TramAI - governed AI workflows for Java and Kotlin

Current Limitations

This page is intentionally blunt. It documents what TramAI does not do yet.

Status Level

TramAI is currently documented against version 0.4.0 — a production-quality pre-1.0 release line.

It is usable for:

  • typed service-style AI integration
  • structured extraction and classification
  • local and cloud provider experiments
  • Spring Boot and standalone integration
  • tests with deterministic fake providers
  • production pilots with resilience, observability, and orchestration

It is not yet a production-complete 1.0.

Not Implemented Yet

These features are not implemented in the current runtime:

  • provider-native structured output optimizations
  • generated proxy code or KSP compile-time processing

Partially Implemented Or Reserved

These concepts exist in the API shape or planning docs but are not fully realized:

  • OpenAI/Codex auth-file support exists, but it is experimental
  • streaming failover retries only before the first emitted token; TramAI does not attempt partial mid-stream recovery across providers
  • secret references are extensible through SecretValueResolver, but bundled AWS/Vault resolvers are not shipped yet
  • tramai-orchestration is shipped but should still be treated as experimental while its API surface settles

Practical Consequences

Before using TramAI in a serious service, assume you still need to make decisions about:

  • how aggressive your fallback topology should be for your workload
  • whether you want custom cloud secret resolvers beyond env: and file:
  • how much provider-specific behavior you are willing to accept

What Is Solid Already

These parts are already coherent and tested:

  • proxy generation
  • structured-output schema generation and retry flow
  • explicit provider registry behavior
  • provider retry behavior for transient failures
  • per-attempt timeout enforcement
  • raw text streaming with explicit terminal failure semantics
  • engine-owned tool calling
  • standalone builder
  • Spring integration
  • OpenTelemetry observer seam
  • OpenTelemetry metrics for attempt latency, token usage, parse failures, and engine events
  • engine-owned token budget controls based on provider-reported usage
  • deterministic test support

Recommended Usage Today

TramAI is in its best shape for:

  • internal tools
  • developer platforms
  • service-side extraction/classification workloads
  • early production pilots with clear guardrails

If you need heavy agent capabilities, conversation memory, or highly autonomous multi-agent behavior, wait for future milestones or build those layers explicitly on top.

Previous
Next