AIMon helps developers build, ship, and monitor LLM Apps more confidently and reliably with its state-of-the-art, multi-model system for detecting LLM quality issues. It helps seamlessly with both offline evaluations and continuous production monitoring. AIMon offers fast, reliable, and cost-effective hallucination detection. It also supports other important quality metrics such as completeness, conciseness, and toxicity. Read our blog post for more details.
✨ Join our community on Slack
The following is a list of quality metrics that are currently available and on our roadmap. Please reach out to express your interest in any of these.
Metric | Status |
---|---|
Model Hallucination (Passage and Sentence Level) | ✓ |
Completeness | ✓ |
Conciseness | ✓ |
Toxicity | ✓ |
Instruction Adherence | ✓ |
AIMon supports asynchronous instrumentation or synchronous detections for the metrics mentioned above. Use these steps to get started with using the AIMon SDK and the product.
- Step 1: Get access to the beta product by joining the waitlist on our website or by requesting it on Slack or sending an email to info@aimon.ai
- Step 2: Install the AIMon SDK by running
pip install aimon
in your terminal. - Step 3: Here is an example to instrument an LLM application synchronously using the AIMon decorator:
from aimon import Detect
detect = Detect(values_returned=['context', 'generated_text'], config={"hallucination": {"detector_name": "default"}})
@detect
def my_llm_app(context, query):
# my_llm_model is the function that generates text using the LLM model
generated_text = my_llm_model(context, query)
return context, generated_text
- Step 4: For an example of how to instrument an LLM application asynchronously using the SDK, please refer
analyze_prod
decorator. - Step 5: For an example of synchronous detections using the SDK, please refer to the sample streamlit application
To demonstrate the effectiveness of our system, we benchmarked it against popular industry benchmarks for the hallucination detection task. The table below shows our results.
A few key takeaways:
✅ AIMon is 10x cheaper than GPT-4 Turbo.
✅ AIMon is 4x faster than GPT-4 Turbo.
✅ AIMon provides the convenience of a fully hosted API that includes baked-in explainability.
✅ Support for a context length of up to 32,000 tokens (with plans to further expand this in the near future).
Overall, AIMon is 10 times cheaper, 4 times faster, and close to or even better than GPT-4 on the benchmarks making it a suitable choice for both offline and online detection of hallucinations.
Metric | Aimon Rely v1 | GPT-4 Turbo (LLM-as-a-judge) |
---|---|---|
Context Length | 32,000 | 128,000 |
TRUE Dataset Precision/Recall | 0.808 / 0.922 | 0.810 / 0.926 |
SummaC (test) Balanced Accuracy | 0.778 | 0.756 |
SummaC (test) AUC | 0.809 | 0.780 |
AnyScale Ranking Test for Hallucinations Accuracy | 0.665 | 0.741 |
AnyScale Ranking Test for Hallucinations Rel. Accuracy | 0.804 | 0.855 |
Avg. Latency | 417ms | 1800ms |
Cost (15M tokens across all benchmark datasets) excluding free tier | $15 | $158 |
Fully Hosted | ✅ | ✅ |
Explainability | Automatic sentence-level Scores | Detailed reasoning with additional prompt engineering |
There is a lack of industry-standard benchmark datasets for these metrics. We will be publishing an evaluation dataset soon. Stay Tuned! ⌛
Refer to the website aimon.ai for details.
Join our Slack community for the latest updates and discussions on generative AI reliability.