Skip to content

ROADMAP.md

Latest commit

 

History

History
80 lines (56 loc) · 2.58 KB

ROADMAP.md

File metadata and controls

80 lines (56 loc) · 2.58 KB
Raw

Roadmap

Current Priority

Ship the current Cortext release.

Near-term work should favor:

  • release readiness
  • stability
  • benchmark validation
  • documentation cleanup
  • product-path hardening

New retrieval architecture work is deferred until after shipping.

Post-Ship Candidate: ColBERT-Style Late Interaction

Use a Liquid AI CPU-friendly ColBERT model as a text reranker, not as a full replacement for the primary embedding stack.

Rationale:

  • improves exact text retrieval precision
  • preserves the current dense embedding pipeline for recall
  • avoids forcing a full token-matrix storage/index redesign before release

Phase 1: Retrieval Reranker

Primary insertion point:

  • src/operations/graph_retrieval.cpp

Planned integration:

  • keep current dense retrieval and graph expansion for recall
  • rerank top-N eligible text candidates after filter_candidates(...)
  • apply reranked relevance before select_diversified(...)
  • preserve existing association/label seeding and diversification behavior
  • fall back to current dense scoring for non-text or missing-text candidates

Deferred Semantic Evidence Scoring

Do not plan label, relation, or summary reranking until a production semantic evidence subsystem is explicitly reintroduced. The v1 runtime intentionally has no semantic extraction or summarization pass to hook into.

Explicitly Deferred

Do not do this before shipping:

  • replace the primary embedding model with ColBERT
  • make token-space late interaction a hard dependency of the live loop
  • replace the multimodal embedding path with a text-only ColBERT stack

Success Criteria For Post-Ship Evaluation

Measure whether the reranker improves:

  • exact entity retrieval
  • retrieval semantic overlap
  • interrupt precision and recall
  • summary/source grounding quality

Do not accept the change if it materially degrades:

  • per-turn latency
  • write-path simplicity
  • multimodal behavior

Release Positioning

Cortext should be treated as a v1 release candidate.

The current priority is shipping the existing architecture cleanly, documenting it clearly, and learning from real usage before taking on larger runtime substitutions.

Post-v1 Architecture Direction

Post-v1 work is expected to make two major infrastructure moves:

  • move the event-driven system to stateforward/sml.cpp
  • use that change to improve system structure and memory safety
  • move inference to stateforward/emel.cpp once that library is complete and production-ready

These are follow-on goals. They should not block the v1 release unless the current architecture shows a concrete release-blocking defect.