Chroma does all this

[ndrean]

Check this link

If you search your photos for "famous bridge in San Francisco". By embedding this query and comparing it to the embeddings of your photos and their metadata - it should return photos of the Golden Gate Bridge.

[nelsonic]

@ndrean did you open this discussion as a way of comparing features of existing projects...? 🤷‍♂️

Yes. There will always be multiple solutions to all problems. 👌
McDonalds already does burgers. 🍔 ✅
Doesn't mean I'm not going to make my own gourmet ones from scratch to learn how to do it better. 😉

We only wanted a way of describing images we want to upload. 📸 ⬆️ ☁️ 📝
The rest of the features are useless to us. 😒
Chroma certainly looks good but we didn't want to build ours in Python or JavaScript. 🙅

[ndrean]

We only wanted a way of describing images we want to upload. 📸 ⬆️ ☁️ 📝
The rest of the features are useless to us. 😒

I understand that you mean that the semantic search is useless to you as a tool. I'm fine with this, but this repo demonstrated a way to run a semantic search in Elixir as embeddings and similarity search are basic AI/ML related tools.

did you open this discussion as a way of comparing features of existing projects?

In a sense yes, but it goes beyond.
You may be curious about RAG Retrieval Augmented Generation. Suppose you want to deliver a "natural language" response based on say your data, then you may need to store embeddings and run similarity search to inject this into the context of the LLM prompt. I wanted to point out that the Chroma project offers this enhanced vector badatase capabilities and is based on PG_VECTOR which I ruled out to use HNSWLib. It is Python based indeed, and my intention is more to draw attention to this particular solution, even if not in the Elixir world.

[ndrean]

Continuation of this RAG process that highlights vector search. You can find numerous papers on RAG. A selection below.

• Firstly 2 pictures below illustrates how RAG is used to enhance LLM chatbots.

1)Extracted from the NVIDIA blog and I share the link.

It says it reduces hallucination and claims to be easy (!).
It also removes (?) the need of training your model. The later seems to be a hard task in the Elixir world, cf https://elixirforum.com/t/rag-app-using-elixir-feasible/60439/16).

2. Another picture to illustrate where vector search (aka semantic search) is used in RAG process: you enhance the question with a context found by vector search from some controlled data, and let the LLM formulate the response nicely for you.

• a paper on RAG from Meta including a HuggingFace link.

• Lastly, a valuable resource from the https://pincecone.io blog that explains the difference between a standalone Vector Index (such as hnswlib) vs Vector Database using indexing (such as pg_vector or chromadb). Similarity search is at the heart of these technologies. See this other Pinecone blog on vector search. See this link for various managed vector db. It does not include chromadb which can be self hosted (in the same way as Postgres pg_vector).

I will try to experiment this in a Livebook. TODO