feat: Add streaming and non-streaming chat completion handlers

Author: ericc-ch

src/routes/chat-completions/handler-streaming.ts

@@ -0,0 +1,33 @@
+import type { Context } from "hono"
+
+import consola from "consola"
+import { streamSSE } from "hono/streaming"
+
+import type { ChatCompletionsPayload } from "~/services/copilot/chat-completions/types"
+
+import { chatCompletionsStream } from "~/services/copilot/chat-completions/service-streaming"
+
+export async function handlerStreaming(c: Context) {
+  const payload = await c.req.json<ChatCompletionsPayload>()
+
+  consola.info(`Received request: ${JSON.stringify(payload).slice(0, 500)}`)
+
+  const response = await chatCompletionsStream(payload)
+
+  return streamSSE(c, async (stream) => {
+    let index = 0
+    for await (const chunk of response) {
+      if (index === 0) {
+        consola.info(
+          `Streaming response, first chunk: ${JSON.stringify(chunk)}`,
+        )
+      }
+
+      index++
+
+      await stream.writeSSE({
+        data: JSON.stringify(chunk.data),
+      })
+    }
+  })
+}

src/routes/chat-completions/handler.ts

@@ -0,0 +1,40 @@
+import type { Context } from "hono"
+
+import consola from "consola"
+import { streamSSE } from "hono/streaming"
+
+import type { ChatCompletionsPayload } from "~/services/copilot/chat-completions/types"
+
+import { chatCompletions } from "~/services/copilot/chat-completions/service"
+
+import { createContentChunk, createFinalChunk, segmentResponse } from "./utils"
+
+export async function handler(c: Context) {
+  const payload = await c.req.json<ChatCompletionsPayload>()
+  payload.stream = false
+
+  consola.info(`Received request: ${JSON.stringify(payload).slice(0, 500)}`)
+
+  const response = await chatCompletions(payload)
+  consola.info(`Response from Copilot: ${JSON.stringify(response)}`)
+
+  const segments = segmentResponse(response.choices[0].message.content)
+  const chunks = segments.map((segment) =>
+    createContentChunk(segment, response, payload.model),
+  )
+
+  chunks.push(createFinalChunk(response, payload.model))
+
+  consola.info(
+    `Streaming response, first chunk: ${JSON.stringify(chunks.at(0))}`,
+  )
+
+  return streamSSE(c, async (stream) => {
+    for (const chunk of chunks) {
+      await stream.writeSSE({
+        data: JSON.stringify(chunk.data),
+      })
+      await stream.sleep(1) // Simulated latency
+    }
+  })
+}