Skip to content

Roadmap

This roadmap describes where Ontographia is heading. It is organized into phases with rough dependencies, not dates. Each workstream is tracked as an epic issue (label epic); concrete work lives in its sub-issues.

Ontographia's core principle stays fixed throughout: the LLM produces only Intent JSON; a deterministic, ontology-constrained pipeline in the Rust core validates it and emits parameter-bound Cypher. The LLM never writes Cypher.

Evaluation results referenced by this roadmap are published in evaluation.md. The roadmap itself makes no performance claims.

Overview

Phase Theme Epics Depends on
0 Foundations #82, #83 —
1 pip-installable and agent-ready #84, #85 Phase 0
2 Familiar demos #86 Phase 0 (can run in parallel with Phase 1)
3 Reproducible public benchmarks #87 Phase 1 (Intent layer packaging)
4 Ecosystem integrations #88 Phase 1
5 Constraints from existing ontologies #89 Phase 0 (any time after) 
flowchart LR
  P0["Phase 0<br/>Foundations"] --> P1["Phase 1<br/>pip + MCP"]
  P1 --> P3["Phase 3<br/>Public benchmarks"]
  P1 --> P4["Phase 4<br/>Integrations"]
  P0 --> P2["Phase 2<br/>Sample-graph demos"]
  P0 --> P5["Phase 5<br/>OWL/SHACL constraints"]

Phase 0 — Foundations

Repo hygiene & public roadmap — #82

  • Close evaluation issues that are already complete
  • Publish this roadmap
  • Use milestones and epics; attach existing issues as sub-issues
  • Keep dependency updates moving; start a CHANGELOG.md

Getting started quality — #83

  • 5-minute quickstart: pip install ontographia → query the Movies graph
  • One-page explainer: why the LLM never writes Cypher
  • Troubleshooting page and auto-generated API reference
  • Fix outdated content in docs and README (Intent field names, sample outputs)

Phase 1 — pip-installable and agent-ready

Shipped in v0.1.6 (code merged in #93 and #94). Epics #84 and #85 stay open until that release is published.

Intent layer packaging (ontographia[llm]) — #84

  • ~~Promote the Intent layer from examples/llm/ (ontology subset prompts with full-schema fallback, execution-feedback refinement, difficulty-adaptive spend) into an optional extra, e.g. pip install "ontographia[llm]"~~ (done in v0.1.6 / #93)
  • ~~Pluggable LLM backend (OpenAI-compatible endpoint via openai SDK; LiteLLM extra reserved, not implemented)~~ (done in v0.1.6 / #93)
  • ~~Move the related tests into the main test suite~~ (done in v0.1.6 / #93)

MCP server — #85

  • ~~Tools: intent_schema, build_cypher, execute (read-only), list_ontology~~ (done in v0.1.6 / #94)
  • ~~Neo4j connection settings; launch via uvx; configuration examples for Claude Desktop and Cursor~~ (done in v0.1.6 / #94)
  • ~~Policy that rejects write queries; consistent with the existing Agent Skill~~ (done in v0.1.6 / #94)

Phase 2 — Familiar demos

Neo4j sample-graph quickstarts — #86

  • Ontologies and gold questions for Movies, Northwind, and Recommendations (reusing the Track B schema→COM converter)
  • Notebooks that connect directly to demo.neo4jlabs.com, plus Colab versions
  • Sub-issues: #57 (public domain graph), #48 (generalize Intent repair beyond manufacturing heuristics)

Phase 3 — Reproducible public benchmarks

Reproducible public benchmark suite — #87

  • Pin target databases and split of neo4j/text2cypher-2025v1; one command to reproduce
  • Direct LLM→Cypher baseline under identical model settings (#56)
  • Semantic-equivalence / soft execution matching for Cypher paraphrases (#73)
  • Safety metrics first: invented vocabulary, unbound literals, write-query leakage
  • Results page and regression checks in CI

Phase 4 — Ecosystem integrations

Framework integrations (LangChain / LlamaIndex / neo4j-graphrag) — #88

  • LangChain Tool / Retriever
  • LlamaIndex PropertyGraph retriever
  • neo4j-graphrag retriever adapter
  • An example and docs page for each; upstream docs listings where possible

Phase 5 — Constraints from existing ontologies

OWL/SHACL constraint enforcement — #89

  • Enforce SHACL constraints (minCount, maxCount, pattern, in) during Intent validation
  • Stricter OWL domain/range handling
  • Generate Neo4j constraints from the ontology
  • Import sample from a public ontology (e.g. a schema.org subset)
  • Related: #48 (mapping-driven repair / semantic validation)

Contributing

Pick an open sub-issue under any epic, or open a new issue and reference the relevant epic. See AGENTS.md and the docs for repository conventions.