OCI Enterprise AI Agents

OCI has introduced Enterprise AI Agents in OCI Generative AI as a managed way to build, run, and host agentic applications. Generative AI adoption is moving from simple prompt based experimentation to enterprise ready AI assistants that can answer questions using approved business knowledge. In this article we are going to create a knowledge base using business documents and then create OCI Gen AI agents followed by a chatbot. Here we achieve in getting answers from our own documents, policies, architecture notes, runbooks and internal knowledge. The goal of this article is to show how OCI can be used to build a governed enterprise knowledge assistant without writing application code.

1) Create OCI Object storage bucket

2) Upload the documents into Object storage bucket

3) Create a Knowledge base

4) Create an Agent

6) Create a RAG tool

7) Launch Chat and Query.

The assistant will use a set of documents uploaded into OCI Object Storage. These documents are added to a knowledge base. The knowledge base is then connected to an OCI Generative AI Agent through a RAG tool. When we raise prompt in chat, the agent retrieves relevant information from the knowledge base and generates a response using the selected LLM.

Create OCI Object storage bucket

In OCI Console:

Navigation menu → Storage → Object Storage & Archive Storage → Buckets

Create the bucket with default settings.

Once bucket is created, Upload documents to Object storage. These documents may be PDFs, text files, architecture notes, or other supported file types. The documents stored in the bucket become the source content for the knowledge base.

Create a Knowledge Base

As the next step create a Knowledge Base, in OCI Console:

Navigation menu → Analytics & AI → Generative AI Agents

A knowledge base is the collection of source data that the agent can use for retrieval. In this article, the knowledge base points to the documents stored in OCI Object Storage. The knowledge base does not answer questions directly. Instead, it provides the searchable enterprise context that the RAG tool can retrieve from when the user asks a question.

In the OCI console, we could see three options that we discussed earlier. As first step, create a Knowledge base and then Agents using that Knowledge base and then initiate Chat.

For Knowledge base creation specify the Bucket having the documents.

Either all documents in the bucket can be chosen or any specific documents.

Once documents are chosen click Create button to create the knowledge base.

The knowledge base creation will take few minutes to complete. We need to be patient. We can’t alter or delete the creation until it completes.

Once Knowledge Base is created move to creating Agent.

Create Agent

An agent is the main reasoning layer. It receives our prompts and decides how to use the available tool and generates the final response. The agent can be guided using routing instructions so that it responds in a specific style and follows enterprise rules.

To create Agent, specify Name and then Routing instructions. Instructions required to specify how to respond to chat prompts. In this article, we instruct the agent to answer only from the knowledge base, avoid guessing, provide concise responses, and escalate security or production-impacting decisions to a human owner.

In this exercise we choose detail Routing LLM type Llama.

After Basic information, the next step is Add RAG tool. A RAG tool allows the agent to search the knowledge base, retrieve relevant content and then use that retrieved content to generate a better answer. This is important because it helps the agent answer based on approved documents instead of relying only on the model’s general knowledge.

We need to specify name and Routing description. The description specifies how the tool should work with documents.

In the Add Tool we go with Default LLM (Llama)

The next step is setting up Agent endpoint. An agent endpoint is the access point used to interact with the deployed agent. Once the agent and its tools are configured, the endpoint allows users or applications to send questions to the agent. Guardrails are safety and compliance controls for the AI interaction. They help reduce risk by applying checks such as content moderation, prompt injection protection, and personally identifiable information protection. We will go with default settings for Guardrails.