This roadmap keeps long-horizon research and engineering directions visible without turning every idea into an open issue. Open issues should stay reserved for work that is specific enough to start, review, or close.
These are the near-term threads worth keeping as open issues:
- Generator capacity and grounding disambiguation: #4
- MS MARCO-tuned dense encoder comparison: #15
- Artifact registry and lockfile hardening: #18
- Repeatable error-analysis workflow: #19
- End-to-end inference API packaging: #25
- Ablation runners and report tables: #27
- Retrieval and generation profiling: #29
- Static typing and stricter linting: #50
- External versioning for large data and index artifacts: #123
- RAG input hardening and prompt-injection guardrails: #124
- Statistical regression gates for metric deltas: #125
- Experiment-to-report artifact automation: #126
Future research directions that should become issues only when the experiment contract, dataset slice, and acceptance criteria are clear:
- Citation-aware generation and citation-aware decoding.
- NLI and alternative faithfulness metric stress tests.
- Full-sample semantic metrics and judge-based factual consistency protocols.
- Query robustness, unanswerable-query handling, and calibrated abstention.
- Alternative retrieval families such as SPLADE, BGE, ColBERT, and semantic or parent-child chunking.
- Adaptive retrieval sufficiency checks and richer reranker error slicing.
- Larger generator comparisons under matched RAG conditions.
- Optional adapter layers for RAGAS, TruLens, or similar metric suites.
- Full-corpus retrieval leaderboard matrices once storage and runtime are better managed.
The serving direction should grow from the current package and API surface rather than becoming a separate application too early:
- Keep the Python API as the canonical first serving surface.
- Add streaming only after request validation, error payloads, and artifact lookup are stable.
- Treat vector-store backends, Docker Compose, and a Go gateway as later integration work, not immediate project scope.
- A lightweight demo UI is useful after the API path is stable, but it should not replace the batch-evaluation focus of the repository.
Infrastructure work should reduce reproduction friction or review time:
- Separate offline indexing from online retrieval once artifact versioning is settled.
- Extract a shared runner base only after the current entry-point contracts have stabilized enough that a refactor will reduce review burden rather than hide behavioral changes.
- Publish generated API documentation after the public package interfaces stop moving quickly.
- Extend report generation only where it removes manual copy/paste between metrics artifacts, tables, and written reports.
- Keep open issues actionable and closeable.
- Move broad research directions here until they have a concrete experiment contract.
- Prefer one issue per reproducible comparison, runner, report, or validation gate.
- Close roadmap-only issues with a short note pointing back to this file.