Node embeddings extraction

[DiegoOrtego]

Hi!

First of all, congrats for the great work here and for sharing the code. I'm playing with the code and found that that you are using "text-embedding-3-small" as embedding model for vector search, which is not a free model. You are providing embeddings for nodes, but not for the queries, so I cannot run the full code (maybe I miss something). Anyway, I want to have full control of the models being used so wanted to try out with Qwen3-Embeddings to replace "text-embedding-3-small". However, I realized that there is no code for embedding extraction for the nodes. Could you guide me in how to do it? Do I just need to embed the node text ("texts" field from 'node_embeddings_add_rel_not_compact.pt' and 'node_embeddings_no_rel_not_compact.pt')? I guess that "texts" field in 'node_embeddings_add_rel_not_compact.pt' already contains the relational information to expand the node information with 2-hop info as you describe in the paper, right?

Many thanks in advance!

[DiegoOrtego]

Just to clarify and extend a bit more my question. Debugging your code, I can see that the "texts" fields with relations that you are adding add relations of type brand or things like also viewed. However, sometimes some relations are not added, e.g. category relation. Also, I can't see that 2-hop relations are added there. Am I missing something?
Also, I've seen that before reviews you add some summary of the reviews with the average score and number of reviews. Where do you get this from? Just a template?

Thanks again in advance!

[Derian15]

Hi @DiegoOrtego ,
Thank you for your comments and interest! I apologize for overlooking it until now. Are your questions still relevant? I can upload a file with the code used to retrieve the embeddings of the entities, which should answer your questions.

[DiegoOrtego]

Hi @Derian15, no worries! Thanks for getting back to me. I went deeper in the code and bypassed commercial models adding open ones instead. Maybe the remaining question from my side is more related to Amazon data and whether there is any pre-processing not provided. For example, summarizing reviews (or others). Also, a general feeling is that node texts (for products) are really long and summarizing them retaining key unique aspects will make LLM inference more efficient, while maintaining performance.

[Derian15]

No, there was no preprocessing for the affected step on our side.

We directly embedded the "doc info" of any node provided by the STaRK package, which is:
from stark_qa.skb import SKB
compact_mode=True
skb = load_skb(name=dataset)
doc_info = skb.get_doc_info(node_id, add_rel=add_rel, compact=compact_mode)
There is one exception: If the tokenized text for a node exceeded the input size of the embedding model, we set compact=True. If the input size was still exceeded and add_rel=True, we set add_rel=False. However, compact_mode=True delivers minimal differences compared to compact_mode=False only, i.e., omission of certain special characters as far as I noticed.

Hence, I agree that, especially for smaller models, summarizing reviews (or others) until they fit or even further, might be more efficient, and is worth trying.