LLM Firewall Core Operation
This section describes the technology behind the core components of the LLM Firewall.
Named Entity Recognition (NER) Model
Our NER model uses the BERT Transformer-based architecture. The model is fine-tuned for specific entity recognition tasks. In non-technical terms, this model helps to find the nouns and tag it against a known classified category.
Examples:
Sample Sentence 1:
Elon Musk is the CEO of Tesla, and he visited New York last week.
NER Output
Elon Musk → Person (PER) Tesla → Organization (ORG) New York → Location (LOC)
Sample Sentence 2:
Amazon launched a new product in Seattle on June 15, 2023.
NER Output
Amazon → Organization (ORG) Seattle → Location (LOC) June 15, 2023 → Date (DATE)
The NER model performs the core task of identifying common entities like people, organizations, locations, and dates, etc. The core task helps to create filters to hide sensitive information.
Custom Entity Recognition (CEM)
The NER model helps to identify common entities. But ZKCrypt allows you to create customized agents for different use cases and environments. Our solutions allow you to create your own entity model offline and use it while interacting with AI apps.
Our CEM model helps you to create your custom entities. Here are some examples.
Example 1
Development head creates a troubleshooting agent that will be used by his team of developers. The agent must ensure that proprietary variables and functions are not exposed to Internet or LLMs.
Sample Sentence 1:
The algorithm for sorting the user_data_list is found in src/data-processing.py, which uses the quickSort() function and may need optimization.”
CEM Model Output:
user_data_list → Custom Entity: Variable Name (VARIABLE) src/data-processing.py → Custom Entity: File Path (FILE_PATH) quickSort() → Custom Entity: Function Name (FUNCTION)
