Skip to content

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

FIX: Handle Gemini errors and blocked prompts reported in successful responses (#44262)

Régis Hanol3b7add4745d2
diff --git a/plugins/discourse-ai/app/jobs/scheduled/topics_locale_detection_backfill.rb b/plugins/discourse-ai/app/jobs/scheduled/topics_locale_detection_backfill.rb
index 2b9e79fcb1f..2048b5d7ab9 100644
--- a/plugins/discourse-ai/app/jobs/scheduled/topics_locale_detection_backfill.rb
+++ b/plugins/discourse-ai/app/jobs/scheduled/topics_locale_detection_backfill.rb
@@ -31,6 +31,8 @@ module Jobs
return if topics.empty?
topics.each do |topic|
+ next if !DiscourseAi::Translation::TopicLocalizer.has_relocalize_quota?(topic, "")
+
DiscourseAi::Translation::TopicLocaleDetector.detect_locale(topic)
rescue FinalDestination::SSRFDetector::LookupFailedError
# do nothing, there are too many sporadic lookup failures
diff --git a/plugins/discourse-ai/lib/completions/endpoints/gemini.rb b/plugins/discourse-ai/lib/completions/endpoints/gemini.rb
index 486462bfe4f..568e98b93f9 100644
--- a/plugins/discourse-ai/lib/completions/endpoints/gemini.rb
+++ b/plugins/discourse-ai/lib/completions/endpoints/gemini.rb
@@ -409,6 +409,7 @@ module DiscourseAi
def decode(chunk)
json = JSON.parse(chunk, symbolize_names: true)
+ check_for_failure!(json)
update_usage(json)
candidate = json.dig(:candidates, 0)
@@ -423,6 +424,7 @@ module DiscourseAi
streaming_decoder
.decode(chunk)
.map do |parsed|
+ check_for_failure!(parsed)
update_usage(parsed)
candidate = parsed.dig(:candidates, 0)
parts = candidate&.dig(:content, :parts)
@@ -434,6 +436,17 @@ module DiscourseAi
.compact
end
+ # Gemini reports errors and blocked prompts inside a successful (HTTP 200) response.
+ def check_for_failure!(parsed)
+ if (error = parsed[:error])
+ raise CompletionFailed, error[:message].presence || error.to_json
+ end
+
+ if (block_reason = parsed.dig(:promptFeedback, :blockReason))
+ Rails.logger.warn("#{self.class.name}: prompt blocked by provider: #{block_reason}")
+ end
+ end
+
def decode_parts(parts, batch_token:, streaming: false)
idx = -1
(parts || []).each_with_object([]) do |part, result|
diff --git a/plugins/discourse-ai/spec/jobs/scheduled/topics_locale_detection_backfill_spec.rb b/plugins/discourse-ai/spec/jobs/scheduled/topics_locale_detection_backfill_spec.rb
index 4cc96f51992..05ebd7157ee 100644
--- a/plugins/discourse-ai/spec/jobs/scheduled/topics_locale_detection_backfill_spec.rb
+++ b/plugins/discourse-ai/spec/jobs/scheduled/topics_locale_detection_backfill_spec.rb
@@ -77,6 +77,20 @@ describe Jobs::TopicsLocaleDetectionBackfill do
expect { job.execute({ limit: 10 }) }.not_to raise_error
end
+ context "when relocalize quota is exhausted" do
+ before do
+ DiscourseAi::Translation::TopicLocalizer::MAX_QUOTA_PER_DAY.times do
+ DiscourseAi::Translation::TopicLocalizer.has_relocalize_quota?(topic, "")
+ end
+ end
+
+ it "skips locale detection for topics that have exceeded quota" do
+ DiscourseAi::Translation::TopicLocaleDetector.expects(:detect_locale).never
+
+ job.execute({})
+ end
+ end
+
it "logs a summary after running" do
DiscourseAi::Translation::TopicLocaleDetector.stubs(:detect_locale)
DiscourseAi::Translation::VerboseLogger.expects(:log).with(includes("Detected 1 topic locales"))
diff --git a/plugins/discourse-ai/spec/lib/completions/endpoints/gemini_spec.rb b/plugins/discourse-ai/spec/lib/completions/endpoints/gemini_spec.rb
index fbe3bba600e..178567f7dca 100644
--- a/plugins/discourse-ai/spec/lib/completions/endpoints/gemini_spec.rb
+++ b/plugins/discourse-ai/spec/lib/completions/endpoints/gemini_spec.rb
@@ -1559,6 +1559,38 @@ RSpec.describe DiscourseAi::Completions::Endpoints::Gemini do
)
end
+ it "raises when an error event arrives mid-stream instead of returning the partial reply" do
+ response = <<~TEXT
+ data: {"candidates": [{"content": {"parts": [{"text": "Partial"}],"role": "model"}}]}
+
+ data: {"error": {"code": 503,"message": "This model is currently experiencing high demand.","status": "UNAVAILABLE"}}
+
+ TEXT
+
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+ url = "#{model.url}:streamGenerateContent?alt=sse&key=123"
+ stub_request(:post, url).to_return(status: 200, body: response)
+
+ expect { llm.generate("Hello", user: user) { |_| } }.to raise_error(
+ DiscourseAi::Completions::Endpoints::Base::CompletionFailed,
+ /high demand/,
+ )
+ end
+
+ it "logs the block reason when the prompt is blocked" do
+ response = <<~TEXT
+ data: {"promptFeedback": {"blockReason": "PROHIBITED_CONTENT"},"usageMetadata": {"promptTokenCount": 417,"totalTokenCount": 417}}
+
+ TEXT
+
+ llm = DiscourseAi::Completions::Llm.proxy(model)
+ url = "#{model.url}:streamGenerateContent?alt=sse&key=123"
+ stub_request(:post, url).to_return(status: 200, body: response)
+ Rails.logger.expects(:warn).with(includes("PROHIBITED_CONTENT"))
+
+ expect(llm.generate("Hello", user: user) { |_| }).to eq("")
+ end
+
it "Can correctly handle streamed responses even if they are chunked badly" do
data = +""
data << "da|ta: |"