Skip to content

v1.1.12

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Jul 04:35
· 36 commits to main since this release
v1.1.12
This tag was signed with the committer’s verified signature.
MitulShah1 Mitul Shah
GPG key ID: 9A023A4CFFD1D25D
Verified
Learn about vigilant mode.
77d567b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

[1.1.12] — 2026-07-04

Enterprise-endpoint & Cohere fidelity release. The third provider-readiness remediation phase aligns the Vertex AI, Azure OpenAI, Azure AI Foundry, Databricks, and Cohere providers on request/response correctness and consolidates their hand-rolled HTTP onto the shared OpenAI-compatible helpers. No breaking API changes relative to v1.1.11.

Fixed

  • Vertex AI model publisher prefix: chat and streaming requests now send first-party models with the required google/ publisher prefix (an existing publisher prefix — e.g. a Model Garden meta/ model — is left intact), fixing requests the OpenAI-compatible endpoint rejected.
  • Cohere vision: multimodal image_url content is now forwarded to Cohere as content blocks instead of being silently dropped.
  • Cohere streaming usage: the message-end token counts are now surfaced on a stream chunk (previously parsed and discarded).
  • Azure OpenAI response provider: chat responses now carry the provider name, which the hand-rolled decoder dropped.
  • Databricks retired model: databricks-claude-3-7-sonnet is removed from the advertised model list (Databricks retired it).

Changed

  • Azure AI Foundry endpoint (operator-visible): the provider now targets the GA {endpoint}/openai/v1/chat/completions route (OpenAI-shaped, no api-version) instead of the Model Inference /models route the vendor is retiring, and sends extra-parameters: drop so a non-schema field is ignored rather than rejected.
  • Vertex AI credentials (operator-visible): when neither an API key nor service-account JSON is configured, the provider now falls back to Application Default Credentials (GOOGLE_APPLICATION_CREDENTIALS, gcloud, workload identity, or the GCE/GKE metadata server), so managed environments authenticate without an explicit key.
  • Shared OpenAI-compatible request path: Vertex AI, Azure OpenAI, Azure AI Foundry, and Databricks now route chat/embeddings through the shared openaicompat helpers, so their error handling (via core.APIError) and request/response shape stay consistent. Base-URL validation is enforced at construction for all five providers, and a tuned HTTP transport preset was added for Databricks serving-endpoint cold starts.
  • Model catalog refresh: the Vertex AI static fallback drops the retired gemini-2.0-flash; a recognized image size now maps to the Imagen aspect ratio.

Contributors

Thanks to everyone who shipped this release:

Full changelog: https://github.com/ferro-labs/ai-gateway/blob/v1.1.12/CHANGELOG.md

Contributors

MitulShah1
Assets 8
Loading