Here's the data formatted as a markdown table:
| Tool/Library | Violation Handling | Code Example - Initialization | Code Example - Usage/Calling | How the Guardrail is Used |
|---|---|---|---|---|
| AWS Bedrock Guardrails | Exceptions, warnings, output modification | bedrock_client.create_guardrail(name='my_guardrail', ...) |
response = client.invoke_model_with_guardrail(modelId='anthropic.claude-v2', guardrailId='my_guardrail_id', ...) |
Applied to both user inputs and model outputs during inference. Can be used to filter harmful content, block denied topics, redact PII, and filter words. |
| Nvidia Nemo Guardrails | Reject input, alter input, execute actions, reject output, alter output | config = RailsConfig.from_path("path/to/config") |
app = LLMRails(config) new_message = app.generate(messages=[{ "role": "user", "content": "Hello! What can you do for me?" }]) |
Uses Colang to define flows and apply guardrails at various stages in the conversation. Can be used to validate user input, bot messages, and control dialogue flow. |
| Guardrails Python library | Corrective actions, re-prompting, exception | guard = gd.Guard.from_rail_string(rail_str) |
guard.validate("123-456-7890") |
Uses RAIL specifications to define guardrails. Can be used to validate and structure LLM output, enforce types, and apply corrective actions. |
| LangChain | Filter, fix, exception | guard = Guard().use_many( CompetitorCheck(competitors=competitors_list, on_fail="fix"), ToxicLanguage(on_fail="filter"), ) |
chain = promptmodel |
Offers various types of guards, including dataset embeddings, general LLM guards, and RAG LLM guards. Can be used to detect jailbreak attempts, PII, toxicity, and other issues. |
| Arize | Default response, re-prompting, exception | guard = Guard().use(ArizeDatasetEmbeddings, on="prompt", on_fail="exception") |
validated_response = guard( llm_api=openai.chat.completions.create, prompt=prompt, model="gpt-3.5-turbo", max_tokens=1024, temperature=0.5, ) |
Offers various types of guards, including dataset embeddings, general LLM guards, and RAG LLM guards. Can be used to detect jailbreak attempts, PII, toxicity, and other issues. |
| Giskard | Generate Colang rules for mitigation | scan_report = gsk.scan(my_model, my_dataset) |
scan_report.generate_rails("config/generated_rails.co") |
Uses Colang rules to define guardrails. Can be used to detect vulnerabilities and generate Colang rules for mitigation. |