Regenerate embeddings only when the test representation changes

This commit is contained in:
Claudio Ortolina
2026-04-04 08:32:29 +01:00
parent cd52890f86
commit 2f9460d2cb
3 changed files with 78 additions and 19 deletions
+15
View File
@@ -316,6 +316,21 @@ defmodule MusicLibrary.Records.Similarity do
)
end
@spec generate_embedding(Record.t()) :: :noop | {:ok, RecordEmbedding.t()} | {:error, term()}
def generate_embedding(%Record{} = record) do
new_text = text_representation(record)
case get_embedding_text(record.id) do
{:ok, ^new_text} ->
:noop
_ ->
with {:ok, embedding} <- OpenAI.embeddings(new_text) do
store_embedding(record.id, embedding, new_text)
end
end
end
@spec generate_embedding_async(Record.t()) :: {:ok, Oban.Job.t()} | {:error, Ecto.Changeset.t()}
def generate_embedding_async(record) do
meta = %{title: record.title, artists: Enum.map(record.artists, & &1.name)}
@@ -8,25 +8,10 @@ defmodule MusicLibrary.Worker.GenerateRecordEmbedding do
def perform(%Oban.Job{args: %{"record_id" => record_id}}) do
record = Records.get_record!(record_id)
with {:ok, embedding} <- generate_embedding(record),
{:ok, _} <- store_embedding(record, embedding) do
Records.notify_update(record)
case Similarity.generate_embedding(record) do
:noop -> :ok
{:ok, _} -> Records.notify_update(record)
{:error, _} = error -> error
end
end
defp generate_embedding(record) do
text = Similarity.text_representation(record)
case OpenAI.embeddings(text) do
{:ok, embedding} ->
{:ok, {embedding, text}}
error ->
error
end
end
defp store_embedding(record, {embedding, text_representation}) do
Similarity.store_embedding(record.id, embedding, text_representation)
end
end
@@ -238,6 +238,65 @@ defmodule MusicLibrary.Records.SimilarityTest do
end
end
describe "generate_embedding/1" do
test "generates and stores embedding for a record without existing embedding" do
record = record(%{genres: ["rock", "alternative"]})
embedding = Enum.map(1..1536, fn _ -> 0.5 end)
Req.Test.stub(OpenAI.API, fn conn ->
assert conn.request_path == "/v1/embeddings"
Req.Test.json(conn, %{
"data" => [%{"embedding" => embedding}]
})
end)
assert {:ok, _record_embedding} = Similarity.generate_embedding(record)
assert {:ok, _stored} = Similarity.get_embedding(record.id)
end
test "returns :noop when text representation is unchanged" do
record = record(%{genres: ["rock", "alternative"]})
text = Similarity.text_representation(record)
embedding = Enum.map(1..1536, fn _ -> 0.5 end)
{:ok, _} = Similarity.store_embedding(record.id, embedding, text)
assert :noop = Similarity.generate_embedding(record)
end
test "regenerates embedding when text representation changes" do
record = record(%{genres: ["rock"]})
old_text = "Old text representation"
old_embedding = Enum.map(1..1536, fn _ -> 0.5 end)
new_embedding = Enum.map(1..1536, fn _ -> 0.9 end)
{:ok, _} = Similarity.store_embedding(record.id, old_embedding, old_text)
Req.Test.stub(OpenAI.API, fn conn ->
Req.Test.json(conn, %{
"data" => [%{"embedding" => new_embedding}]
})
end)
assert {:ok, _record_embedding} = Similarity.generate_embedding(record)
assert {:ok, stored_text} = Similarity.get_embedding_text(record.id)
refute stored_text == old_text
end
@tag :capture_log
test "returns error when OpenAI API fails" do
record = record(%{genres: ["rock"]})
Req.Test.stub(OpenAI.API, fn conn ->
Plug.Conn.send_resp(conn, 500, JSON.encode!(%{"error" => "internal server error"}))
end)
assert {:error, _reason} = Similarity.generate_embedding(record)
end
end
describe "find_similar/2" do
setup do
# Create test records with embeddings