Safety and Troubleshooting

GenerativeAgent prioritizes safety in its development. ASAPP ensures accuracy and quality through rigorous testing, continuous updates, and advanced validation to prevent hallucinations. Our team has incorporated Safety Layers that provide benefits such as reliability and response trust. You can take steps to align GenerativeAgent with your organization’s goals.

​

Safety Layers

GenerativeAgent uses a four-layer safety strategy to prevent irrelevant or harmful responses to customer support queries. The layers also prevent any type of hallucination response from the GenerativeAgent.

GenerativeAgent’s Safety Layers work as follows:

• Scope: The Scope layer halts any request that is outside of the reach or the context of GenerativeAgent.
• Input safety: This layer defends against any nefarious prompt attempt from a user.
• Core planning loop: This layer is where the GenerativeAgent does all of its magic (handling tasks, calling APIs, researching Knowledge Bases) while also restraining from performing any task or sending any reply that’s either out of scope, contrary to the desired tone of voice, or that goes against any of your organization policies.
• Output safety: This layer defines the response given by GenerativeAgent and assures that any reply protects customer and organization data.

​

Input Safety

ASAPP’s safety bot protects against manipulation, prompt injection, bad API responses, code/encryption, leakings, and toxic safety risks.

Customers can also configure examples that should be classified as safe to improve model accuracy.

By default, GenerativeAgent’s in scope capabilities prevent customers from engaging with GenerativeAgent on topics outside of your organization’s matters. You can configure topics that GenerativeAgent must not engage.

​

Output Safety

ASAPP’s output bot ensure any output is safe for your organization.

Our TaskBot prompts customers to confirm any action before GenerativeAgent calls identified APIs so the Agent is prevented from performing any action that might impact your organization.

​

Ongoing Evaluations

ASAPP runs red team simulations on a periodic basis. This way, we ensure our systems and GenerativeAgents are protected from any type of exploitation or leaks.

These simulations include everything from security exploits to prompts or tasks that might impact your organization in an unintended manner.

Evaluation Solutions

ASAPP implements automated tests designed to define the performance and functionality of GenerativeAgent. The Tests simulate a wide range of scenarios to evaluate GenerativeAgent’s responses.

​

Knowledge Base and APIs

GenerativeAgent’s responses are grounded on Knowledge Base articles and APIs to construct reliable responses. It is important to set up these two factors correctly to prevent any type of hallucination. Our tests comprise:

• Measurement: ASAPP continuously tracks a combination of deterministic metrics and AI-driven evaluators for conversations in production.
• Human Oversight: ASAPP’s Team actively monitors conversation to ensure accurate and relevant responses.

​

Data Security

ASAPP’s security protocols protect data at each point of transmission, from first user authentication to secure communications, to our auditing and logging system, all the way to securing the environment when data is at rest in data storage.

Additionally, ASAPP’s API gateway solution provides rate limiting, input validation and protects endpoints against direct access, injections, and other attacks.

Access to data by ASAPP teams is tightly constrained and monitored. Strict security protocols protect both ASAPP and our customers.

ASAPP utilizes custom redaction logic to remove sensitive data elements from conversations in real-time.