Tutorial: Guardrails (E21)

This tutorial corresponds to the example file examples/E21_guardrails_example.py.

It demonstrates how to enforce safety policies on LLM inputs and outputs. It shows how to: - Configure the built-in KeywordBlocklistGuardrailPlugin. - See how input guardrails can block potentially harmful prompts before they reach the LLM. - See how output guardrails can sanitize or block LLM responses that violate a policy.

Example Code

examples/E21_guardrails_example.py

""" Example: Using Guardrails

---

This example demonstrates how to configure and use Guardrails in Genie Tooling. It uses the built-in KeywordBlocklistGuardrailPlugin.

To Run: 1. Ensure Genie Tooling is installed (poetry install --all-extras). 2. Ensure Ollama is running and 'mistral:latest' is pulled (or configure a different LLM). 3. Run from the root of the project: poetry run python examples/E21_guardrails_example.py Observe how inputs/outputs containing 'secret' or 'forbidden' are handled. """ import asyncio import logging from typing import Optional

from genie_tooling.config.features import FeatureSettings from genie_tooling.config.models import MiddlewareConfig from genie_tooling.genie import Genie

async def run_guardrails_demo(): print("--- Guardrails Example ---") logging.basicConfig(level=logging.INFO)

app_config = MiddlewareConfig(
    features=FeatureSettings(
        llm="ollama",
        llm_ollama_model_name="mistral:latest",
        command_processor="llm_assisted",
        tool_lookup="embedding",
        input_guardrails=["keyword_blocklist_guardrail"],
        output_guardrails=["keyword_blocklist_guardrail"],
    ),
    guardrail_configurations={
        "keyword_blocklist_guardrail_v1": {
            "blocklist": ["secret", "forbidden_word", "sensitive_operation"],
            "case_sensitive": False,
            "action_on_match": "block"
        }
    },
    tool_configurations={
        "calculator_tool": {}
    }
)

genie: Optional[Genie] = None
try:
    print("\nInitializing Genie with Guardrails enabled...")
    genie = await Genie.create(config=app_config)
    print("Genie initialized!")

    print("\n--- Testing Input Guardrail (LLM Chat) ---")
    blocked_input_prompt = "Tell me a secret about the project."
    print(f"Sending potentially blocked input: '{blocked_input_prompt}'")
    try:
        chat_response_blocked = await genie.llm.chat([{"role": "user", "content": blocked_input_prompt}])
        print(f"LLM Response (should not be reached if blocked): {chat_response_blocked['message']['content']}")
    except PermissionError as e_perm:
        print(f"Input Guardrail Blocked: {e_perm}")
    except Exception as e_llm_in:
        print(f"LLM chat error (possibly unrelated to guardrail): {e_llm_in}")

    allowed_input_prompt = "Tell me a fun fact."
    print(f"\nSending allowed input: '{allowed_input_prompt}'")
    try:
        chat_response_allowed = await genie.llm.chat([{"role": "user", "content": allowed_input_prompt}])
        print(f"LLM Response: {chat_response_allowed['message']['content'][:100]}...")
    except Exception as e_llm_allowed:
        print(f"LLM chat error for allowed input: {e_llm_allowed}")


    print("\n--- Testing Output Guardrail (LLM Generate) ---")
    prompt_for_blocked_output = "Write a sentence that includes the word 'secret'."
    print(f"Sending prompt for potentially blocked output: '{prompt_for_blocked_output}'")
    try:
        gen_response = await genie.llm.generate(prompt_for_blocked_output)
        print(f"LLM Generated Text: {gen_response['text']}")
        if "[RESPONSE BLOCKED" in gen_response["text"]:
            print("Output Guardrail successfully blocked the response.")
        elif "secret" in gen_response["text"].lower():
             print("Warning: Output guardrail did not block 'secret' as expected. LLM output was: ", gen_response["text"])
    except Exception as e_gen:
        print(f"LLM generate error: {e_gen}")

    print("\n--- (Illustrative) Tool Usage Guardrail ---")
    print("To test tool usage guardrails, enable 'tool_usage_guardrails' in FeatureSettings")
    print("and try a command that would pass a blocked keyword as a tool parameter.")

except Exception as e:
    print(f"\nAn error occurred: {e}")
    logging.exception("Guardrails demo error details:")
finally:
    if genie:
        await genie.close()
        print("\nGenie torn down.")

if name == "main": asyncio.run(run_guardrails_demo())