Skip to main content
This document outlines the planned milestones for UOM. They are roughly ordered by impact-to-effort, but priorities shift with research needs.

​
1. Latency Optimization (Sandbox Caching)

Today each validation runs inside a clean Daytona container that resolves dependencies and downloads Maven/NuGet packages from scratch — a large part of the ~12-minute run time.
  • Goal: Pre-bake the standard Maven .m2 cache and NuGet fallback folders into the Daytona sandbox image layers.
  • Impact: Cut compile/validation time from minutes to under 10 seconds.

​
2. Faster Runs via Better Context Engineering

The translation prompt is the largest and slowest stage because it injects full configs, multiple few-shot examples, and complete harness skeletons. As target schemas grow, the injected mapping files also grow and threaten the context window.
  • Goal: Add retrieval-augmented context (e.g. a vector store over schema mappings) so only the relevant tables/entities are injected for a given query, rather than the whole mapping.
  • Goal: Trim and templatize few-shot examples and harness skeletons to reduce input tokens without losing grounding.
  • Impact: Lower per-run latency and cost, and support large schemas (thousands of entities) without hitting context limits. See Context Engineering.

​
3. More Translation Pairs (Bidirectional)

UOM currently supports translation in one direction only: .NET source → Java Spring Data target (enforced by the SourceFramework / TargetFramework enums).
  • Goal: Enable the reverse direction (Java Spring Data → .NET) and, eventually, arbitrary source↔target pairs within the supported set.
  • Goal: Surface the new directions as UI suggestion cards. See Design Decisions §2.3.

​
4. More Frameworks & Paradigms

  • Goal: Add ORM/ODM/OGM frameworks beyond the current five — e.g. Python (SQLAlchemy → Beanie/Motor), Node.js (Prisma → Mongoose), Go (GORM → Ent).
  • Goal: Add new target paradigms such as Relational → Time-Series (MSSQL → InfluxDB).
  • The extension path (enums, context files, prompts, validators) is documented in the Contribution Guide.

​
5. Extensive Evaluation Experiments

  • Goal: Build automated benchmark suites over the WideWorldImporters dataset and a curated query set.
  • Goal: Systematically measure pass/fail rates and latency across LLM backends (Ollama, e-INFRA, OpenAI, Anthropic) at temperature=0 to optimize model selection and prompt/context design.
  • Impact: Replace anecdotal confidence with measured accuracy, and ground future prompt/context changes in data.

​
6. Hardening the Human-in-the-Loop Stage

When MAX_TRANSLATION_LOOPS = 3 is reached, the graph routes to human_intervention_node, which raises a LangGraph interrupt(). This path is functional but under-tested.
  • Goal: Add end-to-end tests for the suspend/resume flow (accept, reject-with-feedback, loop-reset) to guarantee the graph reliably resumes from a checkpoint.
  • Goal: Build a richer in-IDE / in-UI widget so developers can edit failing target code in place, see side-by-side compiler errors and DeepDiff output, and submit corrections back into the validation loop.
  • See the User Guide §4 for the current behaviour.

​
7. Stability & Bug Fixing

  • Goal: Fix known edge cases in extraction (ambiguous multi-turn inputs), sandbox provisioning (transient Daytona/Docker timeouts), and equivalence checking (exotic type coercions).
  • Goal: Improve error surfaces so failures are actionable for users rather than raw stack traces.
  • Goal: Expand unit/integration coverage across nodes, validators, and the MCP adapters.

​
Tracking

Issues and progress are tracked on the project repository: github.com/corovcam/Universal-Object-Mapping.