Skip to content

A platform for community discussion. Free, open, simple.

DEV: Let the fake LLM answer structured output requests

Régis Hanolf6e53109e381
diff --git a/plugins/discourse-ai/lib/completions/endpoints/base.rb b/plugins/discourse-ai/lib/completions/endpoints/base.rb
index 7c143194194..4a7ccc5730b 100644
--- a/plugins/discourse-ai/lib/completions/endpoints/base.rb
+++ b/plugins/discourse-ai/lib/completions/endpoints/base.rb
@@ -18,6 +18,10 @@ module DiscourseAi
# Stored in AiApiAuditLog#request_attempts for retried network failures,
# which do not have an HTTP status code. 0 is intentionally outside the HTTP range.
NETWORK_ERROR_RETRY_STATUS = 0
+
+ def self.schema_properties(response_format)
+ response_format&.dig(:json_schema, :schema, :properties)
+ end
RETRIABLE_NETWORK_ERRORS = [
Net::OpenTimeout,
Net::ReadTimeout,
@@ -389,12 +393,10 @@ module DiscourseAi
end
def build_structured_output(model_params)
- return if model_params[:response_format].blank?
-
- schema_properties = model_params[:response_format].dig(:json_schema, :schema, :properties)
- return if schema_properties.blank?
+ properties = self.class.schema_properties(model_params[:response_format])
+ return if properties.blank?
- DiscourseAi::Completions::StructuredOutput.new(schema_properties)
+ DiscourseAi::Completions::StructuredOutput.new(properties)
end
def perform_completion_request_with_retries(
diff --git a/plugins/discourse-ai/lib/completions/endpoints/fake.rb b/plugins/discourse-ai/lib/completions/endpoints/fake.rb
index fda154e8a2a..6177f0de0f2 100644
--- a/plugins/discourse-ai/lib/completions/endpoints/fake.rb
+++ b/plugins/discourse-ai/lib/completions/endpoints/fake.rb
@@ -76,8 +76,34 @@ module DiscourseAi
@fake_content = content
end
- def self.fake_content
- @fake_content || STOCK_CONTENT
+ def self.fake_content(response_format = nil)
+ return @fake_content if @fake_content
+
+ properties = schema_properties(response_format)
+ return STOCK_CONTENT if properties.blank?
+
+ properties.to_h { |key, property| [key, stock_value(key, property)] }.to_json
+ end
+
+ def self.json_object?(content)
+ content.is_a?(String) && JSON.parse(content).is_a?(Hash)
+ rescue JSON::ParserError
+ false
+ end
+
+ def self.stock_value(key, property)
+ case property[:type].to_s
+ when "integer"
+ 42
+ when "number"
+ 4.2
+ when "boolean"
+ true
+ when "array"
+ [stock_value(key, property[:items] || {})]
+ else
+ "fake #{key}"
+ end
end
def self.delays
@@ -132,9 +158,12 @@ module DiscourseAi
# guard memory in test
self.class.previous_calls.shift if self.class.previous_calls.length > 10
- content = self.class.fake_content
+ content = self.class.fake_content(model_params[:response_format])
content = content.shift if content.is_a?(Array)
+ structured_output = build_structured_output(model_params) if self.class.json_object?(
+ content,
+ )
if block_given?
if content.is_a?(DiscourseAi::Completions::ToolCall)
@@ -171,12 +200,21 @@ module DiscourseAi
break if cancel
content << chunk
- yield(chunk, cancel_proc)
+ next yield(chunk, cancel_proc) if structured_output.nil?
+
+ structured_output << chunk
+ yield(structured_output, cancel_proc)
end
end
+ elsif structured_output
+ structured_output << content
end
- content
+ return content if structured_output.nil?
+
+ structured_output.finish
+ yield(structured_output, -> {}) if block_given?
+ structured_output
end
end
end
diff --git a/plugins/discourse-ai/spec/lib/completions/endpoints/fake_spec.rb b/plugins/discourse-ai/spec/lib/completions/endpoints/fake_spec.rb
new file mode 100644
index 00000000000..6c51add864c
--- /dev/null
+++ b/plugins/discourse-ai/spec/lib/completions/endpoints/fake_spec.rb
@@ -0,0 +1,112 @@
+# frozen_string_literal: true
+
+RSpec.describe DiscourseAi::Completions::Endpoints::Fake do
+ fab!(:llm_model, :fake_model)
+
+ let(:response_format) do
+ {
+ type: "json_schema",
+ json_schema: {
+ name: "reply",
+ schema: {
+ type: "object",
+ properties: {
+ verdict: {
+ type: "string",
+ },
+ confident: {
+ type: "boolean",
+ },
+ attempts: {
+ type: "integer",
+ },
+ reasons: {
+ type: "array",
+ items: {
+ type: "string",
+ },
+ },
+ },
+ },
+ },
+ }
+ end
+
+ before { enable_current_plugin }
+
+ it "answers a structured request with a payload shaped like the schema" do
+ result =
+ llm_model.to_llm.generate(
+ "hello",
+ user: Discourse.system_user,
+ feature_name: "test",
+ response_format: response_format,
+ )
+
+ expect(result).to be_a(DiscourseAi::Completions::StructuredOutput)
+ expect(JSON.parse(result.to_s)).to eq(
+ "verdict" => "fake verdict",
+ "confident" => true,
+ "attempts" => 42,
+ "reasons" => ["fake reasons"],
+ )
+ end
+
+ it "streams the structured payload" do
+ partials = []
+
+ llm_model
+ .to_llm
+ .generate(
+ "hello",
+ user: Discourse.system_user,
+ feature_name: "test",
+ response_format: response_format,
+ ) { |partial| partials << partial }
+
+ expect(partials.last).to be_a(DiscourseAi::Completions::StructuredOutput)
+ expect(partials.last.read_buffered_property(:verdict)).to eq("fake verdict")
+ end
+
+ it "keeps returning explicitly set content", :aggregate_failures do
+ described_class.with_fake_content({ verdict: "reject" }.to_json) do
+ result =
+ llm_model.to_llm.generate(
+ "hello",
+ user: Discourse.system_user,
+ feature_name: "test",
+ response_format: response_format,
+ )
+
+ expect(result.to_s).to eq({ verdict: "reject" }.to_json)
+ end
+
+ described_class.with_fake_content("a plain summary") do
+ result =
+ llm_model.to_llm.generate(
+ "hello",
+ user: Discourse.system_user,
+ feature_name: "test",
+ response_format: response_format,
+ )
+
+ expect(result).to eq("a plain summary")
+ end
+ end
+
+ it "answers with the stock content when no schema is requested", :aggregate_failures do
+ expect(
+ llm_model.to_llm.generate("hello", user: Discourse.system_user, feature_name: "test"),
+ ).to include("Discourse Markdown Styles Showcase")
+ expect(
+ llm_model.to_llm.generate(
+ "hello",
+ user: Discourse.system_user,
+ feature_name: "test",
+ response_format: {
+ type: "json_object",
+ },
+ ),
+ ).to include("Discourse Markdown Styles Showcase")
+ end
+end