Juliagomes/arize default response

README.md

@@ -49,7 +49,7 @@ We defer readers to the arxiv paper to see the effectiveness of these prompts ag
 ## Installation
 
 ```bash
-$ guardrails hub install hub://guardrails/validator_template
+$ guardrails hub install hub://arize/dataset_embeddings
 ```
 
 ## Usage Examples
@@ -97,6 +97,7 @@ Initializes a new instance of the ValidatorTemplate class.
 **Parameters**
 - **`threshold`** *(str)*: When the Guard is called on an input message, the Guard checks the cosine distance between the embedded incoming message and the embedded chunks from the source dataset. If any of the chunks are within the `threshold` (defaults to 0.2), then the Guard will take the `on_fail` action.
 - **`on_fail`** *(str, Callable)*: The policy to enact when a validator fails.  If `str`, must be one of `reask`, `fix`, `filter`, `refrain`, `noop`, `exception` or `fix_reask`. Otherwise, must be a function that is called when the validator fails.
+- **`default_response`** *(str)*: The default response to use if the Validator fails. This will override the `on_fail` action and be used as a default response if the Guard fails.
 - **`sources`** *List[str]*: Specifies a source dataset with examples of either user input messages or LLM output messages that we would like to Guard against. We recommend including 10 examples.
 - **`embed_function`** *(Callable)*: The embedding function used to embed both the text chunks from `sources` and the input messages.
 - **`chunk_strategy`** *(EmbeddingChunkStrategy)*: The strategy to use for chunking in the Guardrails AI helper `get_chunks_from_text`.

validator/main.py

@@ -2,8 +2,6 @@
 import numpy as np
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 from enum import Enum
-import pandas as pd
-import os
 import logging
 
 import numpy as np
@@ -15,6 +13,7 @@
     ValidationResult,
     Validator,
     register_validator,
+    OnFailAction
 )
 from llama_index.embeddings.openai import OpenAIEmbedding
 
@@ -44,6 +43,9 @@ class EmbeddingChunkStrategy(Enum):
     TOKEN = 3
 
 
+ARIZE_DEFAULT_RESPONSE = "I'm sorry, I cannot respond to that. Please try rephrasing or let me know if I can help you with something else."
+
+
 @register_validator(name="arize/dataset_embeddings", data_type="string")
 class ArizeDatasetEmbeddings(Validator):
     """Validates that user-generated input does not match dataset of jailbreak
@@ -52,11 +54,15 @@ class ArizeDatasetEmbeddings(Validator):
     def __init__(
         self,
         threshold: float = 0.2,
-        on_fail: Optional[Callable] = None,
+        on_fail: Optional[Union[Callable, OnFailAction]] = None,
         **kwargs,
     ):
+        if kwargs.get("default_response") is not None:
+            on_fail = "fix"
+        self.fix_value = kwargs.get("default_response", ARIZE_DEFAULT_RESPONSE)
+
         super().__init__(
-            on_fail, threshold=threshold, **kwargs
+            on_fail=on_fail, threshold=threshold, **kwargs
         )
         self._threshold = float(threshold)
         if kwargs.get("sources") is None:
@@ -114,6 +120,7 @@ def validate(self, value: Any, metadata: Dict[str, Any]) -> ValidationResult:
                 error_message=(
                     f"The following message triggered the ArizeDatasetEmbeddings Guard:\n\t{user_message}"
                 ),
+                fix_value=self.fix_value
             )
         # All chunks exceeded the cosine distance threshold
         return PassResult(metadata=metadata)