Fiddler Guardrails Free Trial Documentation

Rate Limits

The Fiddler Guardrails free trial experience is subject to the following rate limits. To increase these limits, please contact sales (https://www.fiddler.ai/contact-sales).

Trust Model	Requests Per Second	Requests Per Hour	Requests Per Day
Safety	2	70	200
Faithfulness	2	70	200

Understanding Trust Model Scores

Fiddler Trust Models return scores in the range of 0 to 1. These scores represent the model's confidence that the input belongs to the target class (e.g., toxicity, hallucination).

• Higher scores (closer to 1): Higher confidence the input belongs to the target class

• Lower scores (closer to 0): Lower confidence the input belongs to the target class

Threshold Selection

You must select a threshold value between 0 and 1 to convert these scores into binary decisions. This creates a tradeoff:

• Lower thresholds: Catch more true positives but include more false positives

• Higher thresholds: Reduce false positives but might miss some true positives

Our quickstart examples include default thresholds that work well for many applications, but you should adjust these based on your specific requirements and risk tolerance.

Adjusting Your Thresholds

To find the optimal threshold for your use case:

1. Start with the default threshold

2. Monitor both missed detections and false alarms

3. Adjust gradually based on which type of error is more problematic for your application

Safety Model

Banner: The safety guardrails are restricted to a 4k token length limit for the free trial experience. To increase these limits, please contact sales.

This Fiddler Trust Model evaluates prompt and response safety across ten dimensions:

• Jailbreaking

• Illegal content

• Hateful content

• Harassment

• Racism

• Sexism

• Violence

• Sexual content

• Harmful content

• Unethical content

The model requires a single string input and outputs ten distinct scores (0-1 range). For detailed information, see our official documentation.

How to Use Thresholding

Users can apply thresholds on individual safety dimensions (e.g., harmful, violent, racist) or evaluate all of them collectively. This flexibility allows you to tailor how strictly you filter content based on your unique requirements.

Safety Model OpenAPI Spec

openapi: 3.0.3
info:
  title: Fiddler FTL Safety
  version: 1.0.0
servers:
  - url: "https://{fiddler_endpoint}"
paths:
  /v3/guardrails/ftl-safety:
    post:
      summary: Assess the safety or harmfulness of the provided input text.
      operationId: evaluateSafety
      security:
        - bearerAuth: []
      requestBody:
        required: true
        content:
          application/json:
            schema:
              type: object
              properties:
                data:
                  type: object
                  properties:
                    input:
                      type: string
                      description: The text to be evaluated for various safety indicators.
                      example: "I am a dangerous person who will be wreaking havoc upon the world!!!"
      responses:
        '200':
          description: Successful safety assessment
          content:
            application/json:
              schema:
                type: object
                properties:
                  fdl_harmful:
                    type: number
                    format: float
                    description: Likelihood score for harmful content.
                  fdl_violent:
                    type: number
                    format: float
                    description: Likelihood score for violent content.
                  fdl_unethical:
                    type: number
                    format: float
                    description: Likelihood score for unethical content.
                  fdl_illegal:
                    type: number
                    format: float
                    description: Likelihood score for illegal content.
                  fdl_sexual:
                    type: number
                    format: float
                    description: Likelihood score for sexual content.
                  fdl_racist:
                    type: number
                    format: float
                    description: Likelihood score for racist content.
                  fdl_jailbreaking:
                    type: number
                    format: float
                    description: Likelihood score for jailbreaking attempts (prompt manipulation).
                  fdl_harassing:
                    type: number
                    format: float
                    description: Likelihood score for harassing content.
                  fdl_hateful:
                    type: number
                    format: float
                    description: Likelihood score for hateful content.
                  fdl_sexist:
                    type: number
                    format: float
                    description: Likelihood score for sexist content.
        '400':
          description: Bad request (invalid input data)
        '401':
          description: Unauthorized (missing or invalid Bearer token)

components:
  securitySchemes:
    bearerAuth:
      type: http
      scheme: bearer
      bearerFormat: JWT

Faithfulness Model

Banner: The safety guardrails are restricted to a 3500 token length limit for context, and 350 token length limit for response for the free trial experience. To increase these limits, please contact sales.

This Fiddler Trust Model detects hallucinations by evaluating the accuracy and reliability of facts presented in AI-generated text responses in retrieval-augmented generation (RAG) contexts.

The model requires two inputs:

1. Response: the text generated by your generative application

2. Context Documents: the reference text that the application response must remain faithful to

The output is a single score (float) representing the factual consistency between the response and the provided context.

How to Use Thresholding

Our quick start examples use a default threshold that works well in a variety of applications. You may wish to set this higher or lower to allow you to tailor how strictly you filter content based on your unique requirements.

Faithfulness Model OpenAPI Spec

openapi: 3.0.3
info:
  title: Fiddler FTL Response Faithfulness
  version: 1.0.0
servers:
  - url: "https://{fiddler_endpoint}"
paths:
  /v3/guardrails/ftl-response-faithfulness:
    post:
      summary: Evaluate the faithfulness of a provided response against given context.
      operationId: evaluateFaithfulness
      security:
        - bearerAuth: []
      requestBody:
        required: true
        content:
          application/json:
            schema:
              type: object
              properties:
                data:
                  type: object
                  properties:
                    response:
                      type: string
                      description: The response text to be evaluated.
                      example: "The Yorkshire Terrier and the Cavalier King Charles Spaniel are both small breeds of companion dogs."
                    context:
                      type: string
                      description: The contextual text to compare against.
                      example: "The Yorkshire Terrier is a small dog breed... The Cavalier King Charles Spaniel is a small spaniel..."
      responses:
        '200':
          description: Successfully computed faithfulness score
          content:
            application/json:
              schema:
                type: object
                properties:
                  fdl_faithful_score:
                    type: number
                    format: float
                    description: A numerical measure indicating how faithful the response is to the given context.
        '400':
          description: Bad request (invalid payload or missing parameters)
        '401':
          description: Unauthorized (missing or invalid Bearer token)

components:
  securitySchemes:
    bearerAuth:
      type: http
      scheme: bearer
      bearerFormat: JWT

Fiddler Trust Service Error codes

Error Code	Reason	Resolution
400	Invalid Input	Adjust API input to follow above API specification.
401	Invalid Auth Token	The authentication token is invalid or expired. Please double-check your token if it is invalid, and contact [e-mail] if you’d like to extend your trial.
404	Invalid guardrail endpoint called	API called must be either ftl-safety (safety model) or ftl-response-faithfulness (faithfulness model).
413	Input token length exceeds API token size limits	The safety guardrail has a limit of 4096 tokens. The faithfulness guardrail has a limit of 3500 tokens for the context field, and 350 tokens for the response field. These limits are higher in our paid plans.
429	Rate Limits exceeded	The rate limits for the free trial guardrail experience is 2 requests per second, 70 requests per minute, and 200 requests per day. These limits are higher in our paid plans
500/503/504	Internal Server Error	We are experiencing some internal service errors. Please watch #fiddler-guardrails-support on Slack or contact technical support.