clj-oci.generative-ai — com.soundhound/clj-oci 1.11.0

OCI Generative AI Service Management API wrapper — Models + Imported Models.

Provides full CRUD + list + compartment-move on the two control-plane resource types of the Generative AI service:

• Model (BASE pretrained or CUSTOM fine-tuned): list-models, get-model, create-model, update-model, change-model-compartment, delete-model.

• ImportedModel (externally sourced from Object Storage or Hugging Face): list-imported-models, get-imported-model, create-imported-model, update-imported-model, change-imported-model-compartment, delete-imported-model.

Plus helpers for the polymorphic input slots; the OCI SDK exposes these as abstract base classes with multiple concrete subclasses, and oci-build does not auto-dispatch on JSON discriminators. Callers compose pre-built Java instances via the helpers below and pass them through:

• object-storage-dataset — Dataset → ObjectStorageDataset
• t-few-training-config — TrainingConfig → TFewTrainingConfig
• vanilla-training-config — TrainingConfig → VanillaTrainingConfig
• lora-training-config — TrainingConfig → LoraTrainingConfig
• object-storage-data-source — ModelDataSource → ObjectStorageObject
• hugging-face-data-source — ModelDataSource → HuggingFaceModel
• fine-tune-details — composes a dataset + cluster + config

The service is regional (12 endpoint regions). Operations route via clients/generativeai-client-for, which resolves the region from (:region opts) > the resource OCID's region segment > the singleton default — same precedence as Compute and FileStorage.

Async operations (create-model, delete-model, create-imported-model, update-imported-model, delete-imported-model) return an :opcWorkRequestId in :data. The Generative AI work-request polling API (GetWorkRequest/ListWorkRequests/etc.) is not wrapped in this slice; callers query externally if they need to wait.

Note: oci/http-request discards response headers, so this namespace accepts :if-match outbound but cannot surface etag or opc-request-id inbound — same limitation as compute.clj.

change-imported-model-compartment clj

(change-imported-model-compartment imported-model-id target-compartment-id)

(change-imported-model-compartment imported-model-id target-compartment-id opts)

Moves an ImportedModel to a different compartment. Synchronous (204).

Options: :if-match, :opc-retry-token, :region.

change-model-compartment clj

(change-model-compartment model-id target-compartment-id)

(change-model-compartment model-id target-compartment-id opts)

Moves a custom Model to a different compartment within the same tenancy. Synchronous (204 No Content from the API).

Options: :if-match, :opc-retry-token, :region.

create-imported-model clj

(create-imported-model details)

(create-imported-model details opts)

Imports a model from an Object Storage bucket or a Hugging Face repository.

Async — the API returns 201 with an opc-work-request-id; the :opcWorkRequestId is surfaced in :data.

details is a Clojure map; required keys are :compartment-id and :data-source (pre-built via object-storage-data-source or hugging-face-data-source).

Options: :region, :opc-retry-token.

create-model clj

(create-model details)

(create-model details opts)

Creates a custom Model by fine-tuning a BASE model with your own dataset.

Async — the API returns 201 with an opc-work-request-id header that tracks the fine-tune job. The :opcWorkRequestId is surfaced in :data.

details is a Clojure map; required keys are :compartment-id, :base-model-id, and :fine-tune-details (build via fine-tune-details, which composes a Dataset + DedicatedAiCluster + optional TrainingConfig).

Options: :region — region override; defaults to the singleton's region :opc-retry-token — idempotency token

Returns {:data {...new model... :opcWorkRequestId "..."}} or {:error ...}.

delete-imported-model clj

(delete-imported-model imported-model-id)

(delete-imported-model imported-model-id opts)

Deletes an ImportedModel. Cannot be deleted while active endpoints reference it (409 Conflict).

Async — 202 with opc-work-request-id surfaced in :data as :opcWorkRequestId.

Options: :if-match, :region.

delete-model clj

(delete-model model-id)

(delete-model model-id opts)

Deletes a custom Model. Cannot be deleted while active endpoints reference it (the API returns 409 Conflict).

Async — the API returns 202 with an opc-work-request-id; the :opcWorkRequestId is surfaced in :data.

Options: :if-match, :region.

fine-tune-details clj

(fine-tune-details {:keys [training-dataset dedicated-ai-cluster-id
                           training-config]})

Composes a FineTuneDetails for use as :fineTuneDetails on create-model.

Required: :training-dataset — pre-built Dataset (use object-storage-dataset) :dedicated-ai-cluster-id — OCID of the fine-tuning DedicatedAiCluster Optional: :training-config — pre-built TrainingConfig (use t-few-/vanilla-/lora-training-config)

get-imported-model clj

(get-imported-model imported-model-id)

(get-imported-model imported-model-id opts)

Fetches a single ImportedModel by OCID. Region is OCID-derived; pass :region in opts to override.

Returns {:data {...}} or {:error ...}.

get-model clj

(get-model model-id)

(get-model model-id opts)

Fetches a single Model by OCID (BASE or CUSTOM).

The region is resolved from the model OCID's region segment; pass :region "us-ashburn-1" in opts to override.

Returns {:data {...}} or {:error ...}.

hugging-face-data-source clj

(hugging-face-data-source params)

Builds a HuggingFaceModel for use as :data-source on create-imported-model.

Required: :modelId (the Hugging Face model identifier). Optional: :accessToken, :branch, :commit.

list-imported-models clj

(list-imported-models compartment-id)

(list-imported-models compartment-id opts)

Lists ImportedModels in compartment-id, paginating automatically.

Returns {:data {:items [...]}} or {:error ...}.

Options (all optional unless noted): :region — region override; defaults to the singleton's region :vendor — exact match :capability — vector of keywords from [:text-to-text :image-text-to-text :embedding :rerank :text-to-image] (or Java enum constants) :lifecycle-state — keyword (:active, :creating, :updating, :deleting, :deleted, :failed) or Java constant :display-name — exact match :id — imported-model OCID :sort-by — :display-name, :lifecycle-state, or :time-created :sort-order — :asc or :desc :limit — page size (default 100)

list-models clj

(list-models compartment-id)

(list-models compartment-id opts)

Lists Models in compartment-id, paginating automatically. Includes both BASE pretrained models and CUSTOM fine-tuned models in the compartment.

Returns {:data {:items [model-maps...]}} or {:error ...}. Each map has camelCase keys via from-java.

Options (all optional unless noted): :region — region-id (e.g. "us-ashburn-1") or 3-letter code ("iad"); defaults to the singleton's region :vendor — exact match on vendor string :capability — vector of keywords from [:text-generation :text-summarization :text-embeddings :fine-tune :chat :text-rerank :text-to-image] (or Java enum constants) :lifecycle-state — keyword (:active, :creating, :deleting, :deleted, :failed) or Java constant :display-name — exact match (case-sensitive) :id — model OCID :sort-by — :display-name, :lifecycle-state, or :time-created :sort-order — :asc or :desc :limit — page size (default 100)

lora-training-config clj

(lora-training-config params)

Builds a LoraTrainingConfig for use as :training-config on fine-tune-details. LoRA (low-rank adaptation) is a parameter-efficient fine-tuning method.

Adds :loraR, :loraAlpha, :loraDropout to the common training fields.

object-storage-data-source clj

(object-storage-data-source params)

Builds an ObjectStorageObject for use as :data-source on create-imported-model.

Required: :namespaceName, :bucketName, :prefixName. Optional: :region (string region-id, defaults to the tenancy's home region).

object-storage-dataset clj

(object-storage-dataset params)

Builds an ObjectStorageDataset for use as :training-dataset on fine-tune-details. The only concrete subclass of Dataset shipped by the SDK at present.

Required: :namespaceName, :bucketName, :objectName.

t-few-training-config clj

(t-few-training-config params)

Builds a TFewTrainingConfig for use as :training-config on fine-tune-details. T-Few is a parameter-efficient fine-tuning method.

All fields optional: :totalTrainingEpochs, :learningRate, :trainingBatchSize, :earlyStoppingPatience, :earlyStoppingThreshold, :logModelMetricsIntervalInSteps.

update-imported-model clj

(update-imported-model imported-model-id details)

(update-imported-model imported-model-id details opts)

Updates mutable fields on an ImportedModel.

Unlike update-model, this is async — the API returns 202 with an opc-work-request-id; the :opcWorkRequestId is surfaced in :data.

Mutable: :display-name, :description, :vendor, :version, :freeform-tags, :defined-tags.

Options: :if-match, :region.

update-model clj

(update-model model-id details)

(update-model model-id details opts)

Updates mutable fields on a custom Model. Synchronous — returns the updated Model body.

Mutable: :display-name, :description, :vendor, :version, :freeform-tags, :defined-tags.

Options: :region — region override; defaults to OCID-derived :if-match — etag for optimistic concurrency

Returns {:data {...}} or {:error ...}.

vanilla-training-config clj

(vanilla-training-config params)

Builds a VanillaTrainingConfig for use as :training-config on fine-tune-details. Vanilla is full-parameter fine-tuning.

Same field set as t-few-training-config.