Embedding that fits the budget — pick a model that matches your corpus — step 6 of 9
This is the bug that ships to production most often: the query gets embedded with one model, the documents with another. Cosine similarity is meaningless across two different vector spaces — you get garbage retrieval.
In the editor, two stub "models" (embed_a and embed_b) mimic
the OpenAI vs Voyage situation: same input, completely different
output spaces. The docs were embedded with embed_a. The query
was accidentally embedded with embed_b.
Fix it so the query is embedded with the SAME model as the docs. The user asks "reset my password" — the matching doc should return a cosine of 1.0.
Expected output:
1.0000 reset my password
0.0000 cancel my subscription
0.0000 export my data
This is the bug that ships to production most often: the query gets embedded with one model, the documents with another. Cosine similarity is meaningless across two different vector spaces — you get garbage retrieval.
In the editor, two stub "models" (embed_a and embed_b) mimic
the OpenAI vs Voyage situation: same input, completely different
output spaces. The docs were embedded with embed_a. The query
was accidentally embedded with embed_b.
Fix it so the query is embedded with the SAME model as the docs. The user asks "reset my password" — the matching doc should return a cosine of 1.0.
Expected output:
1.0000 reset my password
0.0000 cancel my subscription
0.0000 export my data
this step needs the editor
on desktop today; in the app (coming soon). save your spot and we'll bring you back here when you're ready.