Add OCI LangChain wrapper, docs, and dependency integration

Wire oci_langchain wrapper using langchain-oci with OCI SDK client
construction, max_completion_tokens for OpenAI models, retry and
timeout support. Add OCI integration docs, test fixtures, and
declare uv extra conflicts for langchain-oci compatibility.

Author: fede-kamel

docs/source/build-workflows/llms/index.md

@@ -28,6 +28,7 @@ NVIDIA NeMo Agent Toolkit supports the following LLM providers:
 | [OpenAI](https://openai.com) | `openai` | OpenAI API |
 | [AWS Bedrock](https://aws.amazon.com/bedrock/) | `aws_bedrock` | AWS Bedrock API |
 | [Azure OpenAI](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/quickstart) | `azure_openai` | Azure OpenAI API |
+| [OCI Generative AI](https://docs.oracle.com/en-us/iaas/Content/generative-ai/home.htm) | `oci` | OCI Generative AI via the OCI SDK-backed `langchain-oci` integration |
 | [LiteLLM](https://github.com/BerriAI/litellm) | `litellm` | LiteLLM API |
 | [Hugging Face](https://huggingface.co) | `huggingface` | Hugging Face API |
 | [Hugging Face Inference](https://huggingface.co/docs/api-inference) | `huggingface_inference` | Hugging Face Inference API, Endpoints, and TGI |
@@ -52,6 +53,15 @@ llms:
   azure_openai_llm:
     _type: azure_openai
     azure_deployment: gpt-4o-mini
+  oci_llm:
+    _type: oci
+    model_name: nvidia/Llama-3.1-Nemotron-Nano-8B-v1
+    region: us-chicago-1
+    compartment_id: ocid1.compartment.oc1..example
+    auth_type: API_KEY
+    auth_profile: DEFAULT
+    auth_file_location: ~/.oci/config
+    provider: meta
   litellm_llm:
     _type: litellm
     model_name: gpt-4o
@@ -118,6 +128,39 @@ The AWS Bedrock LLM provider is defined by the {py:class}`~nat.llm.aws_bedrock_l
 * `credentials_profile_name` - The credentials profile name to use for the model
 * `max_retries` - The maximum number of retries for the request
 
+### OCI Generative AI
+
+You can use the following fields to configure the OCI Generative AI LLM provider:
+
+* `region` - OCI region for the Generative AI service (defaults to `us-chicago-1`). The service endpoint is derived automatically.
+* `endpoint` - Optional explicit endpoint URL. Overrides the region-derived endpoint when set.
+* `compartment_id` - The OCI compartment OCID used for inference requests
+* `auth_type` - OCI SDK auth mode such as `API_KEY`, `SECURITY_TOKEN`, `INSTANCE_PRINCIPAL`, or `RESOURCE_PRINCIPAL`
+* `auth_profile` - OCI config profile name for file-backed auth
+* `auth_file_location` - Path to the OCI config file
+* `provider` - Optional provider override such as `meta`, `google`, `cohere`, or `openai`
+
+The OCI Generative AI LLM provider is defined by the {py:class}`~nat.llm.oci_llm.OCIModelConfig` class.
+
+* `model_name` - The name of the model to use
+* `region` - OCI region (defaults to `us-chicago-1`). The endpoint is derived from `https://inference.generativeai.{region}.oci.oraclecloud.com`.
+* `endpoint` - Optional explicit endpoint URL. Overrides the region-derived endpoint.
+* `compartment_id` - OCI compartment OCID
+* `auth_type` - OCI SDK auth type
+* `auth_profile` - OCI profile name for file-backed auth
+* `auth_file_location` - Path to the OCI config file
+* `provider` - Optional OCI provider override such as `meta`, `google`, `cohere`, or `openai`
+* `temperature` - The temperature to use for the model
+* `top_p` - The top-p value to use for the model
+* `max_tokens` - The maximum number of tokens to generate
+* `seed` - The seed to use for the model
+* `max_retries` - The maximum number of retries for the request
+* `request_timeout` - HTTP request timeout in seconds
+
+:::{note}
+This provider targets OCI Generative AI through the OCI SDK-backed `langchain-oci` path and does not enable the Responses API.
+:::
+
 ### Azure OpenAI
 
 You can use the following environment variables to configure the Azure OpenAI LLM provider:

docs/source/components/integrations/index.md

@@ -23,4 +23,5 @@ limitations under the License.
 ./frameworks.md
 ./a2a.md
 AWS Bedrock <./integrating-aws-bedrock-models.md>
-```
+OCI Generative AI <./integrating-oci-generative-ai-models.md>
+```

docs/source/components/integrations/integrating-oci-generative-ai-models.md

@@ -0,0 +1,99 @@
+<!--
+SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+SPDX-License-Identifier: Apache-2.0
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# NVIDIA NeMo Agent Toolkit OCI Integration
+
+The NeMo Agent Toolkit supports integration with multiple [LLM](../../build-workflows/llms/index.md) providers, including OCI Generative AI. The `oci` provider uses OCI SDK authentication and is designed for OCI Generative AI model and endpoint access. For workflow parity with the AWS Bedrock path, the toolkit also includes a LangChain wrapper built on `langchain-oci`.
+
+To view the full list of supported LLM providers, run `nat info components -t llm_provider`.
+
+## Configuration
+
+### Prerequisites
+Before integrating OCI, ensure you have:
+
+- access to OCI Generative AI in the target region
+- a valid OCI auth method such as `API_KEY`, `SECURITY_TOKEN`, `INSTANCE_PRINCIPAL`, or `RESOURCE_PRINCIPAL`
+- the target compartment OCID
+- the target OCI region (defaults to `us-chicago-1`) or a custom endpoint URL
+
+Common deployment patterns include:
+
+- OCI Generative AI regional endpoints
+- custom OCI Generative AI endpoints
+- OCI-hosted inference for NVIDIA Nemotron used as a live integration target
+
+### Example Configuration
+Add the OCI LLM configuration to your workflow config file:
+
+```yaml
+llms:
+  oci_llm:
+    _type: oci
+    model_name: nvidia/Llama-3.1-Nemotron-Nano-8B-v1
+    region: us-chicago-1
+    compartment_id: ocid1.compartment.oc1..example
+    auth_type: API_KEY
+    auth_profile: DEFAULT
+    temperature: 0.0
+    max_tokens: 1024
+    top_p: 1.0
+    request_timeout: 60
+```
+
+### Configurable Options
+* `model_name`: The name of the OCI-hosted model to use (required)
+* `region`: OCI region for the Generative AI service (defaults to `us-chicago-1`). The service endpoint is derived automatically.
+* `endpoint`: Optional explicit service endpoint URL. Overrides the region-derived endpoint when set.
+* `compartment_id`: OCI compartment OCID
+* `auth_type`: OCI SDK auth type
+* `auth_profile`: OCI profile name for file-backed auth
+* `auth_file_location`: Path to the OCI config file
+* `provider`: Optional OCI provider override such as `meta`, `google`, `cohere`, or `openai`
+* `temperature`: Controls randomness in the output (0.0 to 1.0)
+* `max_tokens`: Maximum number of tokens to generate
+* `top_p`: Top-p sampling parameter (0.0 to 1.0)
+* `seed`: Optional random seed
+* `max_retries`: Maximum number of retries for the request
+* `request_timeout`: HTTP request timeout in seconds
+
+### Limitations
+* This provider targets OCI Generative AI through the OCI SDK-backed `langchain-oci` path.
+* The Responses API is not enabled for this provider in the current release.
+
+## Nemotron On OCI
+
+One strong OCI deployment pattern is NVIDIA Nemotron hosted on OCI and exposed through an OpenAI-compatible route. In that setup, the toolkit can validate live integration behavior against the OCI-hosted Nemotron endpoint while the official provider and LangChain wrapper cover the OCI Generative AI path.
+
+## Usage
+Reference the OCI LLM in your configuration:
+
+```yaml
+llms:
+  oci_llm:
+    _type: oci
+    model_name: nvidia/Llama-3.1-Nemotron-Nano-8B-v1
+    region: us-chicago-1
+    compartment_id: ocid1.compartment.oc1..example
+    auth_profile: DEFAULT
+```
+
+## Troubleshooting
+* `401 Unauthorized`: verify the OCI profile, signer, and IAM permissions for Generative AI.
+* `404 Not Found`: confirm the regional endpoint or custom endpoint URL is correct.
+* `Connection errors`: verify OCI networking and regional endpoint reachability.
+* `Tool calling issues`: verify the served model supports tool calling and that the serving stack is configured for it.

docs/source/conf.py

@@ -379,6 +379,8 @@ def _build_api_tree() -> Path:
         '/extend/custom-components/gated-fields.html',
     'extend/integrating-aws-bedrock-models':
         '/components/integrations/integrating-aws-bedrock-models.html',
+    'extend/integrating-oci-generative-ai-models':
+        '/components/integrations/integrating-oci-generative-ai-models.html',
     'extend/memory':
         '/extend/custom-components/memory.html',
     'extend/object-store':

docs/source/get-started/installation.md

@@ -27,6 +27,7 @@ The following [LLM](../build-workflows/llms/index.md) API providers are supporte
 - OpenAI
 - AWS Bedrock
 - Azure OpenAI
+- OCI Generative AI
 
 ## Packages
 

examples/frameworks/agno_personal_finance/pyproject.toml

@@ -34,7 +34,7 @@ classifiers = ["Programming Language :: Python"]
 [tool.setuptools_dynamic_dependencies]
 dependencies = [
   "nvidia-nat[agno,test] == {version}",
-  "openai~=1.106",
+  "openai>=1.106,<3.0.0",
 ]
 
 [tool.uv.sources]

examples/frameworks/multi_frameworks/pyproject.toml

@@ -38,7 +38,7 @@ dependencies = [
   "beautifulsoup4~=4.13",
   "markdown-it-py~=3.0",
   "nvidia-haystack~=0.3.0",
-  "openai~=1.106",
+  "openai>=1.106,<3.0.0",
 ]
 
 [tool.uv.sources]

packages/nvidia_nat_agno/pyproject.toml

@@ -57,7 +57,7 @@ dependencies = [
   "agno>=1.2.3,<2.0.0",
   "google-search-results>=2.4.2,<3.0.0",
   "litellm>=1.74,<1.82.7",  # pin to known-good version
-  "openai~=1.106",
+  "openai>=1.106,<3.0.0",
 ]
 
 [tool.setuptools_dynamic_dependencies.optional-dependencies]

packages/nvidia_nat_langchain/pyproject.toml

@@ -67,6 +67,7 @@ dependencies = [
   "langchain-litellm>=0.3.5,<1.0.0",
   "langchain-milvus>=0.3.3,<1.0.0",
   "langchain-nvidia-ai-endpoints>=1.0.2,<2.0.0",
+  "langchain-oci>=0.2.4,<1.0.0",
   "langchain-openai>=1.1.6,<2.0.0",
   "langchain-tavily>=0.2.16,<1.0.0",
   "langgraph>=1.0.5,<2.0.0",

packages/nvidia_nat_langchain/src/nat/plugins/langchain/llm.py

@@ -37,6 +37,7 @@
 from nat.llm.huggingface_llm import HuggingFaceConfig
 from nat.llm.litellm_llm import LiteLlmModelConfig
 from nat.llm.nim_llm import NIMModelConfig
+from nat.llm.oci_llm import OCIModelConfig
 from nat.llm.openai_llm import OpenAIModelConfig
 from nat.llm.utils.hooks import _create_metadata_injection_client
 from nat.llm.utils.thinking import BaseThinkingInjector
@@ -54,6 +55,12 @@
 ModelType = TypeVar("ModelType")
 
 
+def _get_langchain_oci_chat_model():
+    from langchain_oci import ChatOCIGenAI
+
+    return ChatOCIGenAI
+
+
 def _patch_llm_based_on_config(client: ModelType, llm_config: "LLMBaseConfig") -> ModelType:
 
     from langchain_core.language_models import LanguageModelInput
@@ -222,6 +229,57 @@ async def openai_langchain(llm_config: OpenAIModelConfig, _builder: Builder):
         yield _patch_llm_based_on_config(client, llm_config)
 
 
+@register_llm_client(config_type=OCIModelConfig, wrapper_type=LLMFrameworkEnum.LANGCHAIN)
+async def oci_langchain(llm_config: OCIModelConfig, _builder: Builder):
+    import oci
+    from langchain_oci.common.auth import create_oci_client_kwargs
+
+    validate_no_responses_api(llm_config, LLMFrameworkEnum.LANGCHAIN)
+
+    ChatOCIGenAI = _get_langchain_oci_chat_model()
+
+    model_kwargs: dict[str, Any] = {}
+    if llm_config.temperature is not None:
+        model_kwargs["temperature"] = llm_config.temperature
+    if llm_config.top_p is not None:
+        model_kwargs["top_p"] = llm_config.top_p
+    if llm_config.max_tokens is not None:
+        if llm_config.provider and llm_config.provider.lower() == "openai":
+            model_kwargs["max_completion_tokens"] = llm_config.max_tokens
+        else:
+            model_kwargs["max_tokens"] = llm_config.max_tokens
+    if llm_config.seed is not None:
+        model_kwargs["seed"] = llm_config.seed
+
+    client_kwargs = create_oci_client_kwargs(
+        auth_type=llm_config.auth_type,
+        service_endpoint=llm_config.endpoint,
+        auth_file_location=llm_config.auth_file_location,
+        auth_profile=llm_config.auth_profile,
+    )
+    client_kwargs["retry_strategy"] = oci.retry.RetryStrategyBuilder(
+        max_attempts=llm_config.max_retries + 1  # OCI SDK counts total attempts (initial + retries)
+    ).get_retry_strategy()
+    if llm_config.request_timeout is not None:
+        client_kwargs["timeout"] = (10, llm_config.request_timeout)
+    oci_client = oci.generative_ai_inference.GenerativeAiInferenceClient(**client_kwargs)
+
+    client = ChatOCIGenAI(
+        client=oci_client,
+        model_id=llm_config.model_name,
+        service_endpoint=llm_config.endpoint,
+        compartment_id=llm_config.compartment_id,
+        auth_type=llm_config.auth_type,
+        auth_profile=llm_config.auth_profile,
+        auth_file_location=llm_config.auth_file_location,
+        provider=llm_config.provider,
+        is_stream=getattr(llm_config, "stream", False),
+        model_kwargs=model_kwargs or None,
+    )
+
+    yield _patch_llm_based_on_config(client, llm_config)
+
+
 @register_llm_client(config_type=DynamoModelConfig, wrapper_type=LLMFrameworkEnum.LANGCHAIN)
 async def dynamo_langchain(llm_config: DynamoModelConfig, _builder: Builder):
     """