Baidu Qianfan SDK

The Qianfan SDK provides access to several categories of capabilities on the Baidu Intelligent Cloud Qianfan platform:

• Model Capabilities (Inference): Chat, Completion, Embedding, Plugin calling, and Text2Image.
• Model Fine-tuning: Training and optimizing models.
• Model Management: Managing model lifecycle and services.
• Dataset Management: Handling training and evaluation datasets.
• Prompt Management: Managing prompt templates.
• Utilities: Tokenizer and interface flow control.

The Qianfan SDK provides a best-practice toolchain for the Baidu Qianfan Large Model Platform, enabling seamless access for AI workflows and AI-native applications. The SDK's core capabilities are divided into three main areas:

• Large Model Inference: Encapsulates interfaces for the ERNIE (一言) series and various open-source models. It supports tasks such as Chat, Completion, and Embedding.
• Large Model Training: Supports end-to-end training processes, including training data management, fine-tuning/pre-training, and model hosting.
• General & Extensions: Includes common AI development tools like Prompt, Debug, and Client, as well as extensions that adapt to common middleware frameworks.

You can manage your prompts using the following operations:

• Create, update, or delete a Prompt.
• Get Prompt details.
• Get a list of Prompts.
• Get a list of Prompt tags.

The Qianfan platform supports training and fine-tuning capabilities for large models. You can manage the lifecycle of a training job through the following operations:

• Create a training task.
• Create a task run.
• Get details of a task run.
• Stop a running task.

The Qianfan platform provides comprehensive APIs for dataset lifecycle management, including:

• Creation & Deletion: Create and delete datasets.
• Data Movement: Initiate dataset publishing, importing, and exporting tasks; get export records and import error details.
• Status: Get dataset details and status details.
• Data Processing: Create and manage data cleaning tasks (view list, view details, delete) and data augmentation tasks (view list, view details, delete).
• Annotation: Entity annotation, deleting entities, and retrieving entity lists.

You can deploy and manage models by publishing them as services. Supported operations include:

• Create a service.
• Query service details.

The Qianfan SDK provides access to several core large model prediction capabilities. For detailed implementation and usage, refer to the inference documentation. Supported capabilities include:

• Chat: Conversational AI interactions.
• Completion: Text completion/continuation.
• Embedding: Converting text into vector representations.
• Plugin: Calling external plugins.
• Text-to-Image: Generating images from text descriptions.

Once training is complete, you can manage your models using the following operations:

• Get model details.
• Get model version details.
• Publish a training task as a model.
• Initiate a model evaluation task.
• View model evaluation details.
• View model evaluation reports.
• Stop a model evaluation task.

The qianfan SDK provides a Trainer to manage the complete fine-tuning lifecycle. This includes:

• Loading local datasets via Dataset and uploading them to the Qianfan platform.
• Fine-tuning specific models (e.g., ERNIE-Speed-8K).
• Running batch evaluations on the fine-tuned model.
• Publishing the model as a service.

The SDK also supports Event Handlers for monitoring training and Resumability, allowing you to register an EventHandler and resume a training task from a failure point.

The Qianfan SDK provides extensions to integrate Baidu Qianfan's large model capabilities into the Microsoft Semantic Kernel (SK) framework. This allows developers to use Qianfan models as services within the SK ecosystem.

Supported integration types include:

• QianfanChatCompletion: For text dialogue/chat.
• QianfanTextCompletion: For text completion/continuation.
• QianfanTextEmbedding: For text vectorization.

If you cannot obtain an IAM Access Key/Secret Key, you can use Application-level API Key (AK) and Secret Key (SK).

Warning: This method is considered legacy and will be removed from the SDK in the future. It is also limited in scope: it cannot be used for training, model management, or dataset management functions.

Set the following environment variables for this method:

• QIANFAN_AK
• QIANFAN_SK

import os
import qianfan

os.environ["QIANFAN_AK"] = "..."
os.environ["QIANFAN_SK"] = "..."

# Only supports inference-related functions
chat = qianfan.ChatCompletion()
resp = chat.do(messages=[{"role": "user", "content": "你好，千帆"}])

print(resp["result"])