Examples not working with Ollama

[dbonattoj]

When I run the example provided:
wdoc --model="ollama/qwen3:8b"
--query_eval_model="ollama/qwen3:8b"
--embed_model="ollama/bge-m3"
--task summarize --path https://situational-awareness.ai/

It fails in two ways:

1. it says that the embed_model should be openai, sentencetransformers or huggingface

AssertionError: Backend of embeddings must be either openai, sentencetransformers or huggingface

2. if I replace the embedded model by --embed_model="sentencetransformers/all-MiniLM-L6-v2"
   It fails asking for my OPENAI_API_KEY

Exception: No environment variable named OPENAI_API_KEY found

I would like to use it completely with ollama.

Also the documentation is not clear about saving.
Suppose you have a big folder of pdfs, you process them, you can save them as an embedding and then ask the model questions by using the same embedding at another time, is that correct?
What happens if I want to increase the number of documents ?

Thanks,

[thiswillbeyourgithub]

Hi!

Thanks for the interest in wdoc!

It appears the fix is simple: you must be using a way outdated version of wdoc. As the weird embedding backend situation was overhauled a while back. Do you have any idea how this happened?

You could try with uvx wdoc but there's a chance uvx fails.

Suppose you have a big folder of pdfs, you process them, you can save them as an embedding and then ask the model questions by using the same embedding at another time, is that correct?
What happens if I want to increase the number of documents ?

If you set --save_embeds_as to point to some dir, then the embeddings will be saved there. If next time you set --load_embeds_from then those embeddings will be loaded. You should not use both at the same time. save_embeds_as will overwrite any previous embeddings at this location.

In the passed I actually implemented a pretty advanced deduplicated setup that on the fly split each faiss vectors into subvectors then checked which should be recomputed or not. Given that wdoc was made to be in theory able to handle hundreds of thousands of documents, I had also made some multiprocessing sorcery to pass those vectors arrounds etc. It worked but I wasn't happy with how brittle it seemed. This would make it probably not work in some constrained setup (like recently inside an open-webui docker container). And this is without mentioning the several layers of cache that existed at some point for various reasons.

And then providers started seriously dropping the cost of embeddings so I figured everyone would be better off paying the millionth of cents for it.

But indeed, this design decision is not appropriate for ollama because although cheap, it takes time to do those embeddings.

Also there is an ongoing plan to support basically all other vector stores too in issue #13 .

So to answer explicitly: if you increase the number of documents it will recompute the embeddings.

Wait no. I had forgotten but I'm using a cached embedder in the backend. It's the one from langchain so I'm not sure if it handles cache for individual documents. By that I mean say you send [doc1, doc2] as a batch, then some day later you send [doc2, doc3], I'm not sure it will not recompute doc2.

Does that answer your question?

[thiswillbeyourgithub]

I updated the doc, thanks!

I'm closing this but feel free to reopen if there's still anything I can do for you.