Building an end-to-end AI RAG pipeline without a single line of python using Dagger

[raven pelican]

https://youtu.be/XXu8JQxE1AU by @late agate

[orchid pawn]

@late agate could you share the latest version of the module from the demo?

[late agate]

dagger call -m github.com/marcosnils/daggerverse/gptools yt-chat --openai-api-key OPENAI_API_KEY --url https://www.youtube.com/watch?v=BURIKrZOiuA --question "what's the secret of chatGPT success?"

Haven't tried on a Mac but it's very likely it works since it only requires python pretty much.

1|marcos:Projects/dagger (main) (⎈ |dagger-ci)$ dagger call -m github.com/marcosnils/daggerverse/gptools yt-chat --openai-api-key OPENAI_API_KEY --url https://www.youtube.com/watch?v=BURIKrZOiuA --question "what's the secret of chatGPT success?"
Saving/Loading from persist_dir:  /tmp/llama_index/rag_cli
> Running module query with input: 
query_str: what's the secret of chatGPT success?

> Running module retriever with input: 
input: what's the secret of chatGPT success?

> Running module summarizer with input: 
query_str: what's the secret of chatGPT success?
nodes: [NodeWithScore(node=TextNode(id_='35da43b2-f43e-4c85-9492-9c44e2dd9e97', embedding=None, metadata={'file_path': '/files/yt-transcript.txt', 'file_name': 'yt-transcript.txt', 'file_type': 'text/plain',...

The secret of chatGPT's success lies in the combination of pre-training and post-training phases. By focusing on supervised fine-tuning and leveraging pre-training to build common sense knowledge, chatGPT is able to create a model that can interact effectively with users, collect data, and continuously improve through post-training iterations. Additionally, exploring architectures like the RAAG architecture and experimenting with small language models trained specifically for reasoning tasks contribute to chatGPT's success by enhancing its ability to reason effectively and efficiently.

I'd suggest picking videos no longer than 10m or otherwise the transcript part could take some time. Lex clips (https://www.youtube.com/@LexClips) is genrally a good soruce of material for demos

[late agate]

this is not an issue if you have the pipeline cached. That's something nice to show as part of Dagger's automatic caching

[orchid pawn]

@late agate thanks, I'm running it now

When I asked "how is inference handled" you mentioned it's open ai only. But actually I see whisper running, so there is a CPU-bound inference step in there right?

Is it worth trying to get GPU support in the mix, to speed things up?

For a future version you could add swap-in support for Gladia 🙂

[late agate]

Indeed. That's where showing how to replace different parts of the pipeline by adhering to the API shines ✨

[late agate]

Inference in this case is handled by RAG as the whisper transcript gets stored in a vector database which is provisioned by llamaindex.

It'd be nice to generalize this eventually so the vector DB could be expressed as a Dagger service

[orchid pawn]

But whisper itself is inference, is what I meant

[late agate]

Yes but it's not executed every time

Only once per audio file to generate the embeddings

[orchid pawn]

yes that part is definitely nice, I'm enjoying the benefit right now 🙂

[late agate]

Does it work on Mac?

[orchid pawn]

Yes, but cpu only so a bit slow

https://dagger.cloud/dagger/traces/6de4ff6c6aee650a6845ccd3344414df

[raven pelican]

@late agate just realized that someone on youtube asked about the example code on youtube? What would be best to share with them? https://www.youtube.com/watch?v=XXu8JQxE1AU