Evaluation

How Know Your AI evaluates and red-teams your AI products.

Know Your AI provides two powerful evaluation modes to red-team and assess your AI products. Whether you expose a model through an API or embed a chatbot on your website, Know Your AI can automatically test it for vulnerabilities, safety issues, and compliance gaps.

Evaluation modes

Model Evaluation (API Mode)

Connect your model’s API endpoint and run automated red-team testing with attack datasets.

Chatbot Evaluation (Website Mode)

Evaluate live chatbot websites using a browser control agent for end-to-end testing.

How evaluations work

Every evaluation is built on three core components:

1. Judgment Model — the LLM used to judge responses (e.g., gemini-2.0-flash, hydrox-firewall)
2. Judgment Prompt — the prompt that tells the judge how to score each response
3. Threshold — the pass/fail cutoff for vulnerability detection

Know Your AI follows a consistent pipeline so results are reproducible and auditable:

1. Compose the test

   Select datasets, configure the number of prompts, and choose your evaluation mode.

2. Send prompts

   Each prompt is sent to your product — via direct API call (Model Evaluation) or through a browser control agent (Chatbot Evaluation).

3. Judge responses

   The judgment model scores each response, producing: isVulnerable, confidenceScore, and judgeAnalysis.

4. Analyze compliance

   Responses are automatically analyzed for CCPA/CPRA compliance violations.

5. Store results

   Every run stores prompts, responses, scores, screenshots, and compliance evidence.

Running evaluations

You can run evaluations in three ways:

• Compose Evaluation — ad-hoc runs with selected datasets and configurable prompt counts
• Scheduled runs — cron-based scheduling (hourly, daily, weekly, monthly, or custom) with enable/disable
• Evaluation Market — pre-configured evaluation templates for Safety, Compliance, Quality, and Performance

Interpreting results

Each evaluation run produces:

• Security score — overall vulnerability percentage with animated chart
• Per-prompt results — pass/fail for each prompt with judge analysis
• Compliance report — CCPA/CPRA violation analysis with evidence
• Screenshots — browser screenshots captured during chatbot evaluations
• Execution console — real-time streaming logs of prompts, responses, and judgments

Run status lifecycle

Evaluation runs progress through these stages:

pending → queued → container_creating → running → completed or failed

Related docs

Model Evaluation

Learn about API-based red-team testing.

Chatbot Evaluation

Learn about browser-based chatbot testing.

Datasets

Browse attack datasets and upload your own.

Monitoring

Turn evaluations into ongoing health signals.