mirror of
https://github.com/discourse/discourse-ai.git
synced 2025-11-04 00:18:39 +00:00
5f0c617880
Single and multi-chunk summaries end using different prompts for the last summary. This change detects when the summarized content fits in a single chunk and uses a slightly different prompt, which leads to more consistent summary formats. This PR also moves the chunk-splitting step to the `FoldContent` strategy as preparation for implementing streamed summaries.
96 lines
3.0 KiB
Ruby
96 lines
3.0 KiB
Ruby
# frozen_string_literal: true
|
|
|
|
RSpec.describe DiscourseAi::Summarization::Models::Discourse do
|
|
subject(:model) { described_class.new(model_name, max_tokens: max_tokens) }
|
|
|
|
let(:model_name) { "bart-large-cnn-samsum" }
|
|
let(:max_tokens) { 20 }
|
|
|
|
let(:content) do
|
|
{
|
|
resource_path: "/t/1/POST_NUMBER",
|
|
content_title: "This is a title",
|
|
contents: [{ poster: "asd", id: 1, text: "This is a text" }],
|
|
}
|
|
end
|
|
|
|
before { SiteSetting.ai_summarization_discourse_service_api_endpoint = "https://test.com" }
|
|
|
|
def stub_request(prompt, response)
|
|
WebMock
|
|
.stub_request(
|
|
:post,
|
|
"#{SiteSetting.ai_summarization_discourse_service_api_endpoint}/api/v1/classify",
|
|
)
|
|
.with(body: JSON.dump(model: model_name, content: prompt))
|
|
.to_return(status: 200, body: JSON.dump(summary_text: response))
|
|
end
|
|
|
|
def expected_messages(contents, opts)
|
|
contents.reduce("") do |memo, item|
|
|
memo += "(#{item[:id]} #{item[:poster]} said: #{item[:text]} "
|
|
end
|
|
end
|
|
|
|
def as_chunk(item)
|
|
{ ids: [item[:id]], summary: "(#{item[:id]} #{item[:poster]} said: #{item[:text]} " }
|
|
end
|
|
|
|
describe "#summarize_in_chunks" do
|
|
context "when the content fits in a single chunk" do
|
|
it "performs a request to summarize" do
|
|
opts = content.except(:contents)
|
|
|
|
stub_request(expected_messages(content[:contents], opts), "This is summary 1")
|
|
|
|
chunks = content[:contents].map { |c| as_chunk(c) }
|
|
summarized_chunks = model.summarize_in_chunks(chunks, opts).map { |c| c[:summary] }
|
|
|
|
expect(summarized_chunks).to contain_exactly("This is summary 1")
|
|
end
|
|
end
|
|
|
|
context "when the content fits in multiple chunks" do
|
|
it "performs a request for each one to summarize" do
|
|
content[:contents] << {
|
|
poster: "asd2",
|
|
id: 2,
|
|
text: "This is a different text to summarize",
|
|
}
|
|
opts = content.except(:contents)
|
|
|
|
content[:contents].each_with_index do |item, idx|
|
|
stub_request(expected_messages([item], opts), "This is summary #{idx + 1}")
|
|
end
|
|
|
|
chunks = content[:contents].map { |c| as_chunk(c) }
|
|
summarized_chunks = model.summarize_in_chunks(chunks, opts).map { |c| c[:summary] }
|
|
|
|
expect(summarized_chunks).to contain_exactly("This is summary 1", "This is summary 2")
|
|
end
|
|
end
|
|
end
|
|
|
|
describe "#concatenate_summaries" do
|
|
it "combines all the different summaries into a single one" do
|
|
messages = ["summary 1", "summary 2"].join("\n")
|
|
|
|
stub_request(messages, "concatenated summary")
|
|
|
|
expect(model.concatenate_summaries(["summary 1", "summary 2"])).to eq("concatenated summary")
|
|
end
|
|
end
|
|
|
|
describe "#summarize_with_truncation" do
|
|
let(:max_tokens) { 9 }
|
|
|
|
it "truncates the context to meet the token limit" do
|
|
opts = content.except(:contents)
|
|
|
|
stub_request("( 1 asd said : this is", "truncated summary")
|
|
|
|
expect(model.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
|
|
end
|
|
end
|
|
end
|