// AzureChatModel: wire bodies per dialect and parsing of the documented // examples (chat completion, stream transcript, response object). #include "rig.h" #include "vendors/azure/chat_model.h" using namespace molo::azure; using namespace az_test; namespace { // Documented chat completion example (OpenAI / Azure OpenAI reference). const char kChat[] = R"({"id":"chatcmpl-123","object":"chat.completion","created":1677652288,"model":"gpt-4o-mini","choices":[{"index":0,"message":{"role":"assistant","content":"Hello there, how may I assist you today?"},"finish_reason":"stop"}],"usage":{"prompt_tokens":9,"completion_tokens":12,"total_tokens":21}})"; // Documented stream transcript (chat.completion.chunk). const char kStream[] = "data: {\"id\":\"chatcmpl-123\",\"object\":\"chat.completion.chunk\",\"created\":1694268190,\"model\":\"gpt-4o-mini\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"finish_reason\":null}]}\n\n" "data: {\"id\":\"chatcmpl-123\",\"object\":\"chat.completion.chunk\",\"created\":1694268190,\"model\":\"gpt-4o-mini\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"Hello\"},\"finish_reason\":null}]}\n\n" "data: {\"id\":\"chatcmpl-123\",\"object\":\"chat.completion.chunk\",\"created\":1694268190,\"model\":\"gpt-4o-mini\",\"choices\":[{\"index\":0,\"delta\":{},\"finish_reason\":\"stop\"}]}\n\n" "data: [DONE]\n\n"; // Documented Responses example. const char kResponse[] = R"({"id":"resp_67ccd2bed1ec8190b14f964abc0542670bb6a6b452d3795b","object":"response","created_at":1741476542,"status":"completed","model":"gpt-4.1-2025-04-14","output":[{"type":"message","id":"msg_67ccd2bf17f0819081ff3bb2cf6508e60bb6a6b452d3795b","status":"completed","role":"assistant","content":[{"type":"output_text","text":"In a peaceful grove beneath a silver moon, a unicorn named Lumina discovered a hidden pool.","annotations":[]}]}],"usage":{"input_tokens":36,"output_tokens":87,"total_tokens":123}})"; molo::llm::ChatRequest req(const std::string& model = "gpt4o") { molo::llm::ChatRequest r; r.model = model; r.system = "Be brief."; r.messages.push_back(molo::llm::Message::user("Say hi")); r.max_tokens = 32; return r; } } // namespace TEST(azure_chat_model, generate_over_chat_completions_v1) { Rig rig(key_settings()); rig.openai().answer = kChat; AzureChatModel m(rig.client); ASSERT_EQ(m.vendor(), std::string("azure")); auto out = m.generate(req()); ASSERT_EQ(out.text(), std::string("Hello there, how may I assist you today?")); ASSERT_EQ(out.finish_reason, std::string("stop")); ASSERT_EQ(out.usage.total_tokens, 21); ASSERT_EQ(rig.last().target, std::string("/openai/v1/chat/completions")); Json b = rig.last_body(); ASSERT_EQ(b.find("model")->as_string(), std::string("gpt4o")); ASSERT_TRUE(b.has("messages")); ASSERT_FALSE(b.has("stream")); } TEST(azure_chat_model, generate_over_the_dated_api_uses_the_deployment_url) { Rig rig(key_settings("2024-10-21")); rig.openai().answer = kChat; AzureChatModel m(rig.client); m.generate(req()); ASSERT_EQ(rig.last().target, std::string("/openai/deployments/gpt4o/chat/completions?api-version=2024-10-21")); ASSERT_FALSE(rig.last_body().has("model")); } TEST(azure_chat_model, stream_decodes_the_documented_transcript) { Rig rig(key_settings()); rig.openai().content_type = "text/event-stream"; rig.openai().answer = kStream; AzureChatModel m(rig.client); std::string text, finish; m.stream(req(), [&](const molo::llm::ChatChunk& c) { text += c.text_delta; if (!c.finish_reason.empty()) finish = c.finish_reason; return true; }); ASSERT_EQ(text, std::string("Hello")); ASSERT_EQ(finish, std::string("stop")); Json b = rig.last_body(); ASSERT_TRUE(b.find("stream")->as_bool()); ASSERT_TRUE(b.find("stream_options")->find("include_usage")->as_bool()); } TEST(azure_chat_model, stream_stops_when_the_consumer_says_so) { Rig rig(key_settings()); rig.openai().content_type = "text/event-stream"; rig.openai().answer = kStream; AzureChatModel m(rig.client); int calls = 0; m.stream(req(), [&](const molo::llm::ChatChunk&) { return ++calls < 2; }); ASSERT_EQ(calls, 2); } TEST(azure_chat_model, responses_dialect) { Rig rig(key_settings()); rig.openai().answer = kResponse; ChatModelConfig cfg; cfg.dialect = Dialect::kResponses; AzureChatModel m(rig.client, cfg); auto out = m.generate(req("gpt41")); ASSERT_TRUE(out.text().rfind("In a peaceful grove", 0) == 0); ASSERT_EQ(out.usage.total_tokens, 123); ASSERT_EQ(rig.last().target, std::string("/openai/v1/responses")); ASSERT_EQ(rig.last_body().find("model")->as_string(), std::string("gpt41")); } TEST(azure_chat_model, foundry_dialect_goes_to_the_model_inference_host) { Rig rig(key_settings()); rig.foundry().answer = kChat; ChatModelConfig cfg; cfg.dialect = Dialect::kFoundry; cfg.extra_parameters = "drop"; AzureChatModel m(rig.client, cfg); auto out = m.generate(req("Phi-4")); ASSERT_TRUE(out.text().rfind("Hello there", 0) == 0); const Seen& s = rig.foundry().seen.back(); ASSERT_EQ(s.target, std::string("/models/chat/completions?api-version=2024-05-01-preview")); ASSERT_EQ(s.headers.get_or("extra-parameters", ""), std::string("drop")); Json b = Json::parse(s.body); ASSERT_TRUE(b.has("max_tokens")); ASSERT_FALSE(b.has("max_completion_tokens")); ASSERT_EQ(rig.openai().seen.size(), static_cast(0)); } TEST(azure_chat_model, validation_before_traffic) { Rig rig(key_settings()); AzureChatModel m(rig.client); molo::llm::ChatRequest r = req(); r.model = ""; ASSERT_TRUE(rejects([&] { m.generate(r); }, "deployment")); r.model = "bad name"; ASSERT_TRUE(rejects([&] { m.generate(r); }, "deployment")); r = req(); r.messages.clear(); ASSERT_TRUE(rejects([&] { m.generate(r); }, "messages")); ASSERT_EQ(rig.openai().seen.size(), static_cast(0)); } TEST(azure_chat_model, not_configured_is_a_503_before_traffic) { Settings none; none.endpoint = "https://res.openai.azure.com"; Rig rig(none); AzureChatModel m(rig.client); ASSERT_FALSE(m.configured()); try { m.generate(req()); FAIL("expected not configured"); } catch (const molo::llm::NotConfiguredError& e) { ASSERT_EQ(e.http_status(), 503); } ASSERT_EQ(rig.openai().seen.size(), static_cast(0)); } TEST(azure_chat_model, wire_body_is_exposed_without_traffic) { Rig rig(key_settings()); AzureChatModel m(rig.client); Json b = m.wire_body(req(), false); ASSERT_EQ(b.find("model")->as_string(), std::string("gpt4o")); ASSERT_EQ(rig.openai().seen.size(), static_cast(0)); }