Video summary

n8n: как это работает? | Контент с локальным n8n и LMStudio | готовим новости для ТГ-канала

Main summary

Key takeaways

Technology

Overview

The video demonstrates an experiment/build tutorial for an automated pipeline:

news → AI-filter → Russian translation → Telegram post

It uses n8n (referenced in subtitles as “Nathan/Naton”) and a local LLM running in LM Studio.

Core Idea / Workflow

1) Input source: a tech news page

  • The pipeline starts with a tech news webpage.
  • In the example, TechCrunch is used as the news source.

2) Fetch HTML

  • Use an HTTP request node to download the page HTML.

3) Parse HTML to extract items

  • Use an HTML extract node with CSS-like selectors to pull:
    • headlines/titles
    • links/URLs

4) Combine title + link

  • A JavaScript code node combines two parallel arrays (titles and links) into one array of objects:
{ title, link }
  • The speaker notes n8n is generally more low-code than no-code:
    • code nodes are small snippets
    • they can be generated via GPT chat

5) Fetch each article page

  • For every extracted link:
    • request the article page
    • extract its raw content (a large payload)

6) Extract main article text

  • Another HTML extraction node pulls the article text using pre-selected selectors.

7) Merge enriched items back together

  • A Merge node ensures each item includes:
{ title, link, content }
  • A final code step formats the results into the Items array structure expected by the next node.

Local LLM Stage (Filter + Translation + Formatting)

Model source: LM Studio (local)

  • The LLM runs locally via LM Studio.
  • The subtitles mention “GPTOSS20B” (an auto-generated name; likely a GPT-style local model with ~20B parameters and a large context window).

n8n “Language model” node

Configured with:

  • Base URL: LM Studio endpoint
  • System prompt: created with the help of GPT chat and includes instructions to:
    • translate selected items into Russian
    • deduplicate URLs
    • ignore irrelevant clickbait titles
    • act like an editor that selects relevant AI news
  • User prompt: provides the context for each item (title/link/content), so the model can:
    • pick items related to AI
    • translate them
    • format the output for Telegram

The model’s output becomes the text that will be sent to Telegram.

Telegram Output

  • Uses an n8n Telegram node (e.g., “Send a text message”).
  • Requires:
    • Telegram bot token
    • Telegram chat ID (channel)
  • Sends the model’s output as the message content.

Results / Review Notes (Important Nuance)

  • The pipeline works end-to-end: it posts AI-related news in Russian to Telegram.
  • Noted issues:
    • The model sometimes left headlines in English, despite translation instructions.
    • It sometimes did not fully follow headline-related instruction details.
  • The speaker attributes this more to model behavior and settings than to the n8n pipeline itself.

Conclusions from the Experiment

  • The example demonstrates a practical n8n pipeline that:
    • parses an unstructured webpage
    • extracts and enriches items (title/link + full article text)
    • uses a local LLM to filter/translate/format
    • posts automatically to Telegram
  • The video also mentions n8n supports other ingestion methods like RSS feeds and APIs, but this specific example uses the simplest HTML-parser-based flow for learning/testing.

Main Speakers / Sources (as referenced in the video)

  • Speaker: the video host/instructor (not explicitly named in subtitles)
  • Tools / sources used:
    • n8n
    • LM Studio
    • Telegram Bot API
    • TechCrunch (example news source)
    • GPT chat (used to help generate prompts/code)

Original video