Fix: update Vertex AI model path handling - #21
Merged
Conversation
- Remove query params from Vertex AI websocket path - Add model_prefix for Vertex AI models with project/location path
Add support for GOOGLE_APPLICATION_CREDENTIALS_JSON environment variable which contains the service account JSON content directly. This is useful for containerized environments (Heroku, Fly.io, etc.) where mounting credential files is not practical.
Add documentation for the new GOOGLE_APPLICATION_CREDENTIALS_JSON environment variable option for containerized environments.
This commit introduces support for GOOGLE_APPLICATION_CREDENTIALS_JSON, allowing service account credentials to be passed directly as JSON content via environment variables. This serves as a gemini_ex extension to standard Application Default Credentials (ADC) for easier deployment in containerized environments. Key changes: - Added GOOGLE_APPLICATION_CREDENTIALS_JSON as the highest priority in the ADC discovery chain. - Upgraded the Vertex AI Live WebSocket endpoint from v1beta1 to v1. - Refactored service account JSON parsing to use a map-based atom key conversion for better reliability. - Improved model normalization in Gemini.Types.Live.Setup to handle fully qualified project resource names without duplication. - Updated documentation regarding auth priority, billing requirements for Vertex Live, and versioning differences between Gemini and Vertex. - Refactored test suites to isolate environment variables by setting async: false and adding cleanup logic. - Added RUN_BILLED_VERTEX_LIVE_TESTS flag to gate tests that require active GCP billing. - Adjusted rate limiter tests to reduce timing flakiness during backoff assertions.
Introduces Gemini.ModelRegistry as the central source of truth for model capabilities, modalities, and version aliases. This change improves the reliability of model selection, especially for the Live API, by moving away from static lists to a capability-based lookup system. Core enhancements: - Introduced Gemini.ModelRegistry to track features like thinking, live_api, and structured outputs across the model catalog. - Refactored Gemini.Live.Models.resolve/2 to use registry metadata combined with runtime API discovery for smarter fallback behavior. - Updated Gemini.Client.HTTP to allow per-request authentication overrides, enabling dynamic switching of API keys or GCP projects. - Improved Vertex AI compatibility by handling publisherModels payloads and normalizing full resource names in the Coordinator. - Expanded the configuration to support Gemini 3.1 Pro, updated Live API defaults, and added specialized models for image/video generation. - Added helper functions to Gemini.Config for capability checks and metadata retrieval. - Fixed model path building for Vertex AI to correctly support location and publisher prefixes. This update ensures that as new models are rolled out, the library can automatically discover and utilize compatible versions based on requested capabilities.
…ncements This release introduces a centralized Model Registry, expands authentication options for containerized environments, and improves Vertex AI integration consistency across the library. Model Registry and Configuration: - Added Gemini.ModelRegistry for tracking model capabilities, modalities, and support states. - Implemented capability checks for live API, structured outputs, and modality-based model selection. - Added Config functions for model metadata lookup and capability filtering. Authentication and ADC: - Added support for GOOGLE_APPLICATION_CREDENTIALS_JSON to allow service account configuration via environment variables. - Updated HTTP client to respect per-request authentication overrides, including API keys, project IDs, and service accounts. - Improved service account JSON parsing with an explicit key allowlist. Vertex AI and Live API: - Upgraded Vertex Live API endpoint to v1. - Standardized Vertex AI resource paths for model listing and retrieval. - Normalized response shapes between Gemini and Vertex AI model endpoints. - Improved Live API model resolution to prioritize registry-backed candidates based on modality. Generation API Improvements: - Updated Video generation request schemas for compatibility with latest API requirements, including Veo 3.1 support. - Refactored person generation and resolution parameters in video requests. - Standardized Vertex AI configuration validation for Image and Video APIs. - Improved streaming response collection with configurable timeouts. Cleanup and Maintenance: - Removed approximately 50,000 lines of internal planning documents and legacy development artifacts. - Added integration tests for ADC-authenticated Live API sessions. - Fixed path handling in the HTTP client to prevent double-prefixing of absolute URLs.
This update introduces a centralized Model Registry, upgrades the Vertex Live API to v1, and aligns Video Generation with Veo 3.1 requirements. Model Registry and Resolution: - Introduced ModelRegistry to track capabilities and modalities. - Live API model resolution now uses registry-backed auth-aware selection. - Separate candidate lists implemented for gemini vs vertex_ai strategies. - Updated default models to native audio versions for Live sessions. Vertex AI and Authentication: - Upgraded Vertex Live API to v1 and corrected resource name formatting. - Fixed list_models pagination to use query parameters for Vertex AI. - Automated Vertex AI authentication for all Images API requests. - Revised ADC discovery order: JSON env -> Path env -> gcloud -> metadata. - Added RUN_BILLED_VERTEX_LIVE_TESTS gate to prevent accidental GCP costs. Video and Image Generation: - Refactored Video API to use flattened request keys for Veo 3.1 compatibility. - Updated binary encoding from bytesBase64Encoded to inlineData format. - Changed person_generation default to :dont_allow and moved to snake_case. - Retained legacy fields in structs for BC but excluded them from API payloads. Streaming and Reliability: - Added :stream_timeout option (default 60s) to Gemini.stream_generate/2. - Implemented automatic cleanup of orphaned streams on timeout via Coordinator. - UnifiedManager now passes advanced streaming knobs to the HTTP client. Documentation and Maintenance: - Relocated guide files to the guides/ directory and updated all internal links. - Updated README to reflect new v0.10.0 features and configuration options. - Normalized publisherModels responses to match Gemini API shapes.
Owner
Author
|
Released as https://hex.pm/packages/gemini_ex/0.10.0 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR mirrors the latest maintainer commit from #20 (fork branch) so it can be reviewed/merged without fork-write permissions.
Why: PR #20 has maintainer edits disabled (maintainerCanModify=false) and the branch owner must enable edits to accept direct pushes.
Includes commit: 716f9e4
Supersedes: #20