IA-RAG Core

Governed, structure-aware retrieval and execution planning for agentic systems.

IA-RAG Core makes the information architecture of a question explicit before execution. A domain describes entities, relations, capabilities, bounds, and world semantics. The runtime then turns a question into an inspectable semantic program, a typed logical plan, and one bounded executable plan before producing evidence, claims, a receipt, and a deterministic response.

Why information architecture comes first

Ordinary vector RAG is useful for ranking relevant text, but relevance is not the same as authoritative execution semantics. A similarity result does not by itself say which entities and relations are allowed, whether a traversal is complete, or what an empty result means.

IA-RAG Core puts those decisions in an explicit domain and plan. Retrieval strategies can supply candidates through neutral protocols; the canonical runtime still validates the structure, applies bounds, records evidence, and binds the result to the exact plan that ran.

Canonical pipeline

DomainPack
  -> PlannerRegistry
  -> Semantic interpretation
  -> CanonicalCompiler
  -> LogicalPlan
  -> deterministic lowering
  -> ExecutablePlan
  -> Admission and assurance
  -> PRE enforcement
  -> bounded execution
  -> POST enforcement
  -> Evidence and provenance
  -> Claims and truth state
  -> Receipt and lineage
  -> deterministic rendering

There is one selected planner, one logical-plan lineage, one executable plan, and one receipt bound to that executed plan. Plan and receipt digests are deterministic for the same query, domain configuration, and input facts.

Install and run

python -m pip install ia-rag-core

The offline reference demo is available from a repository checkout. From the checkout, install the development tools and run the tests and demo with the same Python interpreter:

python -m pip install '.[dev]'
python -m pytest -q
python examples/reference_demo.py

Using python -m pytest ensures the tests run with the same Python interpreter and virtual environment into which IA-RAG Core was installed.

The quick-start demo is offline, uses no external credentials, and runs on a small synthetic research-network domain. Its output includes the selected planner, plan digests, evidence, receipt, and response.

Truth and bounded execution

The reference domains demonstrate three explicit semantics:

• OPEN_WORLD: an empty search does not prove that a fact is false, so the result may be UNKNOWN.
• CLOSED_WORLD: within the declared domain and complete evidence boundary, an empty search can support a FALSE result.
• REQUIRED_SUBGRAPH: when a declared required relation or traversal cannot be established, execution remains incomplete and the runtime does not emit a positive claim from the partial result.

These semantics describe the evidence boundary and execution state. They do not guarantee truth about the real world beyond the supplied domain and evidence.

Feature claim ledger

Deliberate scope

The base distribution has zero runtime dependencies and contains only the canonical runtime, neutral contracts, deterministic reference assurance, in-memory execution, and synthetic reference domains. External services, credentials, application data, and deployment-specific integrations are not part of this candidate.

IA-RAG Core is licensed under the Apache License 2.0; see LICENSE.

See the public architecture and the package boundary for the supported scope.