Configuration YAML Schema Reference

1
models:
2
  - type: main                    # Required: Model type
3
    engine: openai                # Required: LLM provider
4
    model: gpt-4                  # Required: Model name
5
    mode: chat                    # Optional: "chat" or "text" (default: "chat")
6
    api_key_env_var: OPENAI_KEY   # Optional: Environment variable for API key
7
    parameters:                   # Optional: Provider-specific parameters
8
      temperature: 0.7
9
      max_tokens: 1000
10
    cache:                        # Optional: Caching configuration
11
      enabled: false
12
      maxsize: 50000

1
models:
2
  - type: my_safety_model
3
    engine: self-hosted
4
    model: my-org/custom-safety-model
5
6
rails:
7
  input:
8
    flows:
9
      - content safety check input $model=my_safety_model

1
models:
2
  - type: content_safety
3
    engine: nim
4
    model: nvidia/llama-3.1-nemotron-safety-guard-8b-v3
5
    cache:
6
      enabled: true
7
      maxsize: 50000
8
      stats:
9
        enabled: false
10
        log_interval: null

1
rails:
2
  input:
3
    parallel: false
4
    flows:
5
      - self check input
6
      - check jailbreak
7
8
  output:
9
    parallel: false
10
    flows:
11
      - self check output
12
    streaming:
13
      enabled: false
14
      chunk_size: 200
15
      context_size: 50
16
      stream_first: true
17
18
  retrieval:
19
    flows:
20
      - check retrieval sensitive data
21
22
  dialog:
23
    single_call:
24
      enabled: false
25
      fallback_to_multiple_calls: true
26
    user_messages:
27
      embeddings_only: false
28
29
  actions:
30
    instant_actions: []
31
32
  tool_output:
33
    flows: []
34
    parallel: false
35
36
  tool_input:
37
    flows: []
38
    parallel: false
39
40
  config:
41
    # Rail-specific configurations

1
rails:
2
  input:
3
    parallel: false      # Execute flows in parallel
4
    flows:
5
      - self check input
6
      - check jailbreak
7
      - mask sensitive data on input

1
rails:
2
  output:
3
    parallel: false
4
    flows:
5
      - self check output
6
      - self check facts
7
    streaming:
8
      enabled: false
9
      chunk_size: 200
10
      context_size: 50
11
      stream_first: true

1
rails:
2
  retrieval:
3
    flows:
4
      - check retrieval sensitive data

1
rails:
2
  dialog:
3
    single_call:
4
      enabled: false
5
      fallback_to_multiple_calls: true
6
    user_messages:
7
      embeddings_only: false
8
      embeddings_only_similarity_threshold: null
9
      embeddings_only_fallback_intent: null

1
rails:
2
  actions:
3
    instant_actions:
4
      - action_name_1
5
      - action_name_2

1
rails:
2
  tool_output:
3
    flows:
4
      - validate tool parameters
5
    parallel: false
6
7
  tool_input:
8
    flows:
9
      - filter tool results
10
    parallel: false

1
rails:
2
  config:
3
    jailbreak_detection:
4
      # Heuristics-based detection
5
      server_endpoint: null
6
      length_per_perplexity_threshold: 89.79
7
      prefix_suffix_perplexity_threshold: 1845.65
8
9
      # NIM-based detection
10
      nim_base_url: "http://localhost:8000/v1/"
11
      nim_server_endpoint: "classify"
12
      api_key_env_var: "JAILBREAK_KEY"

1
rails:
2
  config:
3
    sensitive_data_detection:
4
      recognizers: []
5
      input:
6
        entities:
7
          - PERSON
8
          - EMAIL_ADDRESS
9
          - PHONE_NUMBER
10
          - CREDIT_CARD
11
        mask_token: "*"
12
        score_threshold: 0.2
13
      output:
14
        entities:
15
          - PERSON
16
          - EMAIL_ADDRESS
17
      retrieval:
18
        entities: []

1
rails:
2
  config:
3
    injection_detection:
4
      injections:
5
        - sqli
6
        - template
7
        - code
8
        - xss
9
      action: reject    # "reject" or "omit"
10
      yara_path: ""
11
      yara_rules: {}

1
rails:
2
  config:
3
    fact_checking:
4
      parameters:
5
        endpoint: "http://localhost:5000"
6
      fallback_to_self_check: false

1
rails:
2
  config:
3
    content_safety:
4
      multilingual:
5
        enabled: false
6
        refusal_messages:
7
          en: "Sorry, I cannot help with that."
8
          es: "Lo siento, no puedo ayudar con eso."

1
rails:
2
  config:
3
    autoalign:
4
      parameters: {}
5
      input:
6
        guardrails_config: {}
7
      output:
8
        guardrails_config: {}

1
rails:
2
  config:
3
    patronus:
4
      input:
5
        evaluate_config:
6
          success_strategy: all_pass  # or any_pass
7
          params: {}
8
      output:
9
        evaluate_config:
10
          success_strategy: all_pass
11
          params: {}

1
rails:
2
  config:
3
    clavata:
4
      server_endpoint: "https://gateway.app.clavata.ai:8443"
5
      policies: {}
6
      label_match_logic: ANY  # or ALL
7
      input:
8
        policy: "policy_alias"
9
        labels: []
10
      output:
11
        policy: "policy_alias"
12
        labels: []

1
rails:
2
  config:
3
    pangea:
4
      input:
5
        recipe: "recipe_key"
6
      output:
7
        recipe: "recipe_key"

1
rails:
2
  config:
3
    trend_micro:
4
      v1_url: "https://api.xdr.trendmicro.com/beta/aiSecurity/guard"
5
      api_key_env_var: "TREND_MICRO_API_KEY"

1
rails:
2
  config:
3
    ai_defense:
4
      timeout: 30.0
5
      fail_open: false

1
rails:
2
  config:
3
    private_ai_detection:
4
      server_endpoint: "http://localhost:8080/process/text"
5
      input:
6
        entities: []
7
      output:
8
        entities: []
9
      retrieval:
10
        entities: []

1
rails:
2
  config:
3
    fiddler:
4
      fiddler_endpoint: "http://localhost:8080/process/text"
5
      safety_threshold: 0.1
6
      faithfulness_threshold: 0.05

1
rails:
2
  config:
3
    guardrails_ai:
4
      input:
5
        validators:
6
          - name: toxic_language
7
            parameters:
8
              threshold: 0.5
9
            metadata: {}
10
      output:
11
        validators:
12
          - name: pii
13
            parameters: {}

1
prompts:
2
  - task: self_check_input
3
    content: |
4
      Your task is to check if the user input is safe.
5
      User input: {{ user_input }}
6
      Answer [Yes/No]:
7
    output_parser: null
8
    max_length: 16000
9
    max_tokens: null
10
    mode: standard
11
    stop: null
12
    models: null    # Restrict to specific engines/models

1
instructions:
2
  - type: general
3
    content: |
4
      You are a helpful assistant.

1
sample_conversation: |
2
  user "Hello there!"
3
    express greeting
4
  bot express greeting
5
    "Hello! How can I assist you today?"
6
  user "What can you do for me?"
7
    ask about capabilities
8
  bot respond about capabilities
9
    "As an AI assistant, I can help you with a wide range of tasks."

1
knowledge_base:
2
  folder: kb
3
  embedding_search_provider:
4
    name: default
5
    parameters: {}
6
    cache:
7
      enabled: false

1
core:
2
  embedding_search_provider:
3
    name: default
4
    parameters: {}

1
tracing:
2
  enabled: false
3
  adapters:
4
    - name: FileSystem
5
  span_format: opentelemetry
6
  enable_content_capture: false

1
streaming: false

1
import_paths:
2
  - path/to/shared/config

1
models:
2
  # Main application LLM
3
  - type: main
4
    engine: nim
5
    model: meta/llama-3.1-70b-instruct
6
    parameters:
7
      temperature: 0.7
8
9
  # Content safety model
10
  - type: content_safety
11
    engine: nim
12
    parameters:
13
      base_url: "http://localhost:8000/v1"
14
      model_name: "nvidia/llama-3.1-nemotron-safety-guard-8b-v3"
15
16
  # Embeddings
17
  - type: embeddings
18
    engine: FastEmbed
19
    model: all-MiniLM-L6-v2
20
21
rails:
22
  input:
23
    flows:
24
      - content safety check input $model=content_safety
25
26
  output:
27
    flows:
28
      - content safety check output $model=content_safety
29
    streaming:
30
      enabled: true
31
32
  config:
33
    jailbreak_detection:
34
      nim_base_url: "http://localhost:8001/v1/"
35
36
prompts:
37
  - task: content_safety_check_input $model=content_safety
38
    content: |
39
      Check if this content is safe: {{ user_input }}
40
    output_parser: nemoguard_parse_prompt_safety
41
    max_tokens: 50
42
43
instructions:
44
  - type: general
45
    content: |
46
      You are a helpful, harmless, and honest assistant.
47
48
streaming:
49
  enabled: true