Agent Requirements

Agent Objective:

• answers questions over large enetrprise knowledge base
• uses RAG to ingest and retrieve knowledge
• must give grounded and relevant answers, no hallucinations

Autonomy & Constraints:

• must only give answers based on the input knowledge
• no hallucinations
• must have human-in-loop checks before data is ingested. only checked data must be ingested

Tool & Action Design

• tool for ingesting data
• tool for retrieval

Agent Architecture

• single agent
• tool for ingestion
• tool for retrieval

Memory & Context Design

• uses RAG - vector store for storing embeddings
• short-term memory for now using session id for a chat session

Core Reasoning

• for prompt engineering, the agent will act as an expert in the enterprise knowledge
• the agent will just RAG retrieval when answering questions
• if agent cannot answer then it must say "I do not know"

Safety & Permissions

• human-in-loop during ingestion
• grounding the answers using evals
• rate limiting so that we dont overwhelm the api
• llm as a judge to handle adversarial inputs