my-project-public

repository

loading code, commits, and activity

repositories

loading repo index

#1	/**
#2	* DeepSeek Inference Client
#3	*
#4	* Implements the InferenceClient interface using DeepSeek's OpenAI-compatible API.
#5	* Supports thinking mode via reasoning_effort and extra_body.
#6	* Replaces CLAWD Runtime's default inference when using DeepSeek models.
#7	*/
#8	export const DEEPSEEK_BASE_URL = "https://api.deepseek.com";
#9	export const DEEPSEEK_MODEL_PRO = "deepseek-v4-pro";
#10	export const DEEPSEEK_MODEL_FLASH = "deepseek-v4-flash";
#11	/**
#12	* Create a DeepSeek inference client that conforms to the
#13	* InferenceClient interface used throughout the automaton.
#14	*/
#15	export function createDeepSeekInferenceClient(options) {
#16	const { apiKey, baseUrl = DEEPSEEK_BASE_URL, defaultModel = DEEPSEEK_MODEL_PRO, maxTokens: defaultMaxTokens = 4096, flashModel = DEEPSEEK_MODEL_FLASH, proModel = DEEPSEEK_MODEL_PRO, } = options;
#17	let currentModel = defaultModel;
#18	let maxTokens = defaultMaxTokens;
#19	let lowCompute = false;
#20	const chat = async (messages, opts) => {
#21	const model = opts?.model \|\| currentModel;
#22	const tokenLimit = opts?.maxTokens \|\| maxTokens;
#23	const body = {
#24	model,
#25	messages: messages.map(formatMessage),
#26	max_tokens: tokenLimit,
#27	stream: false,
#28	};
#29	if (opts?.temperature !== undefined) {
#30	body.temperature = opts.temperature;
#31	}
#32	const tools = opts?.tools;
#33	if (tools && tools.length > 0) {
#34	body.tools = tools;
#35	body.tool_choice = "auto";
#36	}
#37	// DeepSeek thinking mode: enable reasoning with high effort
#38	body.reasoning_effort = "high";
#39	body.extra_body = { thinking: { type: "enabled" } };
#40	const resp = await fetch(`${baseUrl}/v1/chat/completions`, {
#41	method: "POST",
#42	headers: {
#43	"Content-Type": "application/json",
#44	Authorization: `Bearer ${apiKey}`,
#45	},
#46	body: JSON.stringify(body),
#47	});
#48	if (!resp.ok) {
#49	const text = await resp.text();
#50	throw new Error(`DeepSeek inference error: ${resp.status}: ${text}`);
#51	}
#52	const data = (await resp.json());
#53	const choice = data.choices?.[0];
#54	if (!choice) {
#55	throw new Error("No completion choice returned from DeepSeek inference");
#56	}
#57	const message = choice.message;
#58	const usage = {
#59	promptTokens: data.usage?.prompt_tokens \|\| 0,
#60	completionTokens: data.usage?.completion_tokens \|\| 0,
#61	totalTokens: data.usage?.total_tokens \|\| 0,
#62	};
#63	// Extract reasoning_content if available (DeepSeek thinking mode)
#64	const reasoningContent = message.reasoning_content \|\| "";
#65	const toolCalls = message.tool_calls?.map((tc) => ({
#66	id: tc.id,
#67	type: "function",
#68	function: {
#69	name: tc.function.name,
#70	arguments: tc.function.arguments,
#71	},
#72	}));
#73	// Prepend reasoning content to the message if present
#74	let content = message.content \|\| "";
#75	if (reasoningContent) {
#76	content = `[Thinking]\n${reasoningContent}\n\n[Response]\n${content}`;
#77	}
#78	return {
#79	id: data.id \|\| "",
#80	model: data.model \|\| model,
#81	message: {
#82	role: message.role,
#83	content,
#84	tool_calls: toolCalls,
#85	},
#86	toolCalls,
#87	usage,
#88	finishReason: choice.finish_reason \|\| "stop",
#89	};
#90	};
#91	const setLowComputeMode = (enabled) => {
#92	lowCompute = enabled;
#93	if (enabled) {
#94	currentModel = flashModel;
#95	maxTokens = 2048;
#96	}
#97	else {
#98	currentModel = defaultModel;
#99	maxTokens = defaultMaxTokens;
#100	}
#101	};
#102	const getDefaultModel = () => {
#103	return currentModel;
#104	};
#105	return {
#106	chat,
#107	setLowComputeMode,
#108	getDefaultModel,
#109	};
#110	}
#111	/**
#112	* Format a ChatMessage to the DeepSeek/OpenAI API format.
#113	*/
#114	function formatMessage(msg) {
#115	const formatted = {
#116	role: msg.role,
#117	content: msg.content,
#118	};
#119	if (msg.name)
#120	formatted.name = msg.name;
#121	if (msg.tool_calls)
#122	formatted.tool_calls = msg.tool_calls;
#123	if (msg.tool_call_id)
#124	formatted.tool_call_id = msg.tool_call_id;
#125	return formatted;
#126	}
#127	/**
#128	* DeepSeek cost estimation in cents per million tokens.
#129	*/
#130	export const DEEPSEEK_PRICING = {
#131	"deepseek-v4-pro": { input: 200, output: 800 },
#132	"deepseek-v4-flash": { input: 15, output: 60 },
#133	};
#134	/**
#135	* Estimate cost in cents for a DeepSeek inference call.
#136	*/
#137	export function estimateDeepSeekCostCents(usage, model) {
#138	const p = DEEPSEEK_PRICING[model] \|\| DEEPSEEK_PRICING["deepseek-v4-pro"];
#139	const inputCost = (usage.promptTokens / 1_000_000) * p.input;
#140	const outputCost = (usage.completionTokens / 1_000_000) * p.output;
#141	return Math.ceil(inputCost + outputCost);
#142	}
#143	//# sourceMappingURL=deepseek-inference.js.map

z6Mkq5mY3JWtxoxUobWcfNHm7AkRubgSWEZTkBVqZXJviFZ5/my-project-public