"PII redacted: {result.pii_types}")
else:
    safe_output = raw_tool_output

# safe_output को agent को वापस pass करें

async def safe_agent_turn(user_input, tool_name, tool_input):
    # 1. User input में injection check करें
    inj = client.agent.detect_injection(text=user_input)
    if inj.injection_detected:
        return "Invalid input."

    # 2. Execute करने से पहले tool call evaluate करें
    call_check = client.agent.evaluate_tool_call(
        tool_name=tool_name,
        tool_input=tool_input,
    )
    if call_check.recommendation == "block":
        return "Tool call not allowed."

    # 3. Tool execute करें
    raw_result = await execute_tool(tool_name, tool_input)

    # 4. Agent को pass करने से पहले result scan करें
    scan = client.agent.scan_tool_result(
        tool_name=tool_name,
        tool_result=raw_result,
        redact_pii=True,
    )
    return scan.redacted_result if scan.pii_detected else raw_result