Reasoning Search

Flag-gated preview

Ontology-grounded reasoning search is built and on main, gated behind the ONTOFORGE_SEARCH flag (default-off) — it ships dark and is not yet in a published release. Enable it locally with ONTOFORGE_SEARCH=1. Watch the repository for the release.

Ask your database a question in plain English and get a cited, reasoned answer — grounded in the ontology Ontoforge generates from your data, executed safely (read-only) against live records, running locally (Ollama) or in the cloud (OpenAI).

What it does

• Plans over the ontology, not raw columns — the model can only reference classes and properties that exist, so it can't invent tables or fields.
• Reasons, not just retrieves — derives non-obvious facts via Datalog rules, each with prov:wasDerivedFrom lineage.
• Multi-hop across real relationships — foreign-key neighbours are loaded automatically so rules can chain across tables.
• Cites everything — every claim maps to a source record; the executed read-only SQL is always shown.
• Stays safe — read-only, allow-listed, sensitivity-tier gated; safe to point at a live database.
• Live SPARQL — materializes the result subgraph as RDF and lets you query it.

Try it

ONTOFORGE_SEARCH=1 ontoforge-wizard --port 5051
# then open http://localhost:5051/ask

1. Pose a question

Pick a flavor and provider (Ollama / OpenAI), set your sensitivity-tier ceiling, and optionally build an RDF subgraph for live SPARQL.

2. Get a cited, reasoned answer

The answer cites its sources, shows the exact read-only SQL that ran, materializes the result subgraph, and streams the full reasoning trace (plan → execute → relations → reason → synthesize).

3. Interrogate the subgraph with live SPARQL

Query exactly what grounded the answer with a SPARQL SELECT over the materialized RDF triples.

Surfaces

Surface	How
SDK	from runtime.reasoning_search import search
CLI	ontoforge search "…" --flavor network-ops --provider ollama
HTTP	POST /api/search · streaming POST /api/search/stream (SSE)
SPARQL	POST /api/sparql over a materialized subgraph
Console	the /ask page (embedded as a wizard step)
Metrics	Prometheus ontomesh_search_* via /metrics (Grafana dashboard in deploy/monitoring/)

from runtime.reasoning_search import search

ans = search("Which alarms have a Critical perceived severity?",
             flavor="network-ops", db_path="db/demo.db",
             providers="ollama", max_tier="Restricted", materialize=True)
print(ans.answer)        # cited, reasoned prose
print(ans.executed_query)  # the read-only SQL that ran
print(ans.subgraph["count"])  # RDF triples in the result subgraph

Examples

"Which customers are at risk from the degraded core router crt-07, and why?" → derives the impacted customers by walking Resource → Service → Customer, cross-checks open alarms, and cites each source.

"Which enrolled subjects are now ineligible after their latest labs — and why?" → re-checks each subject's latest results against the trial's eligibility criteria, flags violations with audit-grade lineage, and keeps PHI on-host.

---

Want to shape the design? Open a discussion on the repo.