OCI Generative AI allows to incorporate managed models into an application. Inference service, a custom model, vector search and a SQL consultation have different responsibilities.
Objectives: To recognize Playground, models and endpoints; To distinguish training from hosting; relate AI Vector Search with RAG; and explain how Select AI uses natural language about structured data.
Key concepts
A model of chat generates an answer. a model embedding turns content into a numerical representation. The families, versions, regions and modalities admitted change: check the available catalogue for use. The old names that appear at a demonstration are not an exhaustive list today.
The Playground allows to explore prompts and parameters without starting with code. When a test is useful, you integrate it with the API and the corresponding SDK. A console test does not automatically display the complete application.
Reforms one at a time. Temperature influence the distribution of sampling; top-p limits candidates by accumulated probability; top-kWhen available, by quantity. The output limit restricts length and does not restrict accuracy. A preamble guides conduct without re-training pesos.
Personalization and dedicated clusters
The service offers pre-trained models and customization options for supported models. In the fine-tuning flow presented by the course, you prepare examples, the Object Storage preserves and use compute from a Dedicated AI Cluster To train. The result is a custom model. A resource hosting and a endpoint serve inferences.
Don't confuse his Fine-tuning cluster with hosting. One adjusts the model and the other attends requests. Efficient methods such as T-Few or LoRA update a part of mandateable parameters. their availability depends on the family. Check formats, permissions, evaluation, capacity and costs before performing the lab. The exercises of this guide do not require them to be put into place.
AI Vector Search and Select AI
AI Vector Search embeddings storage in type columns VECTOR with relative data. A distance function allows you to look for similarities and to combine them with business filters. The embeddings can be generated with a ONNX model supported within the base or with an outside service as set up.
Select AI converts questions into natural language into operations about structured data using a AI profile, scheme context and a LLM set up. Can give back SQL, results or a narrative as an action. The consultation generated should respect permissions and be checked. That's not a vector search by definition.
Examples: "He looks for instructions similar to this malfunction" points to semantic retrieval "Add September's repairs by headquarters" needs a proper consultation and aggregation. To write an answer with documents, RAG adds the generation after recovering. The guy. VECTOR by itself does not write or allow access.
Step-by-step example
A assistant should explain a repair policy and have a look at the number of open orders.
- Test the response format at Playground with a fictitious policy and with no personal data.
- Prepare authorized documents, fragment and generate embeddings that are compatible with consultations.
- Reforms the relevant fragment and cites the version used.
- To count orders, use a controlled consultation or Select AI with approved tables inspect SQL, filters and results.
- He gives both results to the chat model, distinguishing documentary policy with computed figure.
- It evaluates ambiguous questions, absent sources and denial of permissions before integrating the application.
Do not use fine-tuning to memorize the number of commands: will change. Also do not assume that the LLM receives provider or retains the same in any configuration. Check the actual data flow.
Common errors
- To merge Object Storage with the compute that trains the model.
- Treat an endpoint as the training dataset.
- To say that Select AI knows all tables without a proper context and permissions.
- Compare vectors generated in incompatible spaces.
- To interpret the target accuracy of an approximate index as an exact guarantee for each consultation.
Design exercise
A user asks "How many orders from my headquarters are outstanding?" and then "What does the policy say about delays?" What two data routes does assistant require?
Solution
The first requires structured consultation with headquarters filter and state, run with permission. The second needs to get back and use the current policy as evidence. The two can be combined with an answer, but you cannot replace an exact count with vector similarity or assume that recovering a document gives access to all orders.
Sources and review
Query OCI Generative AI, AI Vector Search y Select AI. The material uses both Oracle Database 23ai and Oracle AI Database 26ai. check documentation and availability of your version.