DeepSeek

OverviewCopy Anchor Link

DeepSeek's chat models, with control over their thinking mode.

• DeepSeekLLMService: an OpenAILLMService subclass with DeepSeek's base URL and default model. It uses DeepSeekLLMAdapter, which adds an empty reasoning_content to assistant messages that have none, which DeepSeek requires in thinking mode once a tool call is involved.

InstallationCopy Anchor Link

bun add @pipedog/deepseek

It depends on @pipedog/openai, which is installed with it.

import { DeepSeekLLMService } from "@pipedog/deepseek/llm";
import { DeepSeekLLMAdapter } from "@pipedog/deepseek/deepseek_adapter";

PrerequisitesCopy Anchor Link

• A DeepSeek account and an API key from the DeepSeek platform.
• Pass the key as apiKey. It is required, and the service never falls back to OPENAI_API_KEY. The examples read it from DEEPSEEK_API_KEY.

DeepSeekLLMServiceCopy Anchor Link

ConfigurationCopy Anchor Link

The other options (organization, project, defaultHeaders, serviceTier, retryOnTimeout, retryTimeoutSecs, function calling) are OpenAILLMService's. See OpenAI.

SettingsCopy Anchor Link

Runtime-updatable settings, passed as settings: new DeepSeekLLMService.Settings({...}) (DeepSeekLLMSettings). They are OpenAI's settings, with these defaults:

Every other setting keeps OpenAI's default. Unset sampling settings are left out of the request, so the provider's own defaults apply. extra carries provider-specific request fields.

UsageCopy Anchor Link

import { LLMUpdateSettingsFrame } from "@pipedog/core/frames/frames";
import { DeepSeekLLMService } from "@pipedog/deepseek/llm";

const llm = new DeepSeekLLMService({
  apiKey: process.env.DEEPSEEK_API_KEY ?? "",
  settings: new DeepSeekLLMService.Settings({
    model: "deepseek-flash",
    systemInstruction: "You are a helpful assistant in a voice conversation.",
    temperature: 0.7,
  }),
});

Thinking is off by default, because DeepSeek's V4 models otherwise reason before every answer, which delays the first spoken token. Turn it on with DeepSeekLLMService.ThinkingConfig:

const thinking = new DeepSeekLLMService({
  apiKey: process.env.DEEPSEEK_API_KEY ?? "",
  settings: new DeepSeekLLMService.Settings({
    thinking: new DeepSeekLLMService.ThinkingConfig({ type: "enabled" }),
  }),
});

Change settings mid-call with an LLMUpdateSettingsFrame. The next completion uses them:

// Queue it on the worker, or push it from a processor upstream of the service.
const update = new LLMUpdateSettingsFrame({
  delta: new DeepSeekLLMService.Settings({ temperature: 0.2, maxTokens: 300 }),
});

NotesCopy Anchor Link

• Runs a completion for each LLMContextFrame, between an LLMFullResponseStartFrame and an LLMFullResponseEndFrame, and streams the reply as LLMTextFrames, as OpenAILLMService does. Function calls, token usage metrics and runInference work the same way.
• "developer" messages are sent as "user" messages, since the API doesn't support that role (supportsDeveloperRole is false).
• Messages go through DeepSeekLLMAdapter, which adds an empty reasoning_content to assistant messages that have none, which DeepSeek requires in thinking mode once a tool call is involved.
• seed, max_completion_tokens and service_tier are not sent: DeepSeek doesn't support them.
• DeepSeek can't enforce a response schema (supportsResponseSchema is false), so runInference ignores responseSchema, with a warning.

Runtime notesCopy Anchor Link

Generated from pipecat.runtimes and pipecat.compat in packages/deepseek/package.json: ✅ means the compat smoke test passes on that runtime.

• The client is the official openai JS SDK (through @pipedog/openai), which talks through fetch. It works on Bun, Node, Deno and Cloudflare Workers.
• The deprecated model constructor option is not ported. Use settings.
• Tested against a local OpenAI-compatible mock server, not against DeepSeek itself.