Loading...
Agents needs to give accurate and helpful answers for every user in the Enterprise, based on a shared knowledge base of documents.
The answers should arrive in seconds, and have a very high bar for accuracy. We should cite every answer with the source document, and also check for hallucinations (inline or before returning). We must be honest with the user if the source documents are unclear or the question cannot be answered.
The agent also "reasons" or "thinks" before returning a final response. This happens after the relevant documents have been retrieved.
The agent should also have a high bar for retrieving relevant documents. It should intelligently match the user's queries to semantically relevant documents, do the user does not receive incomplete answers.
This agent does not call any other tools other than reading the knowledge base.
It does not write to any data source, or perform any other action.
It's sole purpose is to write back helpful answers in its response.
It may also serves images, if found in the knowledge base.
We must also consider what happens when the source documents have conflicting answers about the same question. We should surface this to the user, so they can verify this contradiction themselves. We should also notify the user if there are too many source documents to view - in which case, we must also state that the agent is at its context limit. (Or additionally, we must summarize the overall conversational context before proceeding). In this case, we return the summarized responses, and "fuzzy" cite the source documents.)
To track metrics, of how useful our agent is.
We should track these metrics:
Documents retrieved per query.
Times we returned no source documents. (Assess TP/FP)
User feedback loop (thumbs up/down) to assess agent quality.
Agent response latency. p50, p90, p99.
Agent Time to First Byte (Aim for below 2 seconds)
Agent Latency after document retrieval + TTFB
Resolution rate of the user query (again if they can escalate to a human through some mechanism).
Hallucination check - a post processor that checks for the answer accuracy or faithfulness based on the retrieved documents.
Multi document - we can have track when a query retrieved documents from multiple sources, and then link what the user's approval rating was of the answer, as well as the hallucination check.
Tools it has access to:
Query the Vector DB
Design the overall agent architecture. Choose an orchestration pattern (single agent, multi-agent, hierarchical). Identify key components and how they interact. Use the diagramming tool to illustrate the architecture.
We need to have a Vector DB.
We need to have a Re Ranker.
We need to have a Semantic Cache for repeated questions.
We need to have a tool to make requests to the Vector DB - and return documents (or document IDs).
We need to have a tool that performs a hallucination check (external GR or post processing for every answer).
We can also have a citation tool. that assembles all citations based on the Retrieved data.
How does the agent manage memory and context? Define short-term (conversation) and long-term (persistent) memory strategies. How is context window managed? What retrieval mechanisms are used (RAG, vector stores, summarization)?
Deep dive into the agent's reasoning strategy. How does it plan, decide, and self-correct? Discuss prompt engineering, chain-of-thought, reflection loops, failure handling, and tradeoffs.
What guardrails and safety mechanisms does the agent have? Define permission boundaries, human-in-the-loop checkpoints, content filtering, rate limiting, and how the agent handles adversarial inputs or unexpected states.