This excellent article showcasing how to build sophisticated LLM applications by @dylantartarini.bsky.social is a prime example of the in-depth, application-focused #LLM insights that we are currently seeking.

this is what that script looks like in the UI btw

Not really.
Create a app.py with streamlit (or whatever u want to name it) at the root (same level as your src folder).
Then import whatever is needed from src to mimic a chat (I suppose similar to what you already do in src/main.py).

Post here an example just to give you an idea: it's just python

Great!
In my experience writing Streamlit could become frustrating for complex projects, but chatbots are really easy to do nowadays, even with callbacks and memory -> docs.streamlit.io/develop/tuto... find here some example :)

Hope I helped, this really is an interesting project!

Have you considered using something like Streamlit Community Cloud (streamlit.io/cloud) or Hugging Face Spaces (huggingface.co/spaces)?
It's a minimal effort but might be worth it to showcase the project to a broader audience imho. In both cases the platform would take care of the hosting for you.

Hi Martina, had the chance to take a look at the tool and was wondering: are you guys planning on releasing a Front-End of some sort? Not all academics are computer scientists, might make sense to spread the tool's usage.

Clearly,

- nr. 1 needs domain knowledge and time;
- nr. 2 is computationally expensive and risks entropy but is grounded in docs;
- nr. 3 is risking entropy even more, but is less grounded.

I'd be interested in how to manage the Ontology when building and evolving the graph from documents:

Current options in my mind are
1. predefine the Ontology
2. infer it from docs
3. leave the Agent in charge of extracting entities and relationships free to do whatever it wants

My workaround at the moment (my problem statement was: I don't want to spend money) was to start using Groq (console.groq.com/docs/overview) with online endpoints of open source LLMs, available for free under certain rate / token limits per day.

I am working with some complex prompts due to the use case (Knowledge representation using Graphs).
In my M1 MacBook Air I find that Llama3.2 (1B) works fine for simple/medium tasks at reasonable speed.
However, for some harder tasks I am yet to find a good alternative that could be hosted locally

This chart is from our Spanish Instagram account: www.instagram.com/ourworldinda...

Read more about our translation efforts: ourworldindata.org/instagram-in...