Changelog

Copy Markdown View Source

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.9.0 - 2026-01-23

Added

Live API (Bidirectional Streaming)

  • WebSocket Infrastructure: New Gemini.Client.WebSocket module using :gun for HTTP/2 + TLS connections with support for both Gemini API and Vertex AI endpoints
  • Session Management: Gemini.Live.Session GenServer for managing WebSocket connection lifecycle with callback-based message routing (on_message, on_error, on_close)
  • Model Resolution: Gemini.Live.Models for runtime model availability detection based on API key capabilities and regional rollout status
  • Function Calling: Full tool call/response protocol with async scheduling modes (INTERRUPT, WHEN_IDLE, SILENT)
  • Session Resumption: Handle-based reconnection support for preserving context across WebSocket disconnections
  • Context Window Compression: Sliding window configuration for managing long conversations
  • Ephemeral Tokens: Client-side WebSocket authentication token generation
  • Audio Utilities: Gemini.Live.Audio module for PCM audio handling with native audio features
  • Native Audio Features: Affective dialog, proactivity configuration, and thinking budgets
  • GoAway Handling: Graceful session termination message support

Live API Type Definitions

  • BidiGenerateContentSetup and SetupComplete for session handshake
  • ClientContent and RealtimeInput for client messages
  • ServerContent, ToolCall, and ToolResponse for server messages
  • Audio transcription, voice activity detection, and go-away types
  • Media resolution and thinking configuration types

Gap Features (Python SDK v1.57.0-v1.60.0 Parity)

  • RegisterFiles API: GCS file registration support (Gemini API only)
  • ModelArmorConfig: Enterprise content filtering with mutual exclusivity validation against safety settings (Vertex AI only)
  • ImageConfig: New aspect ratios (4:5, 5:4) for social media image generation
  • FileSearchCallContent: New type for file search interactions
  • Gemini 3 Pro: Tokenizer mapping support

OTP Supervision

  • TaskSupervisor: Centralized Gemini.TaskSupervisor in application supervision tree for managed async tasks
  • Supervised task creation in ConcurrencyGate, HTTPStreaming, ToolOrchestrator, and Interactions modules

Telemetry

  • Session lifecycle events (start, stop)
  • Message throughput metrics
  • Tool execution tracking
  • Error state observability
  • WebSocket connection lifecycle events

Examples

  • Example 11: Multi-turn text chat demo
  • Example 12: Audio input and output streaming demo
  • Example 13: Session resumption and context preservation demo
  • Example 14: Function calling with telemetry observation
  • examples/live_api_demo.exs: Basic Live API usage
  • examples/live_function_calling.exs: Tool use demonstration
  • examples/run_all.sh: Script for running all examples

Changed

  • Live API Version: Upgraded from v1alpha to v1beta as default (v1alpha still supported for advanced features)
  • WebSocket Handshake: Updated to use HTTP/1.1 for Live API compatibility
  • WebSocket Frames: Added support for binary JSON frames
  • Model Name Normalization: Automatic models/ prefix addition when missing
  • Text Extraction: Refactored to support decoding inline text blobs and application/json parts
  • Live API Configuration: to_api conversions now handle plain maps for dynamic options
  • Setup Types: Extended to include media resolution and thinking configurations

Fixed

  • Empty text parts validation in content generation
  • Potential deadlock in Live.Session resolved via async tool response callbacks
  • Session resumption handle passing during setup phase
  • Bare spawn and spawn_link calls replaced with supervised tasks (prevents resource leaks)

Security

  • Query Parameter Redaction: Sensitive parameters (API keys, access tokens) now redacted in WebSocket connection logging

Documentation

  • Comprehensive Live API guide (docs/guides/live_api.md)
  • Updated README.md with Live API and new features sections
  • Updated examples/README.md with new example descriptions
  • Added OTP_AUDIT.md for process creation site tracking
  • Added OTP_REMEDIATION_PLAN.md for supervision strategy documentation
  • Updated mix.exs :extras with new guides
  • Reorganized and streamlined guides

0.8.8 - 2025-12-29

Added

  • Paid Tier 3 Rate Limit Profile: New :paid_tier_3 profile for maximum throughput workloads

    • Qualification: >$1,000 total spend + 30 days since payment
    • Settings: 30 max concurrency, 4M token budget, 100ms backoff, 50 adaptive ceiling
    • Added corresponding test coverage in rate_limiter_test.exs

  • Use-Case Model Aliases: Implemented Gemini.Config.model_for_use_case/2 and related functions

    • model_for_use_case/2: Resolve use-case atoms to model strings with optional API validation
    • use_case_token_budget/1: Get recommended token budget for a use-case
    • resolved_use_case_models/0: Get all use-case to model mappings
    • available_use_cases/0: List available use-case aliases
    • Use cases: :cache_context (gemini-2.5-flash), :report_section (gemini-2.5-pro), :fast_path (gemini-2.5-flash-lite)

  • Shared Utility Modules: Consolidated duplicated helpers into dedicated modules

Changed

  • Rate Limit Documentation: Updated to reflect current Google API tier structure

    • Added tier qualifications table (Free, Tier 1-3 with spending thresholds)
    • Documented that rate limits are per-project (not per-API key)
    • Added note that RPD quotas reset at midnight Pacific time
    • Added links to AI Studio for viewing actual model-specific limits
    • Removed hardcoded RPM/TPM values (vary by model)

  • Unified Streaming Architecture: Removed redundant streaming managers

    • UnifiedManager is now the canonical implementation
    • Deleted legacy Manager and ManagerV2 modules
    • Updated examples and tests to use unified architecture

  • Authentication Interface: Renamed strategy/1 to Gemini.Auth.get_strategy/1

Removed

Technical

  • 13 modules updated to use shared MapHelpers and PollingHelpers utilities
  • Removed all __test_ prefixed helper functions from ContextCache, Coordinator, and Videos modules
  • Tests now focus on public module interfaces instead of internal helpers
  • Function extraction and modularization across API modules

Notes

  • All changes maintain backward compatibility for public API
  • Zero compilation warnings maintained
  • Documentation generates without warnings

0.8.7 - 2025-12-27

Changed

  • Code Quality Improvements: Extensive refactoring across the codebase for better maintainability

    • Simplified conditional logic patterns (replaced complex cond with clearer if/else)
    • Extracted helper functions to reduce function complexity and improve readability
    • Improved pattern matching and function composition
    • Replaced Enum.map + Enum.join patterns with Enum.map_join for efficiency
    • Better error handling patterns with reduced nesting
    • Enhanced function organization and naming consistency

  • Dependency Updates:

    • altar: 0.1.2 → 0.2.0 (ALTAR protocol improvements)
    • supertester: 0.3.1 → 0.4.0 (testing framework enhancements)
    • Added stream_data ~> 1.2.0 as explicit dependency

Technical

  • Function extraction and modularization throughout all API modules
  • Consistent error handling patterns across coordinators
  • Improved telemetry callback handling with safer invocation patterns
  • Better separation of concerns in authentication and streaming layers
  • Enhanced readability in type serialization and parsing logic

Notes

  • All changes are internal refactoring - no breaking changes to public API
  • Zero compilation warnings maintained
  • Full test suite passing (425+ tests)

0.8.6 - 2025-12-20

Changed

  • Default models updated to Gemini 2.5: All defaults now use current-generation models
    • Vertex AI default: gemini-2.0-flash-lite → gemini-2.5-flash-lite
    • Universal default: gemini-2.0-flash-lite → gemini-2.5-flash-lite
  • Updated all documentation examples to use gemini-2.5-flash instead of gemini-2.0-flash-exp
  • Updated test helpers (universal_model, structured_output_model) to use 2.5 models
  • Refreshed model comparison examples in examples/07_model_info.exs

Fixed

  • Documentation consistency: All guides, examples, and docstrings now reference current-generation models
  • Test model references updated across live session, function calling, and system instruction tests

Notes

  • Gemini 2.0 models remain fully supported and available via explicit model selection
  • All gemini-2.0-* model keys retained in manifest for backward compatibility
  • Context caching still supports both 2.0 and 2.5 model versions

0.8.5 - 2025-12-18

Added

  • response_json_schema structured outputs support with structured_json/2 defaulting to JSON Schema
  • Built-in tools serialization for GenerateContent (googleSearch, urlContext, codeExecution)
  • Gemini 3 thinking levels :minimal/:medium and :ultra_high media resolution
  • Veo 3.x video inputs (image, last_frame, reference_images, video extension, resolution) and Gemini API video generation
  • New model registry entries (Gemini 3 Flash preview, 2.0/2.5 variants, flash image, native audio previews, deep research)

Changed

  • person_generation defaults to :allow_none (with :dont_allow alias) for image/video config
  • total_reasoning_tokens renamed to total_thought_tokens in Interactions usage
  • Removed object field from Interaction

Fixed

  • Vertex Interactions get/cancel/delete paths now include project/location
  • Documentation/examples refreshed for Gemini 3 thinking levels, JSON Schema outputs, and built-in tools

0.8.4 - 2025-12-13

Added

  • Comprehensive examples suite: 10 numbered examples covering all major features
    • 01_basic_generation.exs - Simple/configured/creative vs precise generation
    • 02_streaming.exs - Real-time streaming with timing analysis
    • 03_chat_session.exs - Multi-turn conversations with context retention
    • 04_embeddings.exs - Single/batch embeddings, similarity, task types
    • 05_function_calling.exs - Tool registration and automatic execution
    • 06_structured_outputs.exs - JSON schema, entity extraction, classification
    • 07_model_info.exs - Model listing, details, comparison
    • 08_token_counting.exs - Token counting, code vs prose comparison
    • 09_safety_settings.exs - Safety categories and thresholds
    • 10_system_instructions.exs - Persona, formatting, domain expert instructions
  • examples/run_all.sh - Script to run all examples with full output
  • examples/README.md - Comprehensive documentation for examples

Fixed

  • Function calling now handles API responses with string keys ("name", "args") in addition to atom keys
  • run_all.sh script properly iterates through all examples (fixed ((passed++)) bash arithmetic issue)

0.8.3 - 2025-12-13

Added

  • Interactions API parity: Gemini.APIs.Interactions with create/get/cancel/delete, background polling, and SSE streaming with resumption (last_event_id).
  • Full Interactions type system: Gemini.Types.Interactions.* (resources, content, tools, deltas, streaming events).
  • New guide: docs/guides/interactions.md plus new mocked unit tests and optional live tests (@moduletag :live_api).

Changed

  • Streaming transport supports disabling alt=sse injection for endpoints that enable streaming via request fields (Interactions).

Fixed

  • Vertex auth supports x-goog-user-project when a quota project is configured (quota_project_id, VERTEX_QUOTA_PROJECT_ID, or GOOGLE_CLOUD_QUOTA_PROJECT).

0.8.2 - 2025-12-07

Fixed

  • Prevented silent model fallback and double endpoint suffixing by normalizing/validating models and using auth-aware streaming URLs.
  • Stopped double-prefixing Vertex model paths; preserved fully qualified publisher/project models.
  • Removed id from function_response payloads to match API contract and avoid INVALID_ARGUMENT errors.
  • Guarded live files tests to skip gracefully unless Gemini API auth is present (Files API is Gemini-only).

Added

  • Extended image_config to include Vertex-only fields (output_mime_type, output_compression_quality) and updated serialization/tests.
  • Added regression tests for model path building and live smoke test for Gemini 3 image preview (writes to gitignored generated/).

Changed

  • Increased system instruction live test timeout to reduce flakiness under Vertex rate limiting.
  • Improved AFC live test assertions to report errors clearly.

0.8.1 - 2025-12-06

Changed

  • Updated README to reference v0.8.1 and generalize v0.8.x feature callouts.

0.8.0 - 2025-12-06

🎉 Major Feature Release: Complete API Parity with Python SDK

This release brings the Elixir client to near-complete feature parity with the Python google-genai SDK, adding comprehensive support for Tunings (fine-tuning), FileSearchStores, Live/WebSocket API, Application Default Credentials (ADC), and Image/Video Generation APIs.

Added

🎛️ Tunings API - Model Fine-Tuning

🔍 FileSearchStores API - Semantic Search

Live/WebSocket API - Real-time Communication

🔐 Application Default Credentials (ADC) - GCP Native Auth

🖼️ Image Generation API - Imagen Models

  • Gemini.APIs.Images.generate/3: Generate images from text prompts
  • Gemini.APIs.Images.edit/5: Edit images with inpainting and masks
  • Gemini.APIs.Images.upscale/3: Upscale images (2x, 4x factors)
  • ImageGenerationConfig: number_of_images, aspect_ratio, safety_filter_level, person_generation
  • EditImageConfig: edit_mode, mask handling, reference images
  • UpscaleImageConfig: upscale_factor, output settings
  • Safety filtering: Configurable content safety levels
  • Aspect ratios: 1:1, 16:9, 9:16, 4:3, 3:4

🎬 Video Generation API - Veo Models

📖 New Documentation Guides

  • docs/guides/tunings.md - Complete fine-tuning guide
  • docs/guides/file_search_stores.md - Semantic search stores guide
  • docs/guides/live_api.md - Real-time WebSocket guide
  • docs/guides/adc.md - Application Default Credentials guide
  • docs/guides/image_generation.md - Image generation guide
  • docs/guides/video_generation.md - Video generation guide

Technical Implementation

🏛️ Architecture

  • Vertex AI only: Tunings, FileSearchStores, Images, Videos are Vertex AI APIs
  • WebSocket via :gun: HTTP/2 and WebSocket client for Live API
  • ETS-based caching: Thread-safe token caching with automatic refresh
  • Long-running operations: Proper polling with exponential backoff
  • TypedStruct patterns: Consistent type definitions across all new modules

🧪 Testing

  • 200+ new tests for all new modules
  • Live API tests: Tagged with @moduletag :live_api for integration testing
  • Mox-based mocking: Proper HTTP mocking following supertester principles
  • Zero Process.sleep: Proper OTP synchronization in all tests

📈 Quality

  • Zero compilation warnings maintained
  • Complete @spec annotations for all public functions
  • Comprehensive @moduledoc and @doc documentation
  • Follows CODE_QUALITY.md standards

Dependencies

  • Added :gun ~> 2.1 for WebSocket support (Live API)

Migration Notes

For Existing Users

All changes are additive - existing code continues to work unchanged. New APIs are available immediately:

# Fine-tune a model
{:ok, job} = Gemini.APIs.Tunings.tune(%{
  base_model: "gemini-2.5-flash-001",
  tuned_model_display_name: "my-tuned-model",
  training_dataset_uri: "gs://bucket/training.jsonl"
}, auth: :vertex_ai)

# Create semantic search store
{:ok, store} = Gemini.APIs.FileSearchStores.create(%{
  display_name: "Knowledge Base"
}, auth: :vertex_ai)

# Start real-time session
{:ok, session} = Gemini.Live.Session.start_link(
  model: "gemini-2.0-flash-exp",
  auth: :vertex_ai
)

# Generate images
{:ok, images} = Gemini.APIs.Images.generate(
  "A sunset over mountains",
  %ImageGenerationConfig{aspect_ratio: "16:9"},
  auth: :vertex_ai
)

# Generate videos
{:ok, op} = Gemini.APIs.Videos.generate(
  "A cat playing piano",
  %VideoGenerationConfig{duration_seconds: 5},
  auth: :vertex_ai
)

ADC Auto-Discovery

With ADC support, credentials are automatically discovered:

# On GCE/Cloud Run - no configuration needed!
{:ok, response} = Gemini.generate("Hello", auth: :vertex_ai)

# Or with service account
export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"
{:ok, response} = Gemini.generate("Hello", auth: :vertex_ai)

Gap Analysis Update

This release addresses the critical gaps identified in the v0.7.3 gap analysis:

  • ✅ Tunings Module - 100% missing → Now implemented
  • ✅ FileSearchStores - 100% missing → Now implemented
  • ✅ Live/WebSocket API - 100% missing → Now implemented
  • ✅ ADC Support - Critical for GCP → Now implemented
  • ✅ Image Generation - Imagen models → Now implemented
  • ✅ Video Generation - Veo models → Now implemented

Estimated parity with Python SDK: ~95% (up from ~85% in v0.7.3)

0.7.3 - 2025-12-06

Added

System Instruction Support

  • system_instruction option: Set persistent system prompts that guide model behavior across conversations
  • Supports multiple formats: string, Content struct, or map with parts
  • Reduces token usage compared to inline instructions in conversation history
  • Works with all content generation operations

Enhanced Function Calling Framework

  • Gemini.Types.Schema: Complete JSON Schema type for defining function parameters

    • All standard types: string, integer, number, boolean, array, object
    • Support for enum, format, minimum/maximum, pattern constraints
    • Nested object and array schemas
    • API format conversion with to_api_map/1 and from_api_map/1

  • Gemini.Tools.Executor: Execute function calls from Gemini responses

    • Function registry pattern for managing implementations
    • Sequential execution with execute_all/2
    • Parallel execution with execute_all_parallel/3 for I/O-bound operations
    • Automatic response building with build_responses/2
    • Comprehensive error handling

  • Gemini.Tools.AutomaticFunctionCalling: Complete AFC loop implementation

    • Configurable with max_calls, ignore_call_history, parallel_execution
    • Extract function calls from responses with extract_function_calls/1
    • Check for function calls with has_function_calls?/1
    • Full AFC loop with loop/8 for autonomous multi-step execution
    • Call history tracking

  • Coordinator helpers: extract_function_calls/1 and has_function_calls?/1 convenience functions

Documentation

  • New guide: docs/guides/function_calling.md - Complete function calling guide
  • New guide: docs/guides/system_instructions.md - System instruction usage guide
  • Added guides to Hex documentation

Comprehensive Gap Analysis

  • Python SDK Comparison: Complete analysis of Python genai SDK (v1.53.0) vs Elixir gemini_ex implementation
  • Executive Summary: High-level overview with severity classifications and recommendations
  • Feature Parity Matrix: Detailed feature-by-feature comparison showing 55% current coverage
  • Critical Gaps Document: In-depth analysis of 8 critical/high-priority gaps:
    • Live/Real-time API (WebSocket) - Not implemented
    • Tools/Function Calling - Types only, no execution
    • Automatic Function Calling (AFC) - Not implemented
    • System Instruction - Missing from request building
    • Model Tuning API - Not implemented
    • Grounding/Retrieval - Not implemented
    • Code Execution Tool - Not implemented
    • Image/Video Generation - Not implemented
  • Implementation Priorities: Tiered roadmap with code examples for closing gaps
  • Implementation Prompt: Detailed TDD-based prompt for implementing identified gaps

Documentation

  • New gap analysis documents in docs/20251206/gap_analysis/:

    • README.md - Navigation index and methodology
    • 00_executive_summary.md - High-level overview
    • 01_critical_gaps.md - Detailed critical gaps
    • 02_feature_parity_matrix.md - Complete feature comparison
    • 03_implementation_priorities.md - Implementation roadmap
    • IMPLEMENTATION_PROMPT.md - TDD implementation guide

  • Added gap analysis documents to Hex documentation:

    • New "Gap Analysis" group in documentation navigation
    • All gap analysis docs included in package

Technical

  • Analysis conducted using 21 parallel subagent deep-dive reports
  • Covers all major Python SDK components:
    • Client structure, Models API, Chat sessions
    • Authentication, Streaming, Files API
    • Context caching, Batch processing
    • Type definitions, Tools/Function calling
    • Safety settings, Embeddings, Live API
    • Multimodal, Grounding, Async patterns
    • Model tuning, Permissions, Pagination
    • Error handling, Request/Response transformation

Quantitative Findings

MetricPython SDKElixir PortCoverage
Total Lines (types)18,205~3,000~16%
API Modules12758%
Type Definitions200+~50~25%
Overall Parity--55%
  1. System Instruction (2-4 hours) - Quick win, high impact
  2. Function Calling Types (1 week) - Foundation for AI agents
  3. Function Execution (1 week) - Enable tool integration
  4. Automatic FC Loop (1 week) - Complete the agent loop
  5. Live API (3 weeks) - WebSocket for real-time apps

0.7.2 - 2025-12-06

Fixed

  • Rate limiter race condition: Replaced :global.trans/2 with ETS-based spinlock using :ets.insert_new/2 for proper single-node mutex semantics in both ConcurrencyGate and State modules
  • TOCTOU race in lock cleanup: Use :ets.delete_object/2 instead of :ets.delete/2 to atomically delete only if PID still matches, preventing lock theft
  • ETS table options: Changed lock tables to use write_concurrency: true instead of read_concurrency: true for write-heavy workloads
  • Test synchronization: Removed flaky Process.sleep from atomic reservation test; now awaits non-blocking task2 before releasing task1 for deterministic synchronization

Changed

  • Lock acquisition retry sleep increased from 1ms to 5ms to reduce CPU usage under contention

0.7.1 - 2025-12-05

Added

  • Atomic token budget reservation (try_reserve_budget/3) with safety multiplier, reconciliation, and telemetry events (budget_reserved, budget_rejected)
  • Shared retry window gating with jittered release plus telemetry hooks (retry_window_set/hit/release)
  • Model use-case aliases (cache_context, report_section, fast_path) resolved through Gemini.Config.model_for_use_case/2 with documented token minima
  • Streaming now goes through the rate limiter (UnifiedManager): permits are held for the duration of the stream, budget is reserved up front, and telemetry is emitted for stream start/completion/error/stop

Fixed

  • Concurrency gate TOCTOU race hardened with serialized permit acquisition; default non_blocking remains false for server workloads
  • Rate limiter now pre-flight rejects over-budget bursts before dispatching requests and returns surplus budget after responses

0.7.0 - 2025-12-05

🎉 Major Feature Release: Complete API Parity

This release brings the Elixir client to near-complete feature parity with the Python google-genai SDK, adding comprehensive support for Files, Batches, Operations, and Documents APIs.

Added

📁 Files API - Complete File Management

📦 Batches API - Bulk Processing with 50% Cost Savings

⏱️ Operations API - Long-Running Task Management

📄 Documents API - RAG Store Document Management

🏷️ Enhanced Enum Types - Comprehensive Type Safety

New enum modules in Gemini.Types.Enums with to_api/1 and from_api/1 converters:

  • HarmCategory - 12 harm category values
  • HarmBlockThreshold - 6 threshold levels
  • HarmProbability - 5 probability levels
  • BlockedReason - 7 block reasons
  • FinishReason - 12 finish reasons
  • TaskType - 9 embedding task types
  • FunctionCallingMode - 3 function calling modes
  • DynamicRetrievalMode - 3 retrieval modes
  • ThinkingLevel - 3 thinking budget levels
  • CodeExecutionOutcome - 4 execution outcomes
  • ExecutableCodeLanguage - 2 code languages
  • GroundingAttributionConfidence - 4 confidence levels
  • AspectRatio - 4 image aspect ratios
  • ImageSize - 3 image size options
  • VoiceName - 6 voice options for TTS

📖 New Documentation Guides

  • docs/guides/files.md - Complete Files API guide
  • docs/guides/batches.md - Batch processing guide
  • docs/guides/operations.md - Long-running operations guide

Technical Implementation

🏛️ Architecture

  • Resumable upload protocol with 8MB chunks and automatic retry
  • Consistent polling patterns with configurable timeouts and progress callbacks
  • TypedStruct patterns with @derive Jason.Encoder for all new types
  • Full multi-auth support (:gemini and :vertex_ai) across all new APIs

🧪 Testing

  • 94 new tests for Files, Operations, Batches, and Documents APIs
  • Unit tests for all type parsing and helper functions
  • Live API test infrastructure for integration testing
  • Test fixtures for file uploads

📈 Quality

  • Zero compilation warnings
  • Complete @spec annotations for all public functions
  • Comprehensive @moduledoc and @doc documentation
  • Follows CODE_QUALITY.md standards

Changed

  • Updated README.md with new API sections and examples
  • Version bump from 0.6.4 to 0.7.0

Migration Notes

For Existing Users

All changes are additive - existing code continues to work unchanged. New APIs are available immediately:

# Upload and use a file
{:ok, file} = Gemini.APIs.Files.upload("image.png")
{:ok, ready} = Gemini.APIs.Files.wait_for_processing(file.name)
{:ok, response} = Gemini.generate([
  "Describe this image",
  %{file_uri: ready.uri, mime_type: ready.mime_type}
])

# Create a batch job
{:ok, batch} = Gemini.APIs.Batches.create("gemini-2.0-flash",
  file_name: "files/input123",
  display_name: "My Batch"
)
{:ok, completed} = Gemini.APIs.Batches.wait(batch.name)

# Track long-running operations
{:ok, op} = Gemini.APIs.Operations.get("operations/abc123")
{:ok, completed} = Gemini.APIs.Operations.wait_with_backoff(op.name)

0.6.4 - 2025-12-05

Added

Response Type Enhancements

  • UsageMetadata now includes:

    • thoughts_token_count - Token count for thinking models (Gemini 2.0+)
    • tool_use_prompt_token_count - Tokens used in tool/function prompts
    • prompt_tokens_details - Per-modality breakdown of prompt tokens
    • cache_tokens_details - Per-modality breakdown of cached tokens
    • response_tokens_details - Per-modality breakdown of response tokens
    • tool_use_prompt_tokens_details - Per-modality breakdown of tool prompt tokens
    • traffic_type - Billing traffic type (ON_DEMAND, PROVISIONED_THROUGHPUT)

  • GenerateContentResponse now includes:

    • response_id - Unique response identifier for tracking
    • model_version - Actual model version used (e.g., "gemini-2.0-flash-exp-001")
    • create_time - Response creation timestamp

  • Candidate now includes:

    • finish_message - Human-readable message explaining stop reason
    • avg_logprobs - Average log probability score

  • PromptFeedback now includes:

    • block_reason_message - Human-readable block explanation

  • Part now includes:

    • file_data - URI-based file references (alternative to inline_data)
    • function_response - Function call response data
    • thought - Boolean flag for thinking model thought parts

  • SafetyRating now includes:

    • probability_score - Numeric harm probability (0.0-1.0)
    • severity - Harm severity level
    • severity_score - Numeric severity score

Request Type Enhancements

  • GenerationConfig now includes:
    • seed - Deterministic generation seed for reproducible outputs
    • response_modalities - Control output modalities (TEXT, IMAGE, AUDIO)
    • speech_config - Audio output configuration with voice selection
    • media_resolution - Input media resolution control (LOW, MEDIUM, HIGH)

New Types

  • ModalityTokenCount - Per-modality token breakdown
  • TrafficType - Billing traffic type enum
  • Modality - Response modality enum (TEXT, IMAGE, AUDIO)
  • MediaResolution - Input media resolution enum
  • FileData - URI-based file data struct
  • FunctionResponse - Function call response struct
  • SpeechConfig, VoiceConfig, PrebuiltVoiceConfig - Audio output configuration

Changed

  • Response parsing now handles all new fields from Gemini API
  • GenerationConfig encoding includes new fields when present

Fixed

  • Token usage now correctly reports thinking tokens separately from output tokens

0.6.3 - 2025-12-05

Added

  • Concurrency gate is now partitionable via concurrency_key (e.g., per-tenant or per-location) instead of a single global queue per model.
  • Concurrency permit wait is configurable via permit_timeout_ms; default is now :infinity (no queue drop). Per-call overrides supported.
  • Per-request timeout overrides for HTTP and streaming; global default HTTP/stream timeout raised to 120_000ms.
  • Streaming knobs: max_backoff_ms, connect_timeout, and configurable cleanup delay for ManagerV2 (config :gemini_ex, :streaming, cleanup_delay_ms: ...).
  • Configurable context cache TTL defaults via config :gemini_ex, :context_cache, default_ttl_seconds: ....
  • Configurable retry delay fallback via config :gemini_ex, :rate_limiter, default_retry_delay_ms: ....
  • Permit leak protection: holders are monitored and reclaimed if the process dies without releasing.

Changed

  • Default HTTP/stream timeout increased from 30_000ms to 120_000ms.
  • Concurrency gate uses configurable permit_timeout_ms (default :infinity) instead of a fixed 60s timeout.

Fixed

  • Streaming client no longer leaks :persistent_term state; SSE parse errors now surface instead of being silently dropped.
  • Streaming backoff ceiling and connect timeout are tunable; SSE parsing failures return errors.

0.6.2 - 2025-12-05

Fixed

  • Eliminated recursive retry loop on :over_budget blocking calls; blocking now waits once for the current window to end, then retries through the normal pipeline.
  • Over-budget retry_at is now set to the window end in non-blocking mode instead of nil.
  • Requests whose estimated tokens exceed the configured budget return immediately with request_too_large: true instead of hanging.

Added

  • estimated_cached_tokens option for proactive budgeting with cached contexts; cached token usage (cachedContentTokenCount) is now included in recorded input tokens.
  • Telemetry for over-budget waits/errors now includes token estimates and wait metadata.
  • max_budget_wait_ms config/option to cap how long blocking over-budget calls will sleep before returning a rate_limited error with retry_at.

Documentation

  • README and rate limiting guide updated with over-budget behavior, estimated_cached_tokens, and cached context budgeting notes.

0.6.1 - 2025-12-04

⚠️ Potentially breaking (upgrade note): Token estimation now runs automatically and budget checks fall back to profile defaults. Apps that never set :estimated_input_tokens or :token_budget_per_window can now receive local :over_budget errors. To preserve 0.6.0 behavior, set token_budget_per_window: nil (globally or per-call), or disable the rate limiter.

Added

Proactive Rate Limiting Enhancements (ADR Implementation)

  • Auto Token Estimation (ADR-0001)

    • Automatic input token estimation at the Coordinator boundary before request normalization
    • Token estimates passed to rate limiter via :estimated_input_tokens option
    • Safe handling of API maps (%{contents: [...]}) in Tokens.estimate/1 - returns 0 for unknown shapes instead of raising
    • Supports both atom keys (:contents) and string keys ("contents")

  • Token Budget Configuration (ADR-0002)

    • New token_budget_per_window config field with conservative defaults
    • New window_duration_ms config field (default: 60,000ms)
    • Budget checking falls back to config.token_budget_per_window when not in per-request opts
    • State.record_usage/4 now accepts configurable window duration via opts

  • Enhanced 429 Propagation (ADR-0003)

    • Retry state now captures quota_dimensions and quota_value from 429 responses
    • Enhanced quota metric extraction from nested error details

  • Tier-Based Rate Limit Profiles (ADR-0004)

    • New :free_tier profile - Conservative for 15 RPM / 1M TPM (32,000 token budget)
    • New :paid_tier_1 profile - Standard production 500 RPM / 4M TPM (1,000,000 token budget)
    • New :paid_tier_2 profile - High throughput 1000 RPM / 8M TPM (2,000,000 token budget)
    • Updated :dev and :prod profiles with token budgets
    • Profile type expanded to include all tier options

Changed

  • RateLimiter.Config struct now includes token_budget_per_window and window_duration_ms fields
  • Manager.check_token_budget/3 now falls back to config defaults
  • Manager.record_usage_from_response/3 passes window duration from config to State
  • Updated docs/guides/rate_limiting.md with comprehensive tier documentation

Documentation

  • Added Quick Start section with tier profile selection table
  • Expanded Profiles section with all tier configurations
  • Enhanced Token Budgeting section explaining automatic estimation
  • Added Fine-Tuning section for concurrency vs token budget guidance

0.6.0 - 2025-12-04

Added

  • Context Caching Enhancements

    • Cache creation now supports system_instruction parameter for setting system-level instructions that apply to all cached content usage
    • Cache creation now supports tools parameter for caching function declarations alongside content
    • Cache creation now supports tool_config parameter for configuring function calling behavior in cached contexts
    • Cache creation now supports fileUri in content parts for caching files stored in Google Cloud Storage (gs:// URIs)
    • Cache creation now supports kms_key_name parameter for customer- managed encryption keys (Vertex AI only)
    • Resource name normalization for Vertex AI automatically expands short cache names like "cachedContents/abc" to fully qualified paths like "projects/{project}/locations/{location}/cachedContents/abc"
    • Model name normalization for Vertex AI automatically expands model names to full publisher paths
    • Top-level cache API delegates added to main Gemini module:
    • CachedContentUsageMetadata struct expanded with Vertex AI specific fields: audio_duration_seconds, image_count, text_count, and video_duration_seconds
    • Model validation warning when using models that may not support explicit caching (models without version suffixes)
    • Live test covering system_instruction with fileUri caching

  • Auth-Aware Model Configuration System

    • Model registry organized by API compatibility:
      • Universal models work identically in both Gemini API and Vertex AI
      • Gemini API models include convenience aliases like -latest suffix
      • Vertex AI models include EmbeddingGemma variants
    • Config.default_model/0 automatically selects appropriate model based on detected authentication:
      • Gemini API: gemini-flash-lite-latest
      • Vertex AI: gemini-2.0-flash-lite
    • Config.default_embedding_model/0 selects embedding model by auth:
      • Gemini API: gemini-embedding-001 (3072 dimensions)
      • Vertex AI: embeddinggemma (768 dimensions)
    • Config.default_model_for/1 and Config.default_embedding_model_for/1 for explicit API type selection
    • Config.models_for/1 returns all models available for a specific API
    • Config.model_available?/2 checks if a model key works with an API
    • Config.model_api/1 returns the API compatibility of a model key
    • Config.current_api_type/0 returns detected auth type
    • Embedding configuration system with per-model settings:
      • Config.embedding_config/1 returns full config for embedding models
      • Config.uses_prompt_prefix?/1 checks if model uses prompt prefixes
      • Config.embedding_prompt_prefix/2 generates task-specific prefixes
      • Config.default_embedding_dimensions/1 returns model default dims
      • Config.needs_normalization?/2 checks if manual normalization needed
    • EmbeddingGemma support with automatic prompt prefix formatting for task types (retrieval_query becomes "task: search result | query: ")

  • Test Infrastructure

    • Gemini.Test.ModelHelpers module for centralized model references
    • Gemini.Test.AuthHelpers module for shared auth detection logic
    • Helper functions: auth_available?/0, gemini_api_available?/0, vertex_api_available?/0, default_model/0, embedding_model/0, thinking_model/0, caching_model/0, universal_model/0

Changed

  • Auth.build_headers/2 now returns {:ok, headers} or {:error, reason} instead of always returning headers, enabling proper error propagation
  • Gemini.configure/2 now stores config under :gemini app environment to align with Config.auth_config/0 which reads from both :gemini and :gemini_ex namespaces
  • EmbedContentRequest.new/2 automatically formats text with prompt prefixes when using EmbeddingGemma models on Vertex AI
  • All example scripts updated to use Config.default_model() instead of hardcoded model strings
  • All tests updated to use auth-aware model selection via ModelHelpers
  • Config module default model comment updated to explain auto-detection

Fixed

  • Vertex AI Cache Endpoints: Cache operations now build fully qualified paths (projects/{project}/locations/{location}/cachedContents) instead of calling /cachedContents directly, which was causing 404 errors
  • Config Alignment: Gemini.configure/2 now properly feeds config to Config.auth_config/0 by using the correct app environment key
  • Service Account Auth: Removed placeholder tokens that masked real authentication failures; errors now propagate properly with descriptive messages
  • JWT Token Exchange: Fixed OAuth2 JWT payload to include scope in the JWT claims as required by Google's jwt-bearer grant type specification
  • Content Formatting: Part formatting now handles function calls, function responses, thought signatures, file data, and media resolution correctly instead of leaving them in snake_case struct format
  • Empty Env Vars: Environment variable reading now treats empty strings as unset, preventing configuration issues with GEMINI_API_KEY=""
  • ContextCache.create/2: Now accepts string content directly in addition to lists, matching README documentation examples
  • Model Prefix Handling: Model name normalization no longer double- prefixes when callers pass models/... format

Documentation

  • README updated with enhanced context caching examples showing system_instruction, fileUri, and model selection
  • README includes new Model Configuration System section explaining auth-aware defaults and API differences
  • README includes embedding model differences table
  • Config module documentation expanded with model registry explanation
  • Implementation plan documents added in docs/20251204/

0.5.2 - 2025-12-03

Fixed

  • Fixed a regression where 429 responses lost their http_status, causing the rate limiter to misclassify them as permanent errors. API errors now preserve status and RetryInfo details so automatic backoff/RetryInfo delays are honored by default.

0.5.1 - 2025-12-03

Added

Gemini 3 Pro Support

Full support for Google's Gemini 3 model family with new API features:

  • thinking_level Parameter - New thinking control for Gemini 3 models

    • GenerationConfig.thinking_level(:low) - Fast responses, minimal reasoning
    • GenerationConfig.thinking_level(:high) - Deep reasoning (default for Gemini 3)
    • Note: :medium is not currently supported by the API
    • Cannot be used with thinking_budget in the same request (API returns 400)

  • gemini-3-pro-image-preview Model - Image generation support

    • Generate images from text prompts
    • Configurable aspect ratios: "16:9", "1:1", "4:3", "3:4", "9:16"
    • Output resolutions: "2K" or "4K"
    • GenerationConfig.image_config(aspect_ratio: "16:9", image_size: "4K")

  • media_resolution Parameter - Fine-grained vision processing control

    • :low - 280 tokens for images, 70 for video frames
    • :medium - 560 tokens for images, 70 for video frames
    • :high - 1120 tokens for images, 280 for video frames
    • Part.inline_data_with_resolution(data, mime_type, :high)
    • Part.with_resolution(existing_part, :high)

  • thought_signature Field - Reasoning context preservation

    • Maintains reasoning context across API calls
    • Required for multi-turn function calling in Gemini 3
    • Part.with_thought_signature(part, signature)
    • SDK handles automatically in chat sessions
    • NEW: Automatic extraction via Gemini.extract_thought_signatures/1
    • NEW: Automatic echoing in Chat.add_model_response/2

  • Context Caching API - Cache long context for improved performance

  • New Example: examples/gemini_3_demo.exs - Comprehensive Gemini 3 features demonstration

Updated Validation

  • Gemini.Validation.ThinkingConfig now validates Gemini 3's thinking_level
  • Prevents combining thinking_level and thinking_budget (API constraint)
  • Warns that :medium thinking level is not supported

Changed

Embeddings Documentation Updates

  • Fixed EMBEDDINGS.md: Corrected code examples and removed outdated/confusing information

    • Fixed incorrect module reference (Coordinator.EmbedContentResponse → EmbedContentResponse)
    • Removed confusing legacy model section (there's only gemini-embedding-001 now)
    • Updated model comparison to reflect current API (single model with MRL support)
    • Updated async batch section with working code examples (was marked as "planned")
    • Added deprecation notice for embedding-001, embedding-gecko-001, gemini-embedding-exp-03-07 (October 2025)

  • Updated embed_content_request.ex: Removed deprecated model reference from documentation

Fixed

  • Documentation now accurately reflects the current Gemini Embeddings API specification (June 2025)
  • Clarified that gemini-embedding-001 is the only recommended model with full MRL support

Migration Notes

For Gemini 3 Users

# Use thinking_level instead of thinking_budget for Gemini 3
config = GenerationConfig.thinking_level(:low)  # Fast
config = GenerationConfig.thinking_level(:high) # Deep reasoning (default)

# Image generation
config = GenerationConfig.image_config(aspect_ratio: "16:9", image_size: "4K")
{:ok, response} = Coordinator.generate_content(
  "Generate an image of a sunset",
  model: "gemini-3-pro-image-preview",
  generation_config: config
)

# Media resolution for vision tasks
Part.inline_data_with_resolution(image_data, "image/jpeg", :high)

Temperature Recommendation

For Gemini 3, keep temperature at 1.0 (the default). Lower temperatures may cause looping or degraded performance on complex reasoning tasks.

0.5.0 - 2025-12-03

Added

Rate Limiting System (Default ON)

A comprehensive rate limiting, retry, and concurrency management system that is enabled by default:

  • RateLimitManager - Central coordinator that wraps all outbound requests

    • ETS-based state tracking keyed by {model, location, metric}
    • Tracks retry_until timestamps from 429 RetryInfo responses
    • Token usage sliding windows for budget estimation
    • Configurable via application config or per-request options

  • ConcurrencyGate - Per-model concurrency limiting

    • Default limit of 4 concurrent requests per model
    • Configurable with max_concurrency_per_model (nil/0 disables)
    • Optional adaptive mode: adjusts concurrency based on 429 responses
    • Non-blocking mode returns immediately if no permits available

  • RetryManager - Intelligent retry with backoff

    • Honors 429 RetryInfo.retryDelay from API responses
    • Exponential backoff with jitter for 5xx/transient errors
    • Configurable max attempts (default: 3)
    • Coordinates with rate limiter to avoid double retries

  • TokenBudget - Preflight token estimation

    • Track actual usage from responses
    • Block/queue when over configured budget
    • Sliding window tracking per model/location

Telemetry Events

New telemetry events for rate limit monitoring (consistent with existing [:gemini, ...] namespace):

  • [:gemini, :rate_limit, :request, :start] - Request submitted
  • [:gemini, :rate_limit, :request, :stop] - Request completed
  • [:gemini, :rate_limit, :wait] - Waiting for retry window
  • [:gemini, :rate_limit, :error] - Rate limit error

Structured Errors

New structured error types:

  • {:error, {:rate_limited, retry_at, details}} - Rate limited with retry info
  • {:error, {:transient_failure, attempts, original_error}} - Transient failure after retries

Configuration Options

config :gemini_ex, :rate_limiter,
  max_concurrency_per_model: 4,    # nil/0 disables
  max_attempts: 3,
  base_backoff_ms: 1000,
  jitter_factor: 0.25,
  non_blocking: false,
  disable_rate_limiter: false,
  adaptive_concurrency: false,
  adaptive_ceiling: 8,
  profile: :prod  # :dev | :prod | :custom

Per-Request Options

  • disable_rate_limiter: true - Bypass all rate limiting
  • non_blocking: true - Return immediately if rate limited
  • max_concurrency_per_model: N - Override concurrency
  • estimated_input_tokens: N - For budget checking
  • token_budget_per_window: N - Max tokens per window

Documentation

  • New rate limiting guide: docs/guides/rate_limiting.md
  • Comprehensive module documentation for all rate limiter components
  • Updated README with rate limiting section

Changed

  • HTTP client now routes all requests through rate limiter by default
  • Supervisor now starts RateLimitManager on application boot

Technical Notes

  • Streaming Safe: Rate limiter only gates request submission; open streams are not interrupted
  • Coordinate Retry Layers: Retry logic coordinates between rate limiter and HTTP client to avoid double retries
  • Test Infrastructure: Added Bypass-based fake Gemini endpoint for testing rate limit behavior

Migration Guide

Rate limiting is enabled by default. To disable:

# Per-request
Gemini.generate("Hello", disable_rate_limiter: true)

# Globally (not recommended)
config :gemini_ex, :rate_limiter, disable_rate_limiter: true

The new structured errors are backward compatible - existing error handling will continue to work, but you can now pattern match on rate limit specifics:

case Gemini.generate("Hello") do
  {:ok, response} -> handle_success(response)
  {:error, {:rate_limited, retry_at, _}} -> schedule_retry(retry_at)
  {:error, other} -> handle_error(other)
end

0.4.0 - 2025-11-06

Added

  • Structured Outputs Enhancement - Full support for Gemini API November 2025 updates
    • property_ordering field in GenerationConfig for Gemini 2.0 model support
    • structured_json/2 convenience helper for structured output setup
    • property_ordering/2 helper for explicit property ordering
    • temperature/2 helper for setting temperature values
    • Support for new JSON Schema keywords:
      • anyOf - Union types and conditional structures
      • $ref - Recursive schema definitions
      • minimum/maximum - Numeric value constraints
      • additionalProperties - Control over extra properties
      • type: "null" - Nullable field definitions
      • prefixItems - Tuple-like array structures
    • Comprehensive integration tests for structured outputs
    • Working examples demonstrating all new features

Improved

  • Enhanced documentation for structured outputs use cases
  • Better code examples in README and API reference
  • Expanded test coverage for generation config options

Notes

  • Gemini 2.5+ models preserve schema key order automatically
  • Gemini 2.0 models require explicit property_ordering field
  • All changes are backward compatible - no breaking changes

0.3.1 - 2025-10-15

🎉 Major Feature: Async Batch Embedding API (Phase 4)

This release adds production-scale async batch embedding support with 50% cost savings compared to the interactive API. Process thousands to millions of embeddings asynchronously with Long-Running Operation (LRO) support, state tracking, and priority management.

Added

🚀 Async Batch Embedding API

  • async_batch_embed_contents/2: Submit large batches asynchronously for background processing

    • 50% cost savings vs interactive embedding API
    • Suitable for RAG system indexing, knowledge base building, and large-scale retrieval
    • Returns immediately with batch ID for polling
    • Support for inline requests with metadata tracking

  • get_batch_status/1: Poll batch job status with progress tracking

    • Real-time progress metrics via EmbedContentBatchStats
    • State transitions: PENDING → PROCESSING → COMPLETED/FAILED
    • Track successful, failed, and pending request counts

  • get_batch_embeddings/1: Retrieve results from completed batch jobs

    • Extract embeddings from inline responses
    • Support for file-based output detection
    • Automatic filtering of successful responses

  • await_batch_completion/2: Convenience polling with configurable intervals

    • Automatic polling until completion or timeout
    • Progress callback support for monitoring
    • Configurable poll interval and timeout

📊 Complete Type System

  • BatchState: Job state enum (:unspecified, :pending, :processing, :completed, :failed, :cancelled)

  • EmbedContentBatchStats: Request tracking with progress metrics

    • progress_percentage/1: Calculate completion percentage
    • success_rate/1 and failure_rate/1: Quality metrics
    • is_complete?/1: Completion check

  • Request Types:

    • InlinedEmbedContentRequest: Single request with metadata
    • InlinedEmbedContentRequests: Container for multiple requests
    • InputEmbedContentConfig: Union type for file vs inline input
    • EmbedContentBatch: Complete batch job request with priority

  • Response Types:

    • InlinedEmbedContentResponse: Single response with success/error
    • InlinedEmbedContentResponses: Container with helper functions
    • EmbedContentBatchOutput: Union type for file vs inline output
    • EmbedContentBatch: Complete batch status with lifecycle tracking

🧪 Comprehensive Test Coverage

  • 41 new unit tests for batch types (BatchState, BatchStats)
  • Full TDD approach with test-first implementation
  • 425 total tests passing (up from 384 in v0.3.0)
  • Zero compilation warnings maintained

Technical Implementation

🎯 Production Features

  • Long-Running Operations (LRO): Full async job lifecycle support
  • Priority-based Processing: Control batch execution order with priority field
  • Progress Tracking: Real-time stats on successful, failed, and pending requests
  • Multi-auth Support: Works with both Gemini API and Vertex AI
  • Type Safety: Complete @spec annotations for all new functions
  • Error Handling: Comprehensive error messages and recovery paths

📈 Performance & Cost

  • 50% cost savings: Async batch API offers half the cost of interactive embedding
  • Scalability: Process millions of embeddings efficiently
  • Production-ready: Designed for large-scale RAG systems and knowledge bases
  • Flexible polling: Configurable intervals (default 5s) with timeout (default 10min)

Usage Examples

# Submit async batch for background processing
{:ok, batch} = Gemini.async_batch_embed_contents(
  ["Text 1", "Text 2", "Text 3"],
  display_name: "My Knowledge Base",
  task_type: :retrieval_document,
  output_dimensionality: 768
)

# Poll for status
{:ok, updated_batch} = Gemini.get_batch_status(batch.name)

# Check progress
if updated_batch.batch_stats do
  progress = updated_batch.batch_stats |> EmbedContentBatchStats.progress_percentage()
  IO.puts("Progress: #{Float.round(progress, 1)}%")
end

# Wait for completion (convenience function)
{:ok, completed_batch} = Gemini.await_batch_completion(
  batch.name,
  poll_interval: 10_000,  # 10 seconds
  timeout: 1_800_000,     # 30 minutes
  on_progress: fn b ->
    progress = EmbedContentBatchStats.progress_percentage(b.batch_stats)
    IO.puts("Progress: #{Float.round(progress, 1)}%")
  end
)

# Retrieve embeddings
{:ok, embeddings} = Gemini.get_batch_embeddings(completed_batch)
IO.puts("Retrieved #{length(embeddings)} embeddings")

Changed

  • Enhanced Coordinator module: Added async batch embedding functions alongside existing sync APIs
  • Type system expansion: New types in Gemini.Types.Request and Gemini.Types.Response namespaces

Migration Notes

For v0.3.0 Users

  • All existing synchronous embedding APIs remain unchanged and fully compatible

  • New async batch API is additive - no breaking changes

  • Use async batch API for:

    • Large-scale embedding generation (1000s-millions of texts)
    • Background processing with 50% cost savings
    • RAG system indexing and knowledge base building
    • Non-time-critical embedding workflows

  • Continue using sync API (embed_content/2, batch_embed_contents/2) for:

    • Real-time embedding needs
    • Small batches (<100 texts)
    • Interactive workflows requiring immediate results

Future Enhancements

  • File-based batch input/output support (GCS integration)
  • Batch cancellation and deletion APIs
  • Enhanced progress monitoring with estimated completion times
  • API Specification: oldDocs/docs/spec/GEMINI-API-07-EMBEDDINGS_20251014.md (lines 129-442)
  • Implementation Plan: EMBEDDING_IMPLEMENTATION_PLAN.md (Phase 4 section)

0.3.0 - 2025-10-14

🎉 Major Feature: Complete Embedding Support with MRL

This release adds comprehensive text embedding functionality with Matryoshka Representation Learning (MRL), enabling powerful semantic search, RAG systems, classification, and more.

Added

📊 Embedding API with Normalization & Distance Metrics

  • ContentEmbedding.normalize/1: L2 normalization to unit length (required for non-3072 dimensions per API spec)
  • ContentEmbedding.norm/1: Calculate L2 norm of embedding vectors
  • ContentEmbedding.euclidean_distance/2: Euclidean distance metric for similarity
  • ContentEmbedding.dot_product/2: Dot product similarity (equals cosine for normalized embeddings)
  • Enhanced cosine_similarity/2: Improved documentation with normalization requirements

🔬 Production-Ready Use Case Examples

  • examples/use_cases/mrl_normalization_demo.exs: Comprehensive MRL demonstration

    • Quality vs storage tradeoffs across dimensions (128-3072)
    • MTEB benchmark comparison table
    • Normalization requirements and effects
    • Distance metrics comparison (cosine, euclidean, dot product)
    • Best practices for dimension selection

  • examples/use_cases/rag_demo.exs: Complete RAG pipeline implementation

    • Build and index knowledge base with RETRIEVAL_DOCUMENT task type
    • Embed queries with RETRIEVAL_QUERY task type
    • Retrieve top-K relevant documents using semantic similarity
    • Generate contextually-aware responses
    • Side-by-side comparison with non-RAG baseline

  • examples/use_cases/search_reranking.exs: Semantic reranking for search

    • E-commerce product search example
    • Compare keyword vs semantic ranking
    • Hybrid ranking strategy (keyword + semantic weighted)
    • Handle synonyms and conceptual relevance

  • examples/use_cases/classification.exs: K-NN classification

    • Few-shot learning with minimal training examples
    • Customer support ticket categorization
    • Confidence scoring and accuracy evaluation
    • Dynamic category addition without retraining

📚 Enhanced Documentation

  • Complete MRL documentation in examples/EMBEDDINGS.md:

    • Matryoshka Representation Learning explanation
    • MTEB benchmark scores table (128d to 3072d)
    • Normalization requirements and best practices
    • Model comparison table (gemini-embedding-001 vs gemini-embedding-001)
    • Critical normalization warnings
    • Distance metrics usage guide

  • README.md embeddings section:

    • Quick start guide for embeddings
    • MRL concepts and dimension selection
    • Task types for better quality
    • Batch embedding examples
    • Links to advanced use case examples

🧪 Comprehensive Test Coverage

  • 26 unit tests for ContentEmbedding module:

    • Normalization accuracy (L2 norm = 1.0)
    • Distance metrics validation
    • Edge cases and error handling
    • Zero vector handling

  • 20 integration tests for embedding coordinator:

    • Single and batch embedding workflows
    • Task type variations
    • Output dimensionality control
    • Error scenarios

Technical Implementation

🎯 Key Features

  • MRL Support: Flexible dimensions (128-3072) with minimal quality loss

    • 768d: 67.99 MTEB (25% storage, -0.26% loss) - RECOMMENDED
    • 1536d: 68.17 MTEB (50% storage, same as 3072d!)
    • 3072d: 68.17 MTEB (100% storage, pre-normalized)

  • Critical Normalization: Only 3072-dimensional embeddings are pre-normalized by API

    • All other dimensions MUST be normalized before computing similarity
    • Cosine similarity focuses on direction (semantic meaning), not magnitude
    • Non-normalized embeddings have varying magnitudes that distort calculations

  • Production Quality: 384 tests passing (100% success rate)

  • Type Safety: Complete @spec annotations for all new functions

  • Code Quality: Zero compilation warnings maintained

📈 Performance Characteristics

  • Storage Efficiency: 768d offers 75% storage savings with <0.3% quality loss
  • Quality Benchmarks: MTEB scores prove minimal degradation across dimensions
  • Real-time Processing: Efficient normalization and distance calculations

Changed

  • Updated README.md: Added embeddings section in features list and comprehensive usage guide
  • Enhanced EMBEDDINGS.md: Complete rewrite with MRL documentation and advanced examples
  • Model Recommendations: Updated to highlight gemini-embedding-001 with MRL support

Migration Notes

For New Users

# Generate embedding with recommended 768 dimensions
{:ok, response} = Gemini.embed_content(
  "Your text",
  model: "gemini-embedding-001",
  output_dimensionality: 768
)

# IMPORTANT: Normalize before computing similarity!
alias Gemini.Types.Response.ContentEmbedding

normalized = ContentEmbedding.normalize(response.embedding)
similarity = ContentEmbedding.cosine_similarity(normalized, other_normalized)

Dimension Selection Guide

  • 768d: Best for most applications (storage/quality balance)
  • 1536d: High quality at 50% storage (same MTEB as 3072d)
  • 3072d: Maximum quality, pre-normalized (largest storage)
  • 512d or lower: Extreme efficiency (>1% quality loss)

Future Roadmap

v0.4.0 (Planned): Async Batch Embedding API

  • Long-running operations (LRO) support
  • 50% cost savings vs interactive embedding
  • Batch state tracking and priority support
  • Comprehensive Guide: examples/EMBEDDINGS.md
  • MRL Demo: examples/use_cases/mrl_normalization_demo.exs
  • RAG Example: examples/use_cases/rag_demo.exs
  • API Specification: oldDocs/docs/spec/GEMINI-API-07-EMBEDDINGS_20251014.md

0.2.3 - 2025-10-08

Fixed

  • CRITICAL: Double-encoding bug in multimodal content - Fixed confusing base64 encoding behavior (Issue #11 comment from @jaimeiniesta)
    • Problem: When users passed Base.encode64(image_data) with type: "base64", data was encoded AGAIN internally, causing double-encoding
    • Symptom: Users had to pass raw (non-encoded) data despite specifying type: "base64", which was confusing and counterintuitive
    • Root cause: Blob.new/2 always called Base.encode64(), even when data was already base64-encoded
    • Fix: When source: %{type: "base64", data: ...} is specified, data is now treated as already base64-encoded
    • Impact:
      • ✅ Users can now pass Base.encode64(data) as expected (documentation examples now work correctly)
      • ✅ API behavior matches user expectations: type: "base64" means data IS base64-encoded
      • ✅ Applies to both Anthropic-style format (%{type: "image", source: %{type: "base64", ...}}) and Gemini SDK style (%{inline_data: %{data: ..., mime_type: ...}})
      • ⚠️ Breaking change for workarounds: If you were passing raw (non-encoded) data as a workaround, you must now pass properly base64-encoded data
    • Special thanks to @jaimeiniesta for reporting this confusing behavior!

Changed

  • Enhanced normalize_single_content/1 to preserve base64 data without re-encoding when type: "base64"
  • Enhanced normalize_part/1 to preserve base64 data in inline_data maps
  • Updated tests to verify correct base64 handling
  • Added demonstration script: examples/fixed_double_encoding_demo.exs

0.2.2 - 2025-10-07

Added

  • Flexible multimodal content input - Accept multiple intuitive input formats for images and text (Closes #11)

    • Support Anthropic-style format: %{type: "text", text: "..."} and %{type: "image", source: %{type: "base64", data: "..."}}
    • Support map format with explicit role and parts: %{role: "user", parts: [...]}
    • Support simple string inputs: "What is this?"
    • Support mixed formats in single request
    • Automatic MIME type detection from image magic bytes (PNG, JPEG, GIF, WebP)
    • Graceful fallback to explicit MIME type or JPEG default

  • Thinking budget configuration - Control thinking token usage for cost optimization (Closes #9, Supersedes #10)

    • GenerationConfig.thinking_budget/2 - Set thinking token budget (0 to disable, -1 for dynamic, or fixed amount)
    • GenerationConfig.include_thoughts/2 - Enable thought summaries in responses
    • GenerationConfig.thinking_config/3 - Set both budget and thoughts in one call
    • Gemini.Validation.ThinkingConfig module - Model-aware budget validation
    • Support for all Gemini 2.5 series models (Pro, Flash, Flash Lite)

Fixed

  • Multimodal content handling - Users can now pass images and text in natural, intuitive formats

    • Previously: Only accepted specific Content structs, causing FunctionClauseError
    • Now: Accepts flexible formats and automatically normalizes them
    • Backward compatible: All existing code continues to work

  • CRITICAL: Thinking budget field names - Fixed PR #10's critical bug that prevented thinking budget from working

    • Previously: Sent thinking_budget (snake_case) which API silently ignored, users still charged
    • Now: Sends thinkingBudget (camelCase) as required by official API, actually disables thinking
    • Added includeThoughts support that was missing from PR #10
    • Added model-specific budget validation (Pro: 128-32K, Flash: 0-24K, Lite: 0 or 512-24K)
    • Note: This supersedes PR #10 with a correct, fully-tested implementation

Changed

  • Enhanced Coordinator.generate_content/2 to accept flexible content formats
  • Added automatic content normalization layer
  • Added convert_thinking_config_to_api/1 to properly convert field names to camelCase
  • GenerationConfig.ThinkingConfig is now a typed struct (not plain map)

[Unreleased]

0.2.1 - 2025-08-08

Added

  • ALTAR Integration Documentation: Added detailed documentation for the ALTAR protocol integration, explaining the architecture and benefits of the new type-safe, production-grade tool-calling foundation.
  • ALTAR Version Update: Bumped ALTAR dependency to v0.1.2.

0.2.0 - 2025-08-07

🎉 Major Feature: Automatic Tool Calling

This release introduces a complete, production-grade tool-calling (function calling) feature set, providing a seamless, Python-SDK-like experience for building powerful AI agents. The implementation is architected on top of the robust, type-safe ALTAR protocol for maximum reliability and future scalability.

Added

🤖 Automatic Tool Execution Engine

  • New Public API: Gemini.generate_content_with_auto_tools/2 orchestrates the entire multi-turn tool-calling loop. The library now automatically detects when a model wants to call a tool, executes it, sends the result back, and returns the final, synthesized text response.
  • Recursive Orchestrator: A resilient, private orchestrator manages the conversation, preventing infinite loops with a configurable :turn_limit.
  • Streaming Support: Gemini.stream_generate_with_auto_tools/2 provides a fully automated tool-calling experience for streaming. A new ToolOrchestrator GenServer manages the complex, multi-stage stream, ensuring the end-user only receives the final text chunks.

🔧 Manual Tool Calling Foundation (For Advanced Users)

  • New Gemini.Tools Facade: Provides a clean, high-level API (register/2, execute_calls/1) for developers who need full control over the tool-calling loop.
  • Parallel Execution: Gemini.Tools.execute_calls/1 uses Task.async_stream to execute multiple tool calls from the model in parallel, improving performance.
  • Robust Error Handling: Individual tool failures are captured as a valid ToolResult and do not crash the calling process.

🏛️ Architectural Foundation (ALTAR Integration)

  • ALTAR Dependency: The project now builds upon the altar library, using its robust Data Model (ADM) and Local Execution Runtime (LATER).
  • Supervised Registry: gemini_ex now starts and supervises its own named Altar.LATER.Registry process (Gemini.Tools.Registry), providing a stable, application-wide endpoint for tool management.
  • Formalized Gemini.Chat Module: The chat history management has been completely refactored into a new Gemini.Chat struct and module, providing immutable, type-safe handling of complex multi-turn histories that include function_call and function_response turns.

Changed

  • Part Struct: The Gemini.Types.Part struct was updated to include a function_call field, enabling type-safe parsing of model responses.
  • Response Parsing: The core response parser in Gemini.Generate has been significantly enhanced to safely deserialize functionCall parts from the API, validating them against the Altar.ADM contract.
  • Chat History: The Gemini.send_message/2 function has been refactored to use the new, more powerful Gemini.Chat module.

Fixed

  • CRITICAL: Tool Response Role: The role for functionResponse turns sent to the API is now correctly set to "tool" (was "user"), ensuring API compatibility.
  • Architectural Consistency: Removed an erroneous function_response field from the Part struct. functionResponse parts are now correctly handled as raw maps, consistent with the library's design.
  • Test Consistency: Updated all relevant tests to use camelCase string keys when asserting against API-formatted data structures, improving test accuracy.

📚 Documentation & Examples

  • New Example (auto_tool_calling_demo.exs): A comprehensive script demonstrating how to register multiple tools and use the new automatic execution APIs for both standard and streaming requests.
  • New Example (manual_tool_calling_demo.exs): A clear demonstration of the advanced, step-by-step manual tool-calling loop.

0.1.1 - 2025-08-03

🐛 Fixed

Generation Config Bug Fix

  • Critical Fix: Fixed GenerationConfig options being dropped in Gemini.APIs.Coordinator module
    • Previously, only 4 basic options (temperature, max_output_tokens, top_p, top_k) were supported
    • Now supports all 12 GenerationConfig fields including response_schema, response_mime_type, stop_sequences, etc.
    • Fixed inconsistency between Gemini.Generate and Gemini.APIs.Coordinator modules
    • Both modules now handle generation config options identically

Enhanced Generation Config Support

  • Complete Field Coverage: Added support for all missing GenerationConfig fields:
    • response_schema - For structured JSON output
    • response_mime_type - For controlling output format
    • stop_sequences - For custom stop sequences
    • candidate_count - For multiple response candidates
    • presence_penalty - For controlling topic repetition
    • frequency_penalty - For controlling word repetition
    • response_logprobs - For response probability logging
    • logprobs - For token probability information

Improved Request Building

  • Struct Priority: GenerationConfig structs now take precedence over individual keyword options
  • Key Conversion: Proper snake_case to camelCase conversion for all API fields
  • Nil Filtering: Automatic filtering of nil values to reduce request payload size
  • Backward Compatibility: Existing code using individual options continues to work unchanged

🧪 Testing

Comprehensive Test Coverage

  • 70 New Tests: Added extensive test suite covering all generation config scenarios
  • Bug Reproduction: Tests that demonstrate the original bug and verify the fix
  • Field Coverage: Individual tests for each of the 12 generation config fields
  • Integration Testing: End-to-end tests with real API request structure validation
  • Regression Prevention: Tests ensure the bug cannot reoccur in future versions

Test Categories Added

  • Individual option handling tests
  • GenerationConfig struct handling tests
  • Mixed option scenarios (struct + individual options)
  • Edge case handling (nil values, invalid types)
  • API request structure validation
  • Backward compatibility verification

🔧 Technical Improvements

Code Quality

  • Helper Functions: Added convert_to_camel_case/1 and struct_to_api_map/1 utilities
  • Error Handling: Improved validation and error messages for generation config
  • Documentation: Enhanced inline documentation for generation config handling
  • Type Safety: Maintained strict type checking while expanding functionality

Performance

  • Request Optimization: Reduced API request payload size by filtering nil values
  • Processing Efficiency: Streamlined generation config building process
  • Memory Usage: More efficient handling of large GenerationConfig structs

📚 Documentation

Updated Examples

  • Enhanced examples to demonstrate new generation config capabilities
  • Added response schema examples for structured output
  • Updated documentation to reflect consistent behavior across modules

Migration Notes

For Existing Users

No breaking changes - all existing code continues to work. However, you can now use previously unsupported options:

# These options now work in all modules:
{:ok, response} = Gemini.generate("Explain AI", [
  response_schema: %{"type" => "object", "properties" => %{"summary" => %{"type" => "string"}}},
  response_mime_type: "application/json",
  stop_sequences: ["END", "STOP"],
  presence_penalty: 0.5,
  frequency_penalty: 0.3
])

# GenerationConfig structs now work consistently:
config = %Gemini.Types.GenerationConfig{
  temperature: 0.7,
  response_schema: %{"type" => "object"},
  max_output_tokens: 1000
}
{:ok, response} = Gemini.generate("Hello", generation_config: config)

0.1.0 - 2025-07-20

🎉 Major Release - Production Ready Multi-Auth Implementation

This is a significant milestone release featuring a complete unified implementation with concurrent multi-authentication support, enhanced examples, and production-ready telemetry system.

Added

🔐 Multi-Authentication Coordinator

  • Concurrent Auth Support: Enable simultaneous usage of Gemini API and Vertex AI authentication strategies
  • Per-request Auth Selection: Choose authentication method on a per-request basis
  • Authentication Strategy Routing: Automatic credential resolution and header generation
  • Enhanced Configuration: Improved config system with better environment variable detection

🌊 Unified Streaming Manager

  • Multi-auth Streaming: Streaming support across both authentication strategies
  • Advanced Stream Management: Preserve excellent SSE parsing while adding auth routing
  • Stream Lifecycle Control: Complete stream state management (start, pause, resume, stop)
  • Event Subscription System: Enhanced event handling with proper filtering

🎯 Comprehensive Examples Suite

  • telemetry_showcase.exs: Complete telemetry system demonstration with 7 event types
  • Enhanced demo.exs: Updated with better chat sessions and API key masking
  • Enhanced streaming_demo.exs: Real-time streaming with authentication detection
  • Enhanced multi_auth_demo.exs: Concurrent authentication strategies with proper error handling
  • Enhanced demo_unified.exs: Multi-auth architecture showcase
  • Enhanced live_api_test.exs: Comprehensive API testing for both auth methods

📊 Advanced Telemetry System

  • 7 Event Types: request start/stop/exception, stream start/chunk/stop/exception
  • Helper Functions: Stream ID generation, content classification, metadata building
  • Performance Monitoring: Live measurement and analysis capabilities
  • Configuration Management: Telemetry enable/disable controls

🔧 API Enhancements

  • Backward Compatibility Functions: Added missing functions (model_exists?, stream_generate, start_link)
  • Response Normalization: Proper key conversion (totalTokens → total_tokens, displayName → display_name)
  • Enhanced Error Handling: Better error formatting and recovery
  • Content Extraction: Support for both struct and raw streaming data formats

Changed

🏗️ Architecture Improvements

  • Type System: Resolved module conflicts and compilation warnings
  • Configuration: Updated default model to gemini-flash-lite-latest
  • Code Quality: Zero compilation warnings achieved across entire codebase
  • Documentation: Updated model references and improved examples

🔄 Example Organization

  • Removed Legacy Examples: Cleaned up simple_test.exs, simple_telemetry_test.exs, telemetry_demo.exs
  • Consistent Execution Pattern: All examples use mix run examples/[name].exs
  • Better Error Handling: Graceful credential failure handling with informative messages
  • Security: API key masking in output for better security

📝 Documentation Updates

  • README Enhancement: Added comprehensive examples section with detailed descriptions
  • Model Updates: Updated references to the latest Gemini models (Gemini 3 Pro Preview, 2.5 Flash/Flash-Lite) and new defaults
  • Configuration Examples: Improved auth setup documentation
  • Usage Patterns: Better code examples and patterns

Fixed

🐛 Critical Fixes

  • Type Module Conflicts: Resolved duplicate module definitions preventing compilation
  • Chat Session Context: Fixed send_message to properly handle [Content.t()] arrays
  • Streaming Debug: Fixed undefined variables in demo scripts
  • Response Parsing: Enhanced build_generate_request to support multiple content formats

🔧 Minor Improvements

  • Function Coverage: Implemented all missing backward compatibility functions
  • Token Counting: Fixed response key normalization for proper token count extraction
  • Stream Management: Improved stream event collection and display
  • Error Messages: Better error formatting and user-friendly messages

Technical Implementation

🏛️ Production Architecture

  • 154 Tests Passing: Complete test coverage with zero failures
  • Multi-auth Foundation: Robust concurrent authentication system
  • Advanced Streaming: Real-time SSE with 30-117ms performance
  • Type Safety: Complete @spec annotations and proper error handling
  • Zero Warnings: Clean compilation across entire codebase

📦 Dependencies

  • Maintained stable dependency versions for production reliability
  • Enhanced configuration system compatibility
  • Improved telemetry integration

Migration Guide

For Existing Users

# Old single-auth pattern (still works)
{:ok, response} = Gemini.generate("Hello")

# New multi-auth capability
{:ok, gemini_response} = Gemini.generate("Hello", auth: :gemini)
{:ok, vertex_response} = Gemini.generate("Hello", auth: :vertex_ai)

Configuration Updates

# Enhanced configuration with auto-detection
config :gemini_ex,
  default_model: "gemini-flash-lite-latest",  # Updated default
  timeout: 30_000,
  telemetry_enabled: true  # New telemetry controls

Performance

  • Real-time Streaming: 30-117ms chunk delivery performance
  • Concurrent Authentication: Simultaneous multi-strategy usage
  • Zero Compilation Warnings: Optimized build performance
  • Memory Efficient: Enhanced streaming with proper backpressure

Security

  • Credential Masking: API keys masked in all output for security
  • Multi-auth Isolation: Secure credential separation between strategies
  • Error Handling: No sensitive data in error messages

0.0.3 - 2025-07-07

Fixed

  • API Response Parsing: Fixed issue where usage_metadata was always nil on successful Gemini.generate/2 calls (#3)
    • The Gemini API returns camelCase keys like "usageMetadata" which were not being converted to snake_case atoms
    • Updated atomize_key function in coordinator to properly convert camelCase strings to snake_case atoms
    • Now properly populates usage_metadata with token count information
  • Chat Sessions: Fixed conversation context not being maintained between messages
    • The send_message function was only sending the new message, not the full conversation history
    • Now builds complete conversation history with proper role assignments before each API call
    • Ensures AI maintains context and remembers information from previous messages

0.0.2 - 2025-06-09

Fixed

  • Documentation Rendering: Fixed mermaid diagram rendering errors on hex docs by removing emoji characters from diagram labels
  • Package Links: Removed redundant "Documentation" link in hex package configuration, keeping only "Online documentation"
  • Configuration References: Updated TELEMETRY_IMPLEMENTATION.md to reference :gemini_ex instead of :gemini for correct application configuration

Changed

  • Improved hex docs compatibility for better rendering of documentation diagrams
  • Enhanced documentation consistency across all markdown files

0.0.1 - 2025-06-09

Added

Core Features

  • Dual Authentication System: Support for both Gemini API keys and Vertex AI OAuth/Service Accounts
  • Advanced Streaming: Production-grade Server-Sent Events (SSE) streaming with real-time processing
  • Comprehensive API Coverage: Full support for Gemini API endpoints including content generation, model listing, and token counting
  • Type Safety: Complete TypeScript-style type definitions with runtime validation
  • Error Handling: Detailed error types with recovery suggestions and proper HTTP status code mapping
  • Built-in Telemetry: Comprehensive observability with metrics and event tracking
  • Chat Sessions: Multi-turn conversation management with state persistence
  • Multimodal Support: Text, image, audio, and video content processing

Authentication

  • Multi-strategy authentication coordinator with automatic strategy selection
  • Environment variable and application configuration support
  • Per-request authentication override capabilities
  • Secure credential management with validation
  • Support for Google Cloud Service Account JSON files
  • OAuth2 Bearer token generation for Vertex AI

Streaming Architecture

  • Unified streaming manager with state management
  • Real-time SSE parsing with event dispatching
  • Configurable buffer management and backpressure handling
  • Stream lifecycle management (start, pause, resume, stop)
  • Event subscription system with filtering capabilities
  • Comprehensive error recovery and retry mechanisms

HTTP Client

  • Dual HTTP client system (standard and streaming)
  • Request/response interceptors for middleware support
  • Automatic retry logic with exponential backoff
  • Connection pooling and timeout management
  • Request validation and response parsing
  • Content-Type negotiation and encoding support

Type System

  • Comprehensive type definitions for all API structures
  • Runtime type validation with descriptive error messages
  • Request and response schema validation
  • Content type definitions for multimodal inputs
  • Model capability and configuration types
  • Error type hierarchy with actionable information

Configuration

  • Hierarchical configuration system (runtime > environment > application)
  • Environment variable detection and parsing
  • Application configuration validation
  • Default value management
  • Configuration hot-reloading support

Utilities

  • Content extraction helpers
  • Response transformation utilities
  • Validation helpers
  • Debugging and logging utilities
  • Performance monitoring tools

Technical Implementation

Architecture

  • Layered architecture with clear separation of concerns
  • Behavior-driven design for pluggable components
  • GenServer-based application supervision tree
  • Concurrent request processing with actor model
  • Event-driven streaming with backpressure management

Dependencies

  • req ~> 0.4.0 for HTTP client functionality
  • jason ~> 1.4 for JSON encoding/decoding
  • typed_struct ~> 0.3.0 for type definitions
  • joken ~> 2.6 for JWT handling in Vertex AI authentication
  • telemetry ~> 1.2 for observability and metrics

Development Tools

  • ex_doc for comprehensive documentation generation
  • credo for code quality analysis
  • dialyxir for static type analysis

Documentation

  • Complete API reference documentation
  • Architecture documentation with Mermaid diagrams
  • Authentication system technical specification
  • Getting started guide with examples
  • Advanced usage patterns and best practices
  • Error handling and troubleshooting guide

Security

  • Secure credential storage and transmission
  • Input validation and sanitization
  • Rate limiting and throttling support
  • SSL/TLS enforcement for all communications
  • No sensitive data logging

Performance

  • Optimized for high-throughput scenarios
  • Memory-efficient streaming implementation
  • Connection reuse and pooling
  • Minimal latency overhead
  • Concurrent request processing