93 lines
3 KiB
TypeScript
Executable file
93 lines
3 KiB
TypeScript
Executable file
#!/usr/bin/env -S deno run -A
|
|
|
|
import { TextLineStream } from "https://deno.land/std@0.224.0/streams/mod.ts"
|
|
import ollama, { Message } from "npm:ollama@0.5.15"
|
|
|
|
function main() {
|
|
const names = new Names()
|
|
const messages = new SpillingQueue<Message>(500)
|
|
listen(8080, (conn) => session(conn, names.pick_for(conn.remoteAddr.hostname), messages))
|
|
}
|
|
|
|
class SpillingQueue<T> {
|
|
public items: T[] = []
|
|
|
|
public constructor(
|
|
public limit: number,
|
|
) {}
|
|
|
|
public push(value: T) {
|
|
this.items.push(value)
|
|
if (this.items.length > this.limit) return this.items.splice(0, 1)[0]
|
|
}
|
|
}
|
|
|
|
async function listen(port: number, handler: (conn: Deno.TcpConn) => Promise<unknown>) {
|
|
console.log("Listening on port", port)
|
|
for await (const connection of Deno.listen({ transport: "tcp", port })) {
|
|
handler(connection).catch((except) => console.log("Session closed", except))
|
|
}
|
|
}
|
|
|
|
const header = (name: string) => `
|
|
KUB
|
|
|
|
Kub is a LLM running on TCP which includes all messages in the same context.
|
|
As Kub is having several discussions at once, he will identify you as : ${name}
|
|
───────────────────────────────────────────────────────────────────────────────
|
|
`
|
|
|
|
async function session(conn: Deno.TcpConn, name: string, messages: SpillingQueue<Message>) {
|
|
console.log("Opening session for", name, "at", conn.remoteAddr.hostname, conn.remoteAddr.port)
|
|
conn.setNoDelay(true)
|
|
await conn.write(new TextEncoder().encode(header(name)))
|
|
async function prompt() {
|
|
await conn.write(new TextEncoder().encode("> "))
|
|
}
|
|
const lines = conn.readable
|
|
.pipeThrough(new TextDecoderStream())
|
|
.pipeThrough(new TextLineStream())
|
|
await prompt()
|
|
for await (const line of lines) {
|
|
await conn.write(new TextEncoder().encode("\n< "))
|
|
const user_content = `(message from:${name}) ${line}`
|
|
const user_message = { role: "user", content: user_content }
|
|
const response = await ollama.chat({
|
|
model: "llama3.2",
|
|
messages: request_messages(messages.items, user_message),
|
|
stream: true,
|
|
options: {},
|
|
})
|
|
let content = "", role = "assistant"
|
|
for await (const part of response) {
|
|
await conn.write(new TextEncoder().encode(part.message.content))
|
|
content += part.message.content
|
|
role = part.message.role
|
|
}
|
|
await conn.write(new TextEncoder().encode("\n\n"))
|
|
await prompt()
|
|
messages.push(user_message)
|
|
messages.push({ role, content })
|
|
}
|
|
}
|
|
|
|
function request_messages(history_messages: Message[], user_message: Message) {
|
|
return [
|
|
{ role: "system", content: "You are 'Kub', a bot which talks to people on the internet. Be concise and helpful." },
|
|
...history_messages,
|
|
user_message,
|
|
] as Message[]
|
|
}
|
|
|
|
class Names {
|
|
options = ["Adam", "Bastien", "Corentin", "Dominique", "Ethan"]
|
|
|
|
public pick_for(key: string) {
|
|
let sum = 0
|
|
for (const letter of key) sum += letter.charCodeAt(0)
|
|
const index = Math.floor((sum + Math.random() * 100) % this.options.length)
|
|
return this.options[index]
|
|
}
|
|
}
|
|
|
|
if (import.meta.main) main()
|