aac-req-qa

AaC-Req-QA

An AaC plugin to perform automated quality checks for the shall statements in your AaC model.

This plugin will scan your architecture for any req entries and will use an LLM to evaluate the quality of the shall field. The plugin is configured to evaluate your shall statement using the following attributes:

• Unambiguous: The requirement should be simple, direct, and precise with no room for interpretation.
• Testable (verifiable): The requirement should be testable, and it should be possible to verify that the system meets the requirement. Preferable the requirement should be verifiable by automated acceptance test, automated analysis, or demonstration rather than inspection. If inspection is the only rational means of verification it will have a lower rating.
• Clear: The requirement should be concise, terse, simple, and precise.
• Correct: The requirement should not have any false or invalid assertions.
• Understandable: The requirement should be easily understood by all stakeholders.
• Feasible: The requirement should be realistic and achievable.
• Independent: The requirement should stand-alone and not be dependent on other requirements.
• Atomic: The requirement should be a single, discrete, and indivisible statement.
• Necessary: The requirement should be necessary to the solution and not be redundant or superfluous.
• Implementation-free: The requirement should not specify how the solution will be implemented. It should only specify what the solution should do, not how it should do it.

If the shall is evaluated to be of sufficient quality, the aac check will pass. Otherwise, you will receive a failure message produced by the AI with an assessment of each attribute and an overall score. Failure results from an overall score of C (Medium) or lower from the AI.

Usage

If you haven't already, install Architecture-as-Code (AaC):

pip install aac

Next install this AaC-Req-QA plugin:

pip install aac-req-qa

Set the environment variables needed to access an OpenAI endpoint. This may be a commercial endpoint in OpenAI or Azure OpenAI or a self-hosted endpoint using a tool like Ollama or vLLM.

• AAC_AI_URL: The usl of the LLM endpoint. Example: https://localhost:11434/v1
• AAC_AI_MODEL: The name of the LLM model. Example: mistral for local (i.e. Ollama), gpt-4 for OpenAI or Azure
• AAC_AI_KEY: The access key for the API. If using a local model, any value will work but it must not be empty or missing. Example: not-a-real-key

If you wish to use an Azure OpenAI set the following environment variables.

• AAC_AI_TYPE: Set to "Azure", otherwise standard OpenAI client will be used.
• AAC_AI_API_VERSION: Set to your desired / supported API Version. Example: 2023-12-01-preview

If you have a proxy, set the proxy environment variables.

• AAC_HTTP_PROXY: The HTTP proxy. Example: http://userid:[e-mail]:80
• AAC_HTTPS_PROXY: The HTTPS proxy. Example: https://userid:[e-mail]:443
• AAC_SSL_VERIFY: Allows you to disable SSL verification. Value must be true or false (default: true).

Although this is a bit cumbersome, it is necessary as there is no other way to provide configuration data within AaC, particularly for constraint plugins. Remember to protect your secrets when configuring these environment variables.

Eval-Req Command

This plugin provides a new command called eval-req that will execute the requirements QA on a specified AaC file. This will perform an evaluation of each req in the AaC file based on INCOSE requirements quality guidelines, and will give you all the AI output for each requirement. Be aware, this performs a requirement-by-requirement evaluation with no context of surrounding requirements in the specification. It is often very difficult to meet all INCOSE guidelines in a single requirement statement. It will also perform separate AI evaluation calls for each requirement which can take a lot of time. If you wish to perform fewer AI calls and evaluate your requirements as a set, use the eval-spec command instead.

Eval-Spec Command

This plugin provides a new command called eval-spec that will execute the requirements specification QA on a specified AaC file. This will perform a wholistic review of the requirement specification and all the included requirements against the INCOSE requirements quality guidelines. For instances where the quality of requirements need to be assessed in context, this is a good solution.

Caveat

Because this is using an LLM, it is a non-deterministic process and cannot be guaranteed to perform consistently. The LLM is tuned to reduce variation and provide reliable, repeatable performance to the greatest extent possible, but no guarantees can be made with the current state-of-the art LLM models.

Performance is completely dependent on the performance of the LLM provided by the endpoint. This has been tested with Azure OpenAI using GPT-4 as well as Mistral 7B run within Ollama and had acceptable performance in both. Performance with other models may be better or worse.

Attribution

We're adapting the analyze claims pattern from the open source Fabric project to evaluate requirements. Huge thanks to the Fabric team for the innovation and examples.