Note: as of 22oct24 (v0.3.83), I declare the modernize-langchain-latest branch dead. Code has been succesfully merged into main and now we can go back to our lives and clean up old Modernize stuff. Note: this was the MODENRIZE branch!

Now on:

• Penguin (TODO reconcile yet)
• Derek (who was broken up but I then took all the code and it worked fine)

About ♊️ GemiNews 📰

Self: palladius/gemini-news-crawler (public)

This is a News Slurper that takes News in real time and - hopefully - feeds an LLM with RAG knowledge.

• Session: https://sessionize.com/app/speaker/session/621013
• Slides: Ricc Slides

Apps are on Cloud Run:

• DEV: https://gemini-news-crawler-dev-x42ijqglgq-ew.a.run.app/
• PROD: 🫀 health check PROD

Description

How can we get an LLM to be updated to today’s news? Gen AI is great at answering questions.. from the past. After the LLM was trained, all you can do is RAG. How about crawling the web for latest news with Gemini for multimodal extraction and offering summarization by your favorite topic? It all gets more exciting thanks to Andrei’s langchainrb gem.

Features

• [hot] Gemini function calling tools:
  • NewsRetriever: getting News from online (part of langchainrb gem),
  • ArticleTool: and from Active Record (local under webapp/app/tools/article_tool.json).
• Use langchainrb gem for Tools, Prompts, AI services (mostly Google ones). Note this version is frozen to 0.13.1 as I had to move fast and monkeypatch the gem in my code rather than sending upstream changes. Will fix this in a future version.

Demos

4 juicy demos are available under webapp/docs/demo/:

https://github.com/palladius/gemini-news-crawler/blob/main/webapp/docs/demo/DEMO.md

Other Ideas

My idea is to bring slides and a demo, all done in ruby leveraging nokogiri, langchainrb and possibly some capabilities in Langchainrb that Andrei is now building (*).

Slides: explain the overall idea, empathise with audience, show architecture diagram, why we’re here, and make people laugh.

My idea is to build a demo in two parts:

• A crawler which crawls a few sample web pages, extract information using Gen AI to understand if they’re pertinent to certain topics (eg music, sport, politics, ..) and extract other information (eg Location).

• Then, RAG-style, I’d feed an LLM and ask questions real time hoping to be able to surprise people with last-week news about different news sections. Like: “How are presidential elections going? What’s the latest news?” What’s latest with the ruby community? .. hoping to retrieve very latest news.

Possibly, retrieve similar pictures/articles based on the questions (embedding style).

• P2. AWESOME. Add a research by embedding. Something like "Search something about fun sport" and it calculates the embedding of "fun sport" and returns 5 closest articles. this means creating and declaring one more function.

App info

TODOs

• add Devise for user mgmt
• add Cloud Run IAP: https://blog.cloud66.com/authenticating_users_with_google_iap_in_rails
• Auto feed continuously. Currently manually done on my local machine :(
• Use updated Gemini embedding models, new since May 14th (launched at NEXT ‘24).
  • text-embedding-004
  • text-multilingual-embedding-002.
• Add multimodal embeddings (search by article picture). This can be achieved by simply adding another embedding:
  • picture_embedding: if picture exists, fetch it and calculate it.
  • picture_description: if picture exists, fetch it and ask Gemini to provide an automated picture description. This is also cool for visually impaired people, and can be added automatically to the pic description! Would be probably worth fetching and downloading the picture on GCS, cos you never know, picture could disappear.

Autofeed now

1. cd crawler/ ; $ make crawl-a-lot or make crawl-continuously. This populates XML every 15min (or I get kicked out by the robots :P ) and slurps articles from XML. XML I check on git, articles i dont or theyre too many.
2. cd webapp ; bundle exec make seed-forever (without bundle wont work). this seeds info from (1) into ActiveRecord, hence DB.
3. call an async routing to populate - although since v0.1.5 this should happen automatically before save of Article.
4. This workED: cd webapp ; echo Article.compute_embeddings_for_all | rails c. Note: since I moved from Array to Vector this script is now BROKEN

• Created secret: projects/272932496670/secrets/geminews-key
• Mounted on Crun as /geminews-key/geminews-key
• Now the final bit: GCP_KEY_PATH_FROM_WEBAPP = /geminews-key/geminews-key

Bibliography

• This Github repo: https://github.com/palladius/gemini-news-crawler/
• Langchain gem: https://github.com/patterns-ai-core/langchainrb
• Gemini function calling: https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/function-calling
• Embeddings explained by Sally Goldman: https://developers.google.com/machine-learning/crash-course/embeddings/video-lecture
• Video of this presentation: https://youtu.be/-ieyahrpDpM
• Slides (private): go/ricc-gemini-verona