Keysight Generator’s next step: self-adjusting AI for smarter test generation
AI can generate test cases from natural language – but doing that reliably, at scale, and across messy real-world requirements is much harder.
That’s why Generator doesn’t just “use an LLM”. It engineers the entire system around control, relevance, and adaptability. And with the latest updates, Generator now goes one step further – it adapts itself automatically to the quality of each requirement.
Why requirement quality matters
In the real world, requirements are rarely consistent. Some are clear and precise. Others are long, complex, or open to interpretation. Traditional AI systems treat them all the same, which often leads to:
- Irrelevant context being used
- Overly generic or incomplete test cases
- Extra rework to correct AI-generated output
Generator now solves this problem at the system level.
How Generator adapts automatically
When a requirement is submitted, Generator evaluates it across three dimensions:
- Syntax – how precise and objective the language is
- Semantics – how open the requirement is to multiple interpretations
- Complexity – how many conditions or logical steps it contains
Based on these scores, Generator categorises the requirement as simple, complex, or ambiguous. This classification then drives how Generator behaves behind the scenes.
For example:
- Simple requirements trigger tighter similarity thresholds and fewer retrieved contexts
- Ambiguous requirements widen retrieval and increase context diversity
- Generation parameters such as temperature and top-p are adjusted automatically
All of this happens without any user intervention. At the same time, Generator makes these scores visible to QA teams, providing feedback on syntax, semantics, and complexity so teams can assess and refine requirement quality before test execution begins.
Why this matters for output quality
Large language models are excellent conditional generators – but only when they receive the right information in the right way.
Generator is built around a retrieval-first RAG pipeline, where:
- Retrieval determines what the model knows
- Prompt structure determines how it reasons
- The model determines how the result is expressed
By dynamically tuning retrieval and generation per requirement, Generator ensures the model sees the most relevant domain context and produces structured, predictable outputs. This dramatically reduces hallucination and improves alignment with real application behavior.
Engineered for control and accessibility
Generator uses a quantised, local Mistral-7B model rather than a heavyweight cloud deployment. This means:
- Customers can run Generator on standard hardware
- Sensitive data stays local
- Performance remains consistent and predictable
Structured prompts and schema constraints enforce output formats, so generated test cases remain usable, reviewable, and automation-ready.
Avaialble in the latest Generator release
The adaptive requirement scoring and auto-tuning capabilities are available in the latest version of Generator. Once deployed or upgraded, teams can immediately take advantage of requirement quality assessment and self-adjusting AI behavior as part of their existing workflow.
The result is a system that delivers better test assets from imperfect inputs, reduces manual rework, and builds confidence in AI-assisted testing – without asking users to become AI experts themselves.
Generator doesn’t just generate tests. It understands the intent behind your requirements and adjusts accordingly.
If you have any questions about Keysight Generator or want to connect with one of our test automation experts, get in touch.