Core platform

Reasoning Engine

The reasoning engine turns questions into evidence-linked answers. It retrieves from the graph and documents, grounds each claim in a source, reasons with models and rules, and scores every claim with calibrated confidence.

ForDevelopersKnowledge engineersImplementation teams

The reasoning pipeline

Every reasoning call runs through the same deterministic pipeline. Retrieval is pluggable, but grounding, scoring, and citation are always enforced — the engine will not return a claim it cannot ground.

Reasoning pipeline

Retrieve → ground → reason → score → cite. Grounding and citation are non-optional.

Retrieval strategies

Retrieval is abstracted behind a service interface so the strategy can change without touching application code. Choose per request based on the workload.

Strategy	Best for	Notes
keyword	Exact terms, codes, tags	Fast, deterministic, no embeddings
semantic	Natural-language intent	Requires embeddings + vector index
hybrid	Most production workloads	Fuses keyword + semantic by score

Swap, don't rewrite

The retrieval layer is an interchangeable RetrievalService. Adding vector search later means implementing one interface and switching a strategy flag, not rewriting the reasoning path.

Calling the engine

reason.ts

const answer = await iios.reason({
  question: "Which PLANT-04 pumps are likely to fail this quarter?",
  scope: { site: "PLANT-04", type: "Asset" },
  strategy: "hybrid",        // keyword | semantic | hybrid
  minConfidence: 0.7,
})

for (const claim of answer.claims) {
  console.log(claim.text, claim.confidence, claim.evidence)
}

Confidence & grounding

Grounded retrieval

Each claim is linked to the graph entities and document spans it was derived from.

Calibrated scoring

The Confidence Engine calibrates scores against outcomes so 0.8 means the same thing everywhere.

Abstention

Below the confidence threshold the engine abstains and asks for evidence rather than guessing.