Question - I have observed the

The similarity score thresholds are not absolute and are highly dependent on which embeddings model you're using . so a score of 0.67 might mean not very similar when using embeddings model 1, while 0.67 might mean very similar when using embeddings model 2. so you need to learn the specific threshold of the embeddings model you're using.
Also consider introducing a reranker to your pipeline to improve the performance.

and if the domain is niche, i am assuming not necesarily in your case, then consider finetuning embeddings adapters and see if that helps.

thanks , how does rereanker help ?

rerankers are mostly cross-encoder models.
so without a reranker what you are doing is embedding the docs and the query separately regardless of any relations between them then you are comparing those separately genrated embeddings (which have already lost some of the meaning because embeddings compress information) while cross encoders take two inputs (the query adn similar docs) and uses a classifier layer to see which similar doc is comparatively the most relevant to the original query by comparing pairs of (original query, similar chunk). dunno if this makes sense

thanks. It does , will do more research thanks again

and about what i mentioned earlier

meaning that the cosine similarity profile is model depdendent. maybe thsi can help

see how the same sentence A to sentence B cosine similarity is returned as 0.52 while using one model, while its 0.64 while using another model

i would also try opensource embeddings models instead of ada.. check the hugging face MTEB leaderboard

i stopped using ada a long long time ago 😄

i used to use an open source embedding but running it locally was an additional cost. Are there any free or low cost mebedding apis available that are better than ada

try converting them to onnx so they run on CPU

cool, thanks will take a look