Regular agents will often just do stuff! Ours follows the law.
poetry shell && ./demo.sh
Install with poetry install
Run with poetry run python main.py
We're modifying the Zero-shot ReACT agent with Chain-of-Thought to also check for compliance.
The old agent:
- Observation
- Thought
- Action
Our agent:
- Observation
- Thought
- Proposed Action
- Compliance check
- Action or Exit
We have a vector database for retrieving relevant regulations and case law.
Compliance chain is an agent Chain-of-Thought reasoning for determining if an action is compliant.
v0
- Take in the action, embed it, and then look up the most relevant case law in FAISS.
- Based on the case law, decide whether the action is legal or not.
- If it is, do the action. If it's not, flag with the relevant case law.
This agent is not conversational, it merely implements the Langchain QADocumentRetrieval pattern.
It's packaged as a tool, so that a tool-enabled conversational agent or another agent can invoke it via prompt, rather than having to modify the underlying agent orchestration.
The Example Chat Agent is a conversational Zero-shot ReACT agent.
[] Composability [] Read-teaming