0.3.0 — Memory, Multimodal, RAG, and Provider Expansion

0.3.0 is the repository milestone that expanded TramAI beyond the original typed service, provider, and orchestration surfaces into memory, multimodal input, RAG, vector stores, and a broader provider matrix.

New Modules

tramai-memory

Production-ready chat memory primitives for multi-turn AI services.

• MessageWindowChatMemory for bounded recent-message windows
• TokenAwareChatMemory for token-budget-aware conversation history
• PersistentChatMemory for memory backed by an external store
• system-message deduplication and deterministic eviction behavior

tramai-memory-store

External persistence SPI support for chat memory.

• durable conversation storage boundary
• JDBC-backed store support
• table DDL helpers for operational setup

tramai-rag

Retrieval-augmented generation pipeline support.

• document loading
• chunking
• retrieval
• context injection into AI operations

tramai-embedding

Embedding model SPI and provider-backed embedding implementations.

• OpenAI embedding model support
• Ollama embedding model support
• registry support for embedding use cases

tramai-vectorstore-spi

Vector store abstractions for semantic retrieval.

• text embedding storage contracts
• metadata-aware query models
• in-memory implementation for tests and simple local flows

tramai-vectorstore-chroma

ChromaDB vector store adapter.

• collection creation and lookup
• document upsert
• similarity query support

tramai-vectorstore-pgvector

PostgreSQL pgvector adapter.

• relational vector storage
• pgvector-backed similarity queries
• fit for applications already using PostgreSQL

tramai-azure-openai

Azure OpenAI provider integration.

• Azure endpoint support
• API-key authentication
• model request and response mapping
• usage metric extraction

tramai-bedrock

AWS Bedrock provider integration.

• Bedrock runtime request mapping
• provider error normalization
• image-capable request support where supported by the target model

tramai-gemini

Google Gemini provider integration.

• Gemini request and response mapping
• multimodal content support
• provider error normalization

tramai-deepseek

DeepSeek provider integration.

• DeepSeek chat API support
• provider error normalization
• OpenAI-style request semantics where compatible

Core Features

Multimodal Content

Core message modeling now supports additive content parts instead of only plain text.

• ContentPart
• TextPart
• ImagePart
• ImageUrlContent
• ImageDetail with LOW, HIGH, and AUTO

The engine validates provider capabilities before execution so image input sent to a non-vision provider fails explicitly.

Image Downloading

TramAI includes built-in image download support for URL-backed image content.

• 20MB response-size limit
• 10s connect timeout
• 30s request timeout
• MIME detection from URL extension

Usage Metrics

Usage reporting was expanded for multimodal workloads.

• image count tracking
• estimated image token tracking
• provider-specific usage normalization

Provider Updates

The existing providers were updated for multimodal serialization and capability reporting.

• OpenAI
• Azure OpenAI
• Anthropic
• Bedrock
• Gemini
• Ollama
• DeepSeek

Observability Updates

Worker and distributed execution observability gained broader event coverage.

• shutdown started
• drain progress
• lease renewed
• worker heartbeat
• workflow abandoned
• graceful shutdown bounds

Notes

• TramAI 0.3.x targets Java 21+.
• Structured output remains the default contract for non-String return types.
• Provider routing remains explicit and registry-based.
• Runtime and platform modules remain optional.