Effect Ontology

A functional, type-safe system for extracting structured knowledge graphs from unstructured text using ontology-guided LLM prompting. Built with Effect-TS, implementing a mathematically rigorous pipeline based on topological catamorphism and monoid folding.

Mathematical Foundation

The system transforms OWL ontologies into LLM prompts via a topological catamorphism over a directed acyclic graph (DAG). The ontology is modeled as a dependency graph G = (V, E) where:

• Vertices (V): OWL classes, identified by IRIs
• Edges (E): rdfs:subClassOf relationships, oriented as Child → Parent
• Context (Γ): A mapping from nodes to their data (labels, properties, comments)

The prompt generation is defined as a fold over this graph using an algebra α:

α: D × List<R> → R

where D is the node data domain and R is the result monoid. The algorithm processes nodes in topological order, ensuring dependencies (subclasses) are computed before dependents (superclasses).

Result Monoid: The system uses a KnowledgeIndex monoid (HashMap-based) rather than string concatenation. This enables:

• Queryable structure: O(1) lookup by IRI instead of linear search
• Context pruning: Focus operations select relevant classes without dumping entire ontology
• Deferred rendering: Structure is preserved until final prompt assembly

The monoid operation is HashMap union with custom merge semantics, satisfying associativity and identity laws required for correct folding.

Why Effect

Effect provides the mathematical abstractions and type safety needed for this pipeline:

Typed Error Channels: The E channel in Effect<A, E, R> ensures all failure modes are explicit and composable. Graph cycles, missing nodes, LLM failures, and RDF parsing errors are tracked through the type system.

Dependency Injection: The R channel enables clean service composition via Layers. The extraction pipeline depends on LlmService, RdfService, and ShaclService, all provided through Effect's context system without global state or manual wiring.

Structured Concurrency: Effect's Fiber model provides cancellation and resource management. The extraction pipeline uses scoped services (PubSub) that automatically clean up when the Effect scope ends.

Referential Transparency: All operations are pure or explicitly effectful. The topological solver, algebra application, and prompt rendering are deterministic and testable without mocks.

Architecture

The pipeline follows a three-phase architecture:

Turtle RDF
  ↓ [Graph/Builder]
Graph<NodeId> + OntologyContext
  ↓ [Prompt/Solver + knowledgeIndexAlgebra]
KnowledgeIndex (HashMap<IRI, KnowledgeUnit>)
  ↓ [Prompt/Enrichment]
Enriched KnowledgeIndex (with inherited properties)
  ↓ [Prompt/Render]
StructuredPrompt
  ↓ [Prompt/PromptDoc]
Prompt String
  ↓ [Services/Llm]
KnowledgeGraph (JSON)
  ↓ [Services/Rdf]
N3.Store (RDF quads)
  ↓ [Services/Shacl]
ValidationReport + Turtle

Phase 1: Pure Fold - The graph solver applies the algebra in topological order, building a raw KnowledgeIndex with class definitions and structure (parent/child relationships).

Phase 2: Effectful Enrichment - The InheritanceService computes effective properties (own + inherited) for each class. This is separate from the fold because inheritance flows downward (parent → child) while the fold processes upward (child → parent).

Phase 3: Rendering - The enriched index is rendered to a StructuredPrompt, then to a formatted string using @effect/printer for declarative document construction.

Usage

Basic Extraction

import { ExtractionWorkflow, ExtractionWorkflowLive } from "@effect-ontology/core-v2"
import type { RunConfig } from "@effect-ontology/core-v2/Domain/Model/ExtractionRun"
import { Effect } from "effect"

const text = "Alice is a person who knows Bob. Bob works for Acme Corp."
const config: RunConfig = {
  ontologyPath: "./ontologies/foaf.ttl",
  concurrency: 4,
  chunking: {
    maxChunkSize: 800,
    preserveSentences: true
  }
}

const program = Effect.gen(function* () {
  const workflow = yield* ExtractionWorkflow
  return yield* workflow.extract(text, config)
}).pipe(
  Effect.provide(ExtractionWorkflowLive),
  Effect.scoped
)

const graph = await Effect.runPromise(program)
console.log(graph)

Expected Output

Input Text:

Alice is a person who knows Bob. Bob works for Acme Corp.

Generated Prompt (excerpt):

SYSTEM INSTRUCTIONS

Class: Person
Properties:
  - name (string)
  - knows (Person)

Class: Organization
Properties:
  - name (string)

TASK
Extract knowledge graph from the following text:
Alice is a person who knows Bob. Bob works for Acme Corp.

LLM Output (JSON):

{
  "entities": [
    {
      "@id": "_:person1",
      "@type": "http://xmlns.com/foaf/0.1/Person",
      "properties": [
        { "predicate": "http://xmlns.com/foaf/0.1/name", "object": "Alice" },
        { "predicate": "http://xmlns.com/foaf/0.1/knows", "object": { "@id": "_:person2" } }
      ]
    },
    {
      "@id": "_:person2",
      "@type": "http://xmlns.com/foaf/0.1/Person",
      "properties": [
        { "predicate": "http://xmlns.com/foaf/0.1/name", "object": "Bob" }
      ]
    },
    {
      "@id": "_:org1",
      "@type": "http://xmlns.com/foaf/0.1/Organization",
      "properties": [
        { "predicate": "http://xmlns.com/foaf/0.1/name", "object": "Acme Corp" }
      ]
    }
  ]
}

Final RDF (Turtle):

_:person1 a foaf:Person ;
    foaf:name "Alice" ;
    foaf:knows _:person2 .

_:person2 a foaf:Person ;
    foaf:name "Bob" .

_:org1 a foaf:Organization ;
    foaf:name "Acme Corp" .

LLM Integration

The system uses @effect/ai's LanguageModel.generateObject for structured output generation. The schema is dynamically generated from the ontology vocabulary:

const schema = makeKnowledgeGraphSchema(classIris, propertyIris)

This ensures the LLM can only emit entities with types and properties that exist in the ontology. The schema is a union of literal IRIs, providing type safety at both the schema level (Effect Schema validation) and the LLM level (structured output constraints).

The prompt is constructed from the KnowledgeIndex, which can be pruned using focus operations to reduce token usage. For example, if extracting only Person entities, the context can be limited to Person and its ancestors, excluding unrelated classes like Vehicle or Document.

Project Structure

packages/@core-v2/src/
  Domain/      # Schemas, models, and error types
  Service/     # LLM, RDF, NLP, extraction, entity resolution services
  Workflow/    # StreamingExtraction, TwoStageExtraction, EntityResolutionGraph
  Runtime/     # Production layer composition (tracing, rate limits, caches)
  Telemetry/   # OpenTelemetry attributes/exporters
  Prompt/      # Prompt helpers and renderers
  Schema/      # Shared schema definitions
  Utils/       # Common utilities

Testing

The codebase includes property-based tests verifying monoid laws, topological ordering guarantees, and inheritance correctness. All tests use Effect's test layer pattern for dependency injection.

Tracing

OpenTelemetry tracing can be enabled to capture LLM call metrics and performance insights:

# Enable tracing (default: true)
export TRACING_ENABLED=true

# Jaeger endpoint (default: http://localhost:14268/api/traces)
export JAEGER_ENDPOINT=http://localhost:14268/api/traces

Running Jaeger Locally

To visualize traces, run Jaeger using Docker:

docker run -d --name jaeger \
  -p 16686:16686 \
  -p 14268:14268 \
  jaegertracing/all-in-one:latest

Then view traces at http://localhost:16686

What Gets Traced

The extraction pipeline automatically annotates spans with:

• LLM Provider: Model name and provider (Anthropic, OpenAI, Google, etc.)
• Token Usage: Input/output token counts for cost tracking
• Estimated Cost: Calculated cost in USD based on token usage
• Extraction Metrics: Entity and triple counts per extraction
• Request Details: Prompt text and response text (optional)

This enables:

• Performance debugging of LLM calls
• Cost attribution and billing
• Bottleneck identification
• Quality monitoring of extractions

Disabling Tracing

To disable tracing:

export TRACING_ENABLED=false

Or omit the environment variable (tracing is enabled by default).

References

• Engineering Specification: docs/effect_ontology_engineering_spec.md - Formal mathematical specification
• Higher-Order Monoid: docs/higher_order_monoid_implementation.md - KnowledgeIndex architecture
• Effect Patterns: docs/effect-patterns/ - Idiomatic Effect-TS patterns used throughout