Roman Rizzi 9a79afcdbf
DEV: Better strategies for summarization (#88)
* DEV: Better strategies for summarization

The strategy responsibility needs to be "Given a collection of texts, I know how to summarize them most efficiently, using the minimum amount of requests and maximizing token usage".

There are different token limits for each model, so it all boils down to two different strategies:

Fold all these texts into a single one, doing the summarization in chunks, and then build a summary from those.
Build it by combining texts in a single prompt, and truncate it according to your token limits.

While the latter is less than ideal, we need it for "bart-large-cnn-samsum" and "flan-t5-base-samsum", both with low limits. The rest will rely on folding.

* Expose summarized chunks to users
2023-06-27 12:26:33 -03:00

117 lines
3.7 KiB
Ruby

# frozen_string_literal: true
require_relative "../../../../support/anthropic_completion_stubs"
RSpec.describe DiscourseAi::Summarization::Models::Anthropic do
let(:model) { "claude-v1" }
let(:max_tokens) { 720 }
subject { described_class.new(model, max_tokens: max_tokens) }
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
def expected_messages(contents, opts)
base_prompt = <<~TEXT
Human: Summarize the following forum discussion inside the given <input> tag.
Include only the summary inside <ai> tags.
Try generating links as well the format is #{opts[:resource_path]}.
The discussion title is: #{opts[:content_title]}.
Don't use more than 400 words.
TEXT
text =
contents.reduce("") do |memo, item|
memo += "(#{item[:id]} #{item[:poster]} said: #{item[:text]} "
end
base_prompt += "<input>#{text}</input>\nAssistant:\n"
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)
AnthropicCompletionStubs.stub_response(
expected_messages(content[:contents], opts),
"<ai>This is summary 1</ai>",
)
summarized_chunks =
subject.summarize_in_chunks(content[:contents], 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|
AnthropicCompletionStubs.stub_response(
expected_messages([item], opts),
"<ai>This is summary #{idx + 1}</ai>",
)
end
summarized_chunks =
subject.summarize_in_chunks(content[:contents], 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 = <<~TEXT
Human: Concatenate the following disjoint summaries inside the given input tags, creating a cohesive narrative.
Include only the summary inside <ai> tags.
<input>summary 1</input>
<input>summary 2</input>
Assistant:
TEXT
AnthropicCompletionStubs.stub_response(messages, "<ai>concatenated summary</ai>")
expect(subject.concatenate_summaries(["summary 1", "summary 2"])).to eq(
"concatenated summary",
)
end
end
describe "#summarize_with_truncation" do
let(:max_tokens) { 709 }
it "truncates the context to meet the token limit" do
opts = content.except(:contents)
instructions = <<~TEXT
Human: Summarize the following forum discussion inside the given <input> tag.
Include only the summary inside <ai> tags.
Try generating links as well the format is #{opts[:resource_path]}.
The discussion title is: #{opts[:content_title]}.
Don't use more than 400 words.
<input>(1 asd said: This is a</input>
Assistant:
TEXT
AnthropicCompletionStubs.stub_response(instructions, "<ai>truncated summary</ai>")
expect(subject.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
end
end
end