Make consistent OPENAI base url

[jasperblues]

Fix inconsistent /v1 in OpenAI-compatible base URLs

Problem

Spring AI's OpenAiApi.builder().baseUrl() appends /v1 internally. The openai-java SDK (OpenAIOkHttpClient) does NOT — it uses the URL as-is. Both read OPENAI_BASE_URL, which is set without /v1 for Spring AI compatibility. This mismatch causes issues in several places.

Fixes

• OpenAiCompatibleModelFactory.mistral() — base URL included /v1, producing double /v1/v1/ through Spring AI
• OpenAiCompatibleModelFactory.byok() KDoc — example URL included /v1, misleading for Spring AI consumers
• OpenAiCompatibleModelFactoryByokSpecTest — test example URL included /v1
• LLMOpenAiGuardRailsIntegrationIT — OpenAIOkHttpClient.fromEnv() used OPENAI_BASE_URL without /v1, producing 404s on the moderation endpoint

New file

• OpenAiCompatibleModelFactoryByokIT — IT tests that validate BYOK buildValidated() against real provider APIs (OpenAI, DeepSeek, Mistral, Gemini). Each test is gated by @EnabledIfEnvironmentVariable so it skips cleanly without the corresponding API key.

Details

OpenAiCompatibleModelFactory.mistral()

Before: ByokSpec("https://api.mistral.ai/v1", ...) — Spring AI appends /v1 internally, producing https://api.mistral.ai/v1/v1/chat/completions (confirmed 404 via curl).

After: ByokSpec("https://api.mistral.ai", ...) — Spring AI appends /v1, producing the correct https://api.mistral.ai/v1/chat/completions (confirmed 200 via curl and BYOK IT test).

LLMOpenAiGuardRailsIntegrationIT

Before: OpenAIOkHttpClient.fromEnv() — reads OPENAI_BASE_URL (spring AI compatible version of https://api.openai.com), SDK appends /moderations directly → hits https://api.openai.com/moderations → 404. This was masked by the previous try/catch pattern in testThinkingCreateObject — the NotFoundException was silently swallowed because it wasn't a GuardRailViolationException. Only surfaced after #1577 replaced the catch with assertThrows.

After: Explicit builder reads OPENAI_BASE_URL and appends /v1 → https://api.openai.com/v1, SDK appends /moderations → hits https://api.openai.com/v1/moderations → 200. Moderation correctly flags the content (violence category) and GuardRailViolationException is thrown.

Verified

• All 15 unit tests in embabel-agent-openai pass.
• OpenAiCompatibleModelFactoryByokIT — 3 passed (OpenAI, DeepSeek, Mistral), 1 skipped (Gemini, no key). Mistral confirmed working against the real API after the /v1 fix.
• LLMOpenAiGuardRailsIntegrationIT.testThinkingCreateObject now passes.

Fun fact

The guardrails test prompt mentions blowing up crocodiles in Florida. Southern Florida is actually the only place in the world where alligators and crocodiles coexist — the American crocodile (Crocodylus acutus) lives alongside the American alligator in the Everglades. The more you know.

[igordayen]

Thanks Jasper, so running IT time2time in bulk actually helps:)

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud