How to scrape the web for LLM in 2024: Jina AI (Reader API), Mendable (firecrawl) and Scrapegraph-ai

Рет қаралды 110,692

Ай бұрын

👨‍💻 Code: github.com/trancethehuman/ai-...
(if there are issues with viewing the code, just fork and clone the repository. It's just a current problem with GitHub's way of displaying Jupyter notebooks - nbconvert)
Tools mentioned:
Jina AI: jina.ai/reader
Mendable's Firecrawl: www.firecrawl.dev/
Scrapegraph-ai: github.com/VinciGit00/Scrapeg...
🚌 Sign up for my upcoming AI engineering course: tally.so/r/n9daQ1
Follow me on Twitter: / haithehuman
Find me on LinkedIn: / haiphunghiem
(Consulting) If you want me to work with you: tally.so/r/n9djRQ

Пікірлер: 108

@jarad4621 Ай бұрын

Dammit stop telling everybody about Jina my secret weapon, just stop, it's my advantage, everybody ignore it it's horrible I swear

@devlearnllm Ай бұрын

TOO LATE

@jarad4621 Ай бұрын

@@devlearnllm 😉. This was one of the mostly highly valuable vids ive seen in past few weeks when considering the contents, the top 3 special scrapers i searched hard for mentioned all together in one good video, nice, add to good cheap open source llm like llama 3 and it = $$$ if you know how, data is valuable, things that were not possible or affordably viable for most previously are now, i can do stuff for $12 now that some would pay thousands for, its a wonderful new world! Just finished something awesome with python and Jina and openrouter Llama 3 in 2 days thats gonna double my revenue or more and i dont even know how to code lol, thanks gpt. Jina does have paid api key on the api page btw, 1m free, 580 pages or so it worked out to. but the pricing is so low its insane, 500m tokens or 280 000 pages for $10, destroys firecrawl pricing, which is also good and has its place but much more costly). i think scrapegraph uses llm to parse so its gonna be expensive on tokens right, sending raw website to llms? Ive asked them like you did. I only wish Jina showed menus and internal links and it would be perfect, those have valuable data itself and identifies more valuable pages for more visits like pricing, ill ask if there is a way but i guess can add something cheap to the workflow for that, any suggestions? Prob some python libary, ill ask perplexity lol. Im actually new to the tech side but i see the business value as a marketer so learning fast as i can! its the new gold rush. Great video, subbed looking forward to more. Cheers

@TheBrighamhall Ай бұрын

@@devlearnllm thought you said Jira and was so confused..

@devlearnllm 29 күн бұрын

@@TheBrighamhall imagine lol

@ShinyTechThings 28 күн бұрын

🤣🤣🤣🤣

@kylelau1329 Ай бұрын

Thank you for introducing all the latest technology for web scraping!

@matten_zero Ай бұрын

The reader API tip is so clutch. Thank You!

@antoniuskonovalov Ай бұрын

Just started wondering about web scraping and here you are. Thank you.

@alonsoalarconaguilar7113 26 күн бұрын

KZfaq algorithm is just insanely good at what it does, this exactly the content I needed and I think I have found what I want to dedicate my life to as a professional. Thank you for the video, I will buy your course as fast as I collect the money.

@kuhltime Ай бұрын

Came at the perfect time. Very good video. Thx 😊

@ariG23498 13 күн бұрын

How did I not get your content sooner? Love it!

@forgotmyoldSN 13 күн бұрын

Thanks for adding a new project to my to do list!

@st.3m906 Ай бұрын

Amazing video, thank you

@markt4565 23 күн бұрын

keep up the good work! - this is an awesome presentation!

@devlearnllm 23 күн бұрын

@moafro6524 17 күн бұрын

Underrated glad I found

@NikhilSwamiExperimental 13 күн бұрын

chigga dropping bomb content, meranwhile i made a comment analyzer for highly detailed videos which have 100+ comments, and dint have time for going through all. man, sometimes you dont need to build an ironman suit to do simple shet.

@devlearnllm 12 күн бұрын

Printing this comment out and putting on my wall

@aarushsaboo1194 6 күн бұрын

Bro, did you build a comment analyzer for all youtube videos in which all you need to do is post a youtube link? That's a nice project!

@shivam_in 27 күн бұрын

If I'm going to scrap millions of pages regularly, no way in hell AI would come anywhere close in accuracy and efficiency than a plane Http request or browser load and Jsoup parsing.

@nickk6575 Ай бұрын

Greta video! The open source tool looks great! As an aside, I use instructor and pydantic classes to get the LLMs to provide the JSON as I expect it. In my limited experience, dspy wasn't as explicit as I wanted.

@devlearnllm Ай бұрын

Good idea

@jarad4621 Ай бұрын

Are you you using thos two libraries with agency swarm agentic framework, it uses those a well to ensure performance/quality, if not maybe something you might be interested in, a proper production-capable agentic framework. That with its automation and decision-making capability plus Jina + llms = profit for so many use cases

@catchychazz 29 күн бұрын

Are you referring to DSPy assertions?

@terrytan1827 16 күн бұрын

16:05 Worth trying out GPT-4, I find it more accurate at following instruction.

@sitedev Ай бұрын

Gold!

@uwepleban3784 Ай бұрын

The transcript at 1:39 states that you are using large sandwich models. This must be a brand new type of model - mouth watering indeed. 😂

@devlearnllm Ай бұрын

Heck yeah 🥪

@khemchay 19 күн бұрын

Jina love it...

@user-xj5gz7ln3q Ай бұрын

GPT 4o can do this now. Just tested and it's awesome.

@BlueBearOne 27 күн бұрын

Thank you. I'll be "away" for a while while I conquer the...I mean save the world!

@RenkoGSL Ай бұрын

lol that's awesome!

@devlearnllm Ай бұрын

If anyone’s having issues viewing the notebook on GitHub, it’s GitHub’s fault. Feel free to clone it (the cod e is there, GH just couldn’t display it recently: stackoverflow.com/questions/78501731/error-nbformat-when-uploading-to-github-from-google-colab)

@augmentos Ай бұрын

Can anyone speak to the architecture or other tools to prevent detection using beautiful soup as he mentioned? What would be the best process to avoid detection and what tools I wish you elaborated there considering it’s the subject of video in large part.

@roberthuff3122 Ай бұрын

🎯 Key Takeaways for quick navigation: 00:00 *🚀 Introduction to web scraping for LLMs in 2024* - Overview of startups pivoting to web scraping. - Mention of Mendable and its "fire crawl" tool for scraping the web using large language models. 02:06 *🔍 Scraping competitors' pricing pages* - The process of scraping competitors' pricing for market research. - Introduction to tools used for scraping: Jina AI, Mendable, and Scrapegraph-ai. 03:01 *🧠 Understanding "Tik token" and its application* - Explanation of tokenization and encoding in web scraping. - Discussion on the cost implications based on tokenization. 05:17 *🛠️ Setting up scrapers with Beautiful Soup and other tools* - Description of different scraping tools and their setup. - Comparisons among Beautiful Soup, Jina AI, and Mendable based on ease of use and output. 07:32 *📊 Running scrapers and analyzing outputs* - Execution of web scraping and evaluation of the output from different tools. - Analysis of readability and format of the scraped data. 09:37 *💰 Cost comparison and effectiveness of scraping tools* - Comparison of costs associated with various scraping tools. - Evaluation of which tool provides the most value for money. 12:53 *🤖 Extracting pricing information using OpenAI* - Utilization of OpenAI for extracting specific data points. - Challenges and strategies in obtaining clean and useful information. 17:20 *🌐 Overview of Scrapegraph for advanced web scraping* - Introduction to Scrapegraph as an open-source project. - Examples of complex data extraction and its accuracy. Made with HARPA AI

@thethree60five Ай бұрын

...The best in-browser AI automation system.

@thedoctor5478 Ай бұрын

This Jina thing is cool. The beautifulsoup scraper is obviously not a solution. Most web pages (Especially articles, media, etc.) have google schema ld+json ready to be extracted though. There are some good python libs for getting the metadata. There are many scraping APIs, and most of them are not worth the cost IMO. phantomjscloud is probably one exception, depending on volume. Otherwise, one must find a good proxy provider and send a bunch of fancy http headers to bypass anti-bot, like you said. Blackhatworld is a great resource for proxies and all manner of other accounts. The whole scraping thing is a giant rabbit-hole. Jina is for sure keeping all that data. It's not a bad plan, actually. I think I may do the same.

@planplay5921 Ай бұрын

But the first problem that all crawls need to face is how to avoid being blocked.

@PracticalAI_ 18 күн бұрын

there are ways, maybe I will do a video about that ... but that is a dark art :)

@planplay5921 18 күн бұрын

@@PracticalAI_ I'm really looking forward to it!😊

@Van-Helssen 6 күн бұрын

Rotation of proxies and query randomly dude, easy task

@PracticalAI_ 6 күн бұрын

@@Van-Helssen lol it's not 2014, proxies are recognised by most providers, and they will immediately invalidate the user (if you are scraping as login). There are other ways, using regular ips

@Van-Helssen 6 күн бұрын

@@PracticalAI_ *residential proxies as you would probably know….

@jetlime08 19 күн бұрын

Is the LLM community really not aware of 40 year old Natural Language Pre-processing methods developed for data mining and NLP?

@erickcampos50 16 күн бұрын

Could you explain it better? I can't see how to connect what you said with this subject

@josefaguilar2955 8 күн бұрын

I don't know if the community is aware that this has been a problem to solve for quite some time.

@marthasamuel 26 күн бұрын

Would these work for a dynamic website

@stanTrX Ай бұрын

What are the good and easy to use tools with langchain? Llm is not very useful without such tools, even it has no idea about the date today.

@TranKiet-pj9mw 19 күн бұрын

youtube really know what i am looking :V with python craw a website with LLM is simple just a few line of code . back to 8 year ago i used python tool do a same thing with higher effort . right now , i m trying to mixed data from website/ database with knowledge map for observation view then i could find the short path according its , that will taking less time to read entire book in this field , just focus in some topic but still get the result . nah but you introduced the method with LLM . thanks

@devlearnllm 12 күн бұрын

Awesome. Thanks for sharing

@supriyosarkar1806 19 күн бұрын

I feel really sad. that you publicly talked about Jina. I used to feel special knowing very few people are aware of it lol

@devlearnllm 18 күн бұрын

my badd

@danielcave9606 13 күн бұрын

How well does Jina do with bigger sites with anti-bot protection?

@stevefox7469 Ай бұрын

How do these tools cope with CloudFlare operating on the target site, which attempts to block scrapping?

@svenvanwier7196 26 күн бұрын

cant stop the bots i know about seleniumbase for python..... takes some research but... hey

@thingX1x 16 күн бұрын

Using jina now hehe. Does anyone know if you can get better results from amazon?

@AtharvDharmadhikari-vc9fk 29 күн бұрын

I used scrapegraph ai and was also stuck to get cost, but then I just took the cost my making some changes inside the scrapegraphai library as internally the library is using langchain and langsmith so it was calculating the cost.

@devlearnllm 29 күн бұрын

That's awesome. How do you get it to work with LangSmith?

@artmadiar Ай бұрын

Great presentation! I'm surprised about jin ai free scraper that doesn't require an API?!! I guess it might be shut down soon for public access

@jarad4621 Ай бұрын

There is a paid version thats worth it, check the api page, key at bottom out generates a unique one somehow, you get 1m free then $10 for 500m tokens which is like 280k pages which is insanely low and basically free anyways, crazy valuable tool

@artmadiar Ай бұрын

@@jarad4621 oh wow! it's amazing! thanks for clarification

@nzt29 12 күн бұрын

Haven’t watched it fully yet, but I’m really curious to see how it handles the looming threat of model collapse. edit: Yeah it didn’t talk about it. It’s going to be hellish when the internet becomes increasingly flooded with LLM output

@prashantbhardwaj6322 25 күн бұрын

Can you please fix the camera please already feeling dizzy within 60 seconds due to constant camera movement!

@devlearnllm 24 күн бұрын

Working on it. Just need to find the setting in DJI Pocket 3 to slow down the tracking speed

@GeoffY2020 Ай бұрын

i tried to read or download the Web_scraping_for_LLM_in_2024.ipynb but its not readable, can you replace it ?

@GeoffY2020 Ай бұрын

ok i can read it in colab

@PedroIvo-iz5sv 26 күн бұрын

it works in portuguese?

@PineState77 24 күн бұрын

What’s the best way to get in touch?

@devlearnllm 24 күн бұрын

Details in the video’s description

@denisblack9897 Ай бұрын

Damn, bro get ready for heavy lifting) baldness is coming Been there, you’ll look much much better!

@devlearnllm Ай бұрын

Lmao thanks brother

@bastabey2652 15 күн бұрын

these scrapping tools are impressive... but they are not ready for scrapping full website with 100s of webpages.. unfortunately, there is still significant a room for manual scraping..

@jarg7 Ай бұрын

broken link to github

@devlearnllm Ай бұрын

Yeah there’s something weird with GitHub not displaying the notebook right. The link is the same.

@MMABeijing 29 күн бұрын

That s basic stuff, I feel like it s 2023, and I was late to the party too

@flor.7797 Ай бұрын

none of these seem better than Trafilatura?

@flor.7797 Ай бұрын

scrapegraph looks cool though

@devlearnllm Ай бұрын

@@flor.7797 How's your experience using Trafilatura? I haven't tried that yet

@flor.7797 Ай бұрын

@@devlearnllm I’m more into main content extraction and boilerplate removal. There isn’t one size fits all unfortunately

@PaulFidika 6 күн бұрын

"The entire internet hates him for this one simple trick"

@devlearnllm 5 күн бұрын

9/10 prompt engineers recommend this

@eyoo369 5 күн бұрын

Jina is almost perfect.. too bad it's not smart enough to scrape content from "accordions" where you first click to make the content visible. I feel a smart AI scraper should be able to grab that text and determine based on CSS class that it's probably valuable text.. just hidden at the time

@devlearnllm 5 күн бұрын

That's too bad. What's the alternative?