DeepSeek General Dialogue Document

Overview

This document describes how to call the DeepSeek conversational model via the Nebula API's OpenAI-compatible interface.

Basic Information

Item	Content
Base URL	`https://llm.ai-nebula.com/v1/chat/completions`
Authentication	API Key (Token)
Request Header	`Authorization: Bearer sk-xxxx`, `Content-Type: application/json`

Supported Models (Examples)

deepseek-v3-1-250821

Other DeepSeek series models (subject to routing configuration)

API Endpoints

1. Minimal Example (Non-streaming)

2. Streaming SSE Example

3. Common Parameters

Sampling and Control: temperature, top_p, max_tokens, stop

Structured Output: response_format/json_schema

Tool Calling: tools/tool_choice (follows OpenAI compatible format)

Note: DeepSeek may support more features or differentiated fields across different channels. Nebula will try to pass through and standardize them in the compatibility layer. It is recommended to use only common fields or consult the channel support list.

4. Tool Calling (Functions / Tools)

Complete Tool Calling Process (Two Stages)

Stage 1: The model returns tool_calls (content is usually null, finish_reason=tool_calls). You need to execute the corresponding function on your server based on tool_calls[*].function.name/arguments.

Stage 2: Pass the tool execution result back to the model as a role:"tool" message and continue completion (can be streaming).

Non-streaming continuation example (Stage 2):

Streaming continuation example (Stage 2 also supports streaming):

Note:

tool_call_id must be consistent with the return from Stage 1.

When tool execution fails, readable error messages or degraded results should be returned to avoid blocking subsequent completion.

DeepSeek's support for tool calling may vary by model version. It is recommended to confirm channel support before use.

5. Thinking Capability (Thinking)

DeepSeek supports enabling/disabling thinking capability with the thinking field. Disabled by default:

thinking={
 "type": "disabled" // Default behavior: disable thinking capability
 // "type": "enabled" // Enable thinking capability
}

In OpenAI-compatible requests, you can pass the thinking field directly at the top level:

Explanation:

Different models/versions may output thinking capability differently (e.g., whether to return an explicit reasoning field or only reflect it in the content structure).

If you need to see the thinking process intuitively in the terminal and the channel returns streaming, you can combine it with stream: true for a better interactive experience.

Response and Usage

Non-streaming: Returns choices and usage all at once.

Streaming: SSE chunks return, possibly containing usage at the end; if the channel supports stream_options.include_usage=true, real-time usage may be returned within chunks.

Frequently Asked Questions (FAQ)

Compatibility with OpenAI?

Uses OpenAI Chat Completions format; a few extension fields may not take effect, subject to channel support.

Is structured output supported?

Supports response_format: json_schema; for complex Schemas, it is recommended to lower temperature to improve consistency.

Are Chain of Thought / Search switches supported?

Depends on channel and model version; for exclusive capabilities, contact admin for activation or pass through parameters (if channel supports).

Is tool calling supported?

Supports OpenAI-compatible tool calling format; specific support levels may vary by model version and channel.

Best Practices

Use streaming to improve time-to-first-token and interactive experience.

Lower temperature and control max_tokens for strictly structured output.

Implement fault tolerance, retry, and timeout control for tool calling results.

Overview#

Basic Information#

Supported Models (Examples)#

API Endpoints#

1. Minimal Example (Non-streaming)#

2. Streaming SSE Example#

3. Common Parameters#

4. Tool Calling (Functions / Tools)#

Complete Tool Calling Process (Two Stages)#

5. Thinking Capability (Thinking)#

Response and Usage#

Frequently Asked Questions (FAQ)#

Best Practices#