A properly responding demonstration still needs execution, data access, monitoring and boundaries to serve actual users. OCI Enterprise AI gathers capabilities to build that path.
Objectives: identify model, tools, loop and memory To distinguish chat, embed and re-run; following Responses API and function calling; and separate access controls from content controls.
Key concepts
A LLM-based agent combines model, tools, loop and memory. The model propose the following step. a tool gets data or runs an operation; The runtime coordinates decisions and results. memory retains relevant context. A chatbot can answer without acting and a workflow can follow a fixed sequence. An agent decides some of his steps dynamically.
The platform means three groups: models to infer, agents to coordinate actions; and governance to control access, conduct and operation. Using a managed runtime reduces operational work but does not remove your responsibility for permissions, tools or response quality.
Models, Recovery and Personalization
Chat generates responses. embed production of vectors; rerank Reorder candidates by relevance. In a search, you can recover an extensive set with embeddings, refine with reeling and pass the selected fragments to the chat model. A Rerank model doesn't look by itself at all your documents and doesn't write the final answer.
To prepare for recovery: Extract text, divide it into useful fragments, generate embeddings and preserve text, vector and metadata. In consultation, it combines authorized filters with similarity. The hybrid search can add lexic coincidence to exact code. It measures both recovery and response.
Adjust first instructions and examples. RAG provides an external knowledge to the context fine-tuning suits behavior. The material presents efficient methods as T-Few and LoRA as well as more extensive settings: compatibility depends on model and version. Check training format and reserve examples of evaluation. Do not remember a historic list of models as a permanent catalogue.
Tools, memory and controls
The OCI Responses API use a OpenAI-supported interface to invoke models and coordinate tools at OCI. That doesn't mean that SDK runs request in OpenAI's infrastructure: they matter endpoint and authentication. A project organize resources as responses, conversations, archives and vector stores. Query's requirements official guide of initiation.
In function callingThe model returns name and arguments. Your valid application, permissions and parameters, run the function and return the result associated with the call. The tools run, such as File Search or Code Interpreter, have another owner from execution. MCP standardizes connection with tools servers; It includes host, client and server and tools, resources and prompts. For remote servers Streamable HTTP is used; stdio corresponds to local stage. The remote server keeps applying its own controls.
The context of several shifts can be managed by client or service. previous_response_id links responses; compaction sums up history. The lasting memory should be associated with the right subject and project. It does not change the weights of LLM. The projects have retention settings: you must not claim that nothing is stored when you choose to retain conversations.
Authentication identifies the applicant; authorisation limits operations. IAM, network, encryption and data policies are complemented. A private endpoint does not replace IAM. The guarrails they detect risks as bad content, prompt injection or pII. detection and information do not amount to a blockade. Check what each mode applying to Captain's documentation.
Step-by-step example
One employee requests: "Query politics and open a ticket for my team."
- He's an addict to his employee and only regains his policy.
- The model explain the policy and propose a tool to create the ticket.
- The application valid equipment, identity, fields and permission of creation.
- Implement a single operation; if you re-try, avoid duplicating tickets with an operation identifier.
- It returns the actual number and retains only the context necessary for monitoring.
If a retrieved document requires "ignore permissions," it does not gain authority as a part of RAG. It limits steps, time and spending. stop or scale if you cannot confirm the result.
Common errors
- To state that declare a feature at JSON already runs its code.
- To merge a resources project with model weights.
- Retain memory from two users under the same identifier.
- Treat "report" as if it meant blocking a request.
- Promises that RAG, private network or guarrails remove all risks.
Security exercise
The agent can have tickets. A retrieved page asks you to close all of the organization's tickets. Do you have enough reading and training information?
Solution
No. Read does not allow to modify and the contents of a page do not give permissions. The application should reject an out-of-reach operation, check his or her identity and apply his or her business policy. The model can explain the limit and offer an allowed consultation. It should also avoid saving that instruction as a reliable memory.
Sources and review
Query Building AI Agents and OCI Enterprise AI block from MyLearning. To deepen later, continue with our guide to Agenic AI.