Browse Source
Add new features to the xllm package, including the ability to enable a thinking mode and specify reasoning effort for DeepSeek-compatible providers. Update the environment configuration and demo implementations to showcase these features. Enhance the README and documentation to reflect the new functionality and usage examples.main
14 changed files with 349 additions and 15 deletions
@ -0,0 +1,76 @@ |
|||
import type { XMessage, XProviderName, XReasoningEffortInput } from "@dm/xllm"; |
|||
import { xllmReasoningEffort, xllmThinkingType } from "../env"; |
|||
import type { DemoLog, XllmClient } from "./types"; |
|||
|
|||
function thinkingRequestOptions(provider: XProviderName): { |
|||
thinking?: { type: "enabled" | "disabled" }; |
|||
reasoningEffort?: XReasoningEffortInput; |
|||
} { |
|||
if (provider !== "deepseek") { |
|||
return {}; |
|||
} |
|||
return { |
|||
thinking: { type: xllmThinkingType }, |
|||
reasoningEffort: xllmReasoningEffort ?? "high", |
|||
}; |
|||
} |
|||
|
|||
export async function runThinkingDemo( |
|||
client: XllmClient, |
|||
log: DemoLog, |
|||
ctx: { provider: XProviderName; model: string }, |
|||
): Promise<void> { |
|||
log("========== DeepSeek 思考模式（thinking + reasoning_effort）=========="); |
|||
const extra = thinkingRequestOptions(ctx.provider); |
|||
if (Object.keys(extra).length === 0) { |
|||
log( |
|||
`[thinking] 当前默认 provider 为「${ctx.provider}」，非 deepseek 时不自动附带 thinking 参数（避免部分网关拒绝未知字段）。将 VITE_XLLM_PROVIDER 设为 deepseek 后再试，或只看下方说明。`, |
|||
); |
|||
log("[thinking] 库内字段：thinking + reasoningEffort → 请求体 thinking、reasoning_effort；流式事件 reasoning.delta。"); |
|||
return; |
|||
} |
|||
|
|||
log( |
|||
`[thinking] 请求参数: thinking=${JSON.stringify(extra.thinking)}, reasoningEffort=${extra.reasoningEffort ?? "high"}`, |
|||
); |
|||
|
|||
let reasoning = ""; |
|||
let text = ""; |
|||
log("--- stream() ---"); |
|||
const messages = [ |
|||
{ role: "user", content: [{ type: "text", text: "用一句话说明「思考模式」对回答质量的帮助，不要分点。" }] }, |
|||
] as XMessage[]; |
|||
for await (const event of client.stream({ |
|||
messages, |
|||
...extra, |
|||
})) { |
|||
if (event.type === "response.start") { |
|||
log(`[stream] start: ${event.provider} ${event.model}`); |
|||
} |
|||
if (event.type === "reasoning.delta") { |
|||
reasoning += event.text; |
|||
log(`[stream] reasoning.delta: ${event.text}`); |
|||
} |
|||
if (event.type === "text.delta") { |
|||
text += event.text; |
|||
log(`[stream] text.delta: ${event.text}`); |
|||
} |
|||
if (event.type === "response.done") { |
|||
log(`[stream] done (finishReason=${event.finishReason ?? ""})`); |
|||
} |
|||
} |
|||
log(`[stream] reasoning 拼接长度: ${reasoning.length}，正文: ${text}`); |
|||
messages.push({ role: "assistant", content: [{ type: "text", text: text }] }); |
|||
log("--- generate() ---"); |
|||
messages.push({ role: "user", content: [{ type: "text", text: "同上问题，再答一次，更短。" }] }); |
|||
const result = await client.generate({ |
|||
messages, |
|||
...extra, |
|||
}); |
|||
if (result.reasoning) { |
|||
log(`[generate] reasoning (前 200 字): ${result.reasoning.slice(0, 200)}${result.reasoning.length > 200 ? "…" : ""}`); |
|||
} else { |
|||
log("[generate] reasoning: （本响应未返回 reasoning_content，属供应商行为）"); |
|||
} |
|||
log(`[generate] text: ${result.text}`); |
|||
} |
|||
File diff suppressed because one or more lines are too long
Loading…
Reference in new issue