updated the AI news bot

.env

@@ -1,4 +1,4 @@
 
 AI_NEWS_HUGGINGFACE_API_KEY=""
-
+AI_NEWS_HUGGINGFACE_INFERENCE_API_MODEL="HuggingFaceH4/zephyr-7b-beta"
 AI_NEWS_NEWSDATA_API_KEY=""

src/config.mts

@@ -19,4 +19,6 @@ try {
 }
 
 export const huggingfaceApiKey = `${process.env.AI_NEWS_HUGGINGFACE_API_KEY || ""}`
+export const hfInferenceApiModel = `${process.env.AI_NEWS_HUGGINGFACE_INFERENCE_API_MODEL || ""}`
+
 export const newsdataApiKey = `${process.env.AI_NEWS_NEWSDATA_API_KEY || ""}`

src/createZephyrPrompt.mts

@@ -0,0 +1,26 @@
+
+interface Message {
+  role: "system" | "user" | "assistant";
+  content: string;
+}
+
+/**
+ * Formats the messages for the chat with the LLM model in the style of a pirate.
+ * @param messages - Array of message objects with role and content.
+ * @returns The formatted chat prompt.
+ */
+export function createZephyrPrompt(messages: Message[]): string {
+  let prompt = ``;
+
+  // Iterate over messages and generate corresponding chat entries.
+  messages.forEach(message => {
+    prompt += `<|${message.role}|>\n${message.content.trim()}</s>`;
+  });
+
+  if (messages.at(-1)?.role === "user") {
+    // Append the assistant's tag for the next response but without a closing tag.
+    prompt += `<|assistant|>`;
+  }
+
+  return prompt;
+}

src/getNews.mts

@@ -4,22 +4,26 @@ import { ApiResponse } from "./types.mts"
 export async function getNews(): Promise<ApiResponse> {
 
   const q = [
-    "ai",
-    "artificial intelligence",
+    // "ai",
+    // "artificial intelligence",
     "openai",
     "huggingface",
-    "hugging face",
-    "stable diffusion",
-    "ai ethics",
-    "tesla optimus",
-    "Optimus Gen2",
+    // "hugging face",
+    // "stable diffusion",
+    // "ai ethics",
+    // "tesla optimus",
+   // "Optimus Gen2",
+   "LLM",
+   "chatgpt",
+   "runwayml",
+   "pikalabs",
     "sdxl",
-    "gaussian splatting",
-    "latent space",
-    "gpu",
-    "nvidia",
-    "spatial computing",
-    "apple vision pro"
+   //  "gaussian splatting",
+   //  "latent space",
+    // "gpu",
+    // "nvidia",
+   // "spatial computing",
+    // "apple vision pro"
   ].join(" OR ")
 
   const country = [

src/main.mts

@@ -2,23 +2,29 @@ import YAML from "yaml"
 
 import { sampleDataFreePlan } from "./sampleDataFreePlan.mts"
 import { getNews } from "./getNews.mts"
+import { summarizeWithZephyr } from "./summarizeWithZephyr.mts"
+import { sleep } from "./sleep.mts"
 
 export const main = async () => {
+  // console.log(JSON.stringify(await getNews(), null, 2))
+
+  let delayInSeconds = 15 * 60
 
-  let delayInSeconds = 15 * 60 // let's check every 5 minutes
   console.log(`-------------- ${delayInSeconds} sec have elapsed --------------`)
   console.log("- checking NewsData.io API..")
 
 
-  let newsItems: any[] = []
+  let news: any[] = []
   
+  let nbMaxNews = 5
+
   // const data = await getNews()
   const data = sampleDataFreePlan
 
   const interestingNews = data.results.filter(news =>
     news.language === "english" &&
     news.country.includes("united states of america")
-  )
+  ).slice(0, nbMaxNews)
 
   interestingNews.forEach(({
     article_id, // string
@@ -40,9 +46,7 @@ export const main = async () => {
     sentiment, // string
     sentiment_stats // SentimentStats
   }) => {
-
-    console.log("language:", language)
-    newsItems.push({
+    news.push({
       title,
       data: pubDate,
       countries: country,
@@ -51,11 +55,59 @@ export const main = async () => {
     })
   })
 
-  const newsPrompt = YAML.stringify(newsItems)
-  console.log("news:", newsPrompt)
-  console.log("")
+  const summaries: string[] = []
+  for (const item of news) {
+    console.log("- generating summary..")
+
+    const summary = await summarizeWithZephyr({
+      news: YAML.stringify(item) 
+    })
+    console.log(" -> summary: " + summary)
+    if (summary) {
+      summaries.push(summary)
+      await sleep(1000)
+    } else {
+      console.log("(!) got no data!")
+      await sleep(5000)
+    }
+  }
+
+  const today = new Date().toLocaleDateString('en-us', {
+    weekday: "long",
+    year: "numeric",
+    month: "short",
+    day: "numeric"
+  }) 
+  // "Friday, Jul 2, 2021"
+
+  const markdown = `
+# Title
+
+News Report ${today} 🔥
+
+# Description
+
+A summary of what happened today in the world of tech and AI
+
+# Prompt
+
+You are a TV news channel agent.
+Your mission is to summarize news (which will be given to you as YAML) into a compact, dense news report suitable for a news anchor.
+Please limit yourself to about 20 paragraphs.
+
+Here is the summary of today's news:
+
+article:
+${summaries.join("\n\narticle:")}
+`
+
+  console.log("final markdown:", markdown)
+
+
+
+
+  // TODO: generate a markdow file and upload it to Hugging Face
 
-  // loop every hour
   setTimeout(() => {
     main()
   }, delayInSeconds * 1000)

src/sleep.mts

@@ -0,0 +1,6 @@
+export const sleep = async (durationInMs: number) =>
+  new Promise((resolve) => {
+    setTimeout(() => {
+      resolve(true)
+    }, durationInMs)
+  })

src/summarizeWithZephyr.mts

@@ -0,0 +1,94 @@
+import { HfInference } from "@huggingface/inference"
+
+import { createZephyrPrompt } from "./createZephyrPrompt.mts"
+import { hfInferenceApiModel, huggingfaceApiKey } from "./config.mts"
+
+
+const hf = new HfInference(huggingfaceApiKey)
+
+export async function summarizeWithZephyr({
+  news,
+  neverThrow
+}: {
+  news: string
+  neverThrow?: boolean
+}): Promise<string> {
+  try {
+    const inputs = createZephyrPrompt([
+      {
+        role: "system",
+        content: `Your must summarize the content into 2 or 3 sentence. DO NOT write more than than. Keep it dense and simple, and short.`,
+      },
+      {
+        role: "user",
+        content: news,
+      }
+    ]) //+ "\n["
+
+    const nbMaxNewTokens = 250
+
+    let rawBufferString = ""
+    try {
+      for await (const output of hf.textGenerationStream({
+        model: hfInferenceApiModel,
+        inputs,
+        parameters: {
+          do_sample: true,
+          max_new_tokens: nbMaxNewTokens,
+          return_full_text: false,
+        }
+      })) {
+        rawBufferString += output.token.text
+        // process.stdout.write(output.token.text)
+        if (
+          rawBufferString.includes("</s>") || 
+          rawBufferString.includes("<s>") ||
+          rawBufferString.includes("/s>") ||
+          rawBufferString.includes("[INST]") ||
+          rawBufferString.includes("[/INST]") ||
+          rawBufferString.includes("<SYS>") ||
+          rawBufferString.includes("<<SYS>>") ||
+          rawBufferString.includes("</SYS>") ||
+          rawBufferString.includes("<</SYS>>") ||
+          rawBufferString.includes("<|user|>") ||
+          rawBufferString.includes("<|end|>") ||
+          rawBufferString.includes("<|system|>") ||
+          rawBufferString.includes("<|assistant|>")
+        ) {
+          break
+        }
+      }
+    } catch (err) {
+      // console.error(`error during generation: ${err}`)
+
+      if (`${err}` === "Error: Model is overloaded") {
+        rawBufferString = ``
+      }
+    }
+
+    const tmpResult = 
+      rawBufferString.replaceAll("</s>", "") 
+      .replaceAll("<s>", "")
+      .replaceAll("/s>", "")
+      .replaceAll("[INST]", "")
+      .replaceAll("[/INST]", "")
+      .replaceAll("<SYS>", "")
+      .replaceAll("<<SYS>>", "")
+      .replaceAll("</SYS>", "")
+      .replaceAll("<</SYS>>", "")
+      .replaceAll("<|user|>", "")
+      .replaceAll("<|end|>", "")
+      .replaceAll("<|system|>", "")
+      .replaceAll("<|assistant|>", "")
+    
+  
+    return tmpResult
+  } catch (err) {
+    if (neverThrow) {
+      console.error(`summarizeWithZephyr():`, err)
+      return ""
+    } else {
+      throw err
+    }
+  }
+}