Roadmap¶
A living plan for the sensemaking harness. The canonical source is
ROADMAP.md in the
repo; this page mirrors it for the docs site.
Stance: ground every architectural choice in current best practice.
Web-search the specific target's primary docs at the planning step; prefer an
architectural fix over an expedient patch. Each committed design decision carries
a # research(YYYY-MM): provenance note.
Open architecture questions¶
Resolved (or partly resolved) by the best-practice research. Treat the still-open items as unsettled; don't harden code against one answer.
- MVP toolset boundary: smallest set of tools/skills/hooks that demonstrates end-to-end value; build vs. expose a stub for a sibling project.
- Graph / multi-hop reasoning: GraphRAG vs. agent-driven Cypher traversal vs. hybrid; entity resolution across stores. (Research: hybrid graph+text + agentic correction; GraphRAG for multi-hop only.)
- Connector strategy: MCP servers vs. in-process tools per store. (Research: MCP tool families per store.)
- Multimodal processing: extraction tools and cross-modal fusion. (Research: convert imagery/video to structured text, then reason.)
- Analytic rigor & eval: provenance, grounding, confidence, success metrics, structured analytic techniques. (Researched and shipped: citation gate, ICD-203 tradecraft lint, planted-needle + rubric eval, governance audit; see the analytic-rigor research.)
- Cloud vs. air-gapped fork: per-component swap points. (Resolved as a single seam at the orchestrator model; see ADR-0012. Remaining work is open-weight validation, not architecture.)
Phased build order¶
| Phase | Focus | Status |
|---|---|---|
| 0 | Scaffold & research | ✅ complete: toolchain, docs site, captured research |
| 1 | Single-store vertical slice | ✅ complete: Neo4j connector + entity-workup skill + provenance hook + ariadne workup (entity → cited note) |
| 2 | Heterogeneous retrieval | mostly complete: Postgres SQL connector + hybrid (full-text + vector, RRF) wired into the loop; dataset-agnostic seam; subagent fan-out deferred (ADR-0005) |
| 3 | Multimodal fusion | planned: multimodal-to-text extraction; cross-modal evidence fusion |
| 4 | Rigor, eval & integration | mostly complete: citation gate, ICD-203 tradecraft lint, planted-needle + rubric eval, reconciliation scoring, read-only governance gate, integration interfaces |
| 5 | Deployment hardening | mostly complete: cloud-vs-air-gapped fork resolved (ADR-0012); OpenTelemetry observability; model profiles (ADR-0013); published to PyPI as ariadne-sensemaking via trusted publishing; open-weight validation remains |
| 6 | Adaptive & self-improving harness | first slices complete (ADR-0020): adapt to a user's own Postgres (introspect, propose a mapping, ratify, run the existing pipeline unchanged); a declarative user ontology; a dynamic MCP surface; and bounded, audited self-improvement, distil / reflect / compare / distil --into (learn from a good run, reflect on a poor one, measure net effect, deepen a skill) on a propose-ratify-freeze spine |
Recently shipped¶
- 2026-06-08: Published to PyPI as
ariadne-sensemaking(GitHub trusted publishing); first-run UX polish (corrected quickstart paths, surfaced the interactive report, actionable run-dir errors, pre-flight store-reachability before the live loop). - 2026-06-07: Adaptive & self-improving harness (Phase 6, first slices): schema
introspection into a ratified mapping that the existing pipeline runs unchanged; a
declarative user ontology; a dynamic MCP surface; and bounded, audited self-improvement
(
distil,reflect,compare,distil --into) on a propose-ratify-freeze spine. - 2026-06-04/05: LLM-rubric analytic-standards eval, reconciliation scoring, read-only governance hard-fail gate, user-selectable model profiles, integration interfaces + reusable workflow patterns, OpenTelemetry observability.
- 2026-06-03: Dataset-agnostic pipeline (canonical schema + adapters), live indexing + hybrid full-text/semantic retrieval, Enron adapter.
- 2026-06-02: Heterogeneous (graph + SQL) retrieval in the live loop; planted-needle eval harness; citation gate v2 (recall + entailment).
- 2026-06-01: Repo scaffolded; Claude Agent SDK reference captured; best-practice research synthesized; Zensical docs site stood up.
The canonical ROADMAP.md
carries the full per-item ledger with research provenance.