rdflib-reasoning-middleware 0.4.0

LangChain Custom Middleware for Semantic Web Reasoning

RDFLib Reasoning Middleware

This package exposes this repository's functionality to DeepAgents (runtime/Research Agents) as Custom Middleware.

• GraphBacked and StructuralElement instances (from rdflib-reasoning-axioms) MUST be usable as tool argument/response models and as values carried inside middleware state. They are stateless, immutable domain snapshots.
• GraphBacked and StructuralElement MUST NOT subclass LangGraph AgentState. Middleware MAY map them into AgentState fields (e.g. as payloads or lists of axioms).

Demo notebooks

The package includes checked-in demonstration notebooks:

• demo-baseline-ontology-extraction.ipynb: prompt-only baseline for RDF extraction without middleware (negative control)
• demo-dataset-middleware.ipynb: end-to-end walkthrough of DatasetMiddleware plus live notebook tracing of agent activity
• demo-vocabulary-middleware.ipynb: DatasetMiddleware combined with RDFVocabularyMiddleware for vocabulary-grounded extraction

DatasetMiddleware tutorial

DatasetMiddleware is the baseline middleware layer for giving a Research Agent an RDF-backed working memory.

Current baseline capabilities:

• list triples in the default graph
• add triples to the default graph
• remove triples from the default graph
• serialize the default graph as RDF text
• reset the in-memory dataset session

Minimal usage:

from deepagents import create_deep_agent
from rdflib import Namespace
from rdflib_reasoning.middleware import (
    DatasetMiddleware,
    DatasetMiddlewareConfig,
    VocabularyConfiguration,
    VocabularyDeclaration,
)

vocabulary_context = VocabularyConfiguration.bundled_plus(
    VocabularyDeclaration(prefix="ex", namespace=Namespace("urn:example:"))
).build_context()

agent = create_deep_agent(
    model=llm,
    middleware=[
        DatasetMiddleware(
            DatasetMiddlewareConfig(vocabulary_context=vocabulary_context)
        )
    ],
)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "Represent this text as RDF."}]}
)

Coordinated task-specific vocabulary setup:

from rdflib import FOAF, Namespace
from rdflib_reasoning.middleware import (
    DatasetMiddleware,
    DatasetMiddlewareConfig,
    RDFVocabularyMiddleware,
    RDFVocabularyMiddlewareConfig,
    VocabularyConfiguration,
    VocabularyDeclaration,
)
from rdflib_reasoning.middleware.namespaces.spec_cache import UserVocabularySource

EX = Namespace("urn:example:vocab#")

vocabulary_context = VocabularyConfiguration.bundled_plus(
    VocabularyDeclaration(
        prefix="ex",
        namespace=EX,
        user_spec=UserVocabularySource(
            graph=domain_terms,
            vocabulary=EX,
            description="Task-specific terms for this extraction run.",
        ),
    )
).build_context()

# Opt in to the extended Dublin Core vocabularies when the task benefits from them.
vocabulary_context = VocabularyConfiguration.bundled_plus(
    VocabularyDeclaration(
        prefix="ex",
        namespace=EX,
        user_spec=UserVocabularySource(
            graph=domain_terms,
            vocabulary=EX,
            description="Task-specific terms for this extraction run.",
        ),
    )
).plus_dublin_core().build_context()

agent = create_deep_agent(
    model=llm,
    middleware=[
        DatasetMiddleware(
            DatasetMiddlewareConfig(vocabulary_context=vocabulary_context)
        ),
        RDFVocabularyMiddleware(
            RDFVocabularyMiddlewareConfig(vocabulary_context=vocabulary_context)
        ),
    ],
)

The middleware contributes both tool implementations and capability-specific system guidance. The Research Agent sees a tool surface such as add_triples and serialize_dataset, while the live RDFLib dataset remains middleware-owned runtime infrastructure.

Dataset tool inputs accept RDF terms in canonical N3 form. For IRIs, the middleware also accepts bare RFC 3987 IRI strings such as urn:example:Person as an input convenience and serializes them back in canonical N3 form such as <urn:example:Person>.

When vocabulary middleware is also present, the intended workflow is to search for a fitting indexed term first, inspect a candidate only when confirmation is needed, and then continue modeling rather than manually browsing large vocabulary slices by default.

For single-pass or notebook-style harnesses, you MAY also add ContinuationGuardMiddleware to provide continuation control for single-run harnesses. It can re-prompt after unfinished planning or recovery narration, switch into a final-answer-only mode after strong completion signals, and stop deterministically once a valid Turtle answer is present. This middleware is role-aware about when it injects HumanMessage reminders and does not force a tool -> user transition for providers that reject that message order. This middleware is optional and is usually unnecessary for deliberately multi-round threaded agents.

Tracing tutorial

The tracing support is split into a core callback-based recorder and an optional notebook renderer.

• Core tracing lives in rdflib_reasoning.middleware.tracing and does not require IPython
• Notebook rendering lives in rdflib_reasoning.middleware.tracing_notebook and is optional via the notebook extra

Notebook example:

from deepagents import create_deep_agent
from rdflib import Namespace
from rdflib_reasoning.middleware import (
    ContinuationGuardMiddleware,
    DatasetMiddleware,
    DatasetMiddlewareConfig,
    VocabularyConfiguration,
    VocabularyDeclaration,
)
from rdflib_reasoning.middleware.tracing_notebook import LiveNotebookTrace

vocabulary_context = VocabularyConfiguration.bundled_plus(
    VocabularyDeclaration(prefix="ex", namespace=Namespace("urn:example:"))
).build_context()

with LiveNotebookTrace(heading="Dataset Trace") as trace:
    agent = trace.attach(
        create_deep_agent(
            model=llm,
            middleware=[
                DatasetMiddleware(
                    DatasetMiddlewareConfig(vocabulary_context=vocabulary_context)
                ),
                ContinuationGuardMiddleware(),
            ],
        )
    )
    result = agent.invoke(
        {"messages": [{"role": "user", "content": "Represent this text as RDF."}]}
    )

The notebook renderer is intended for tutorial and demo scenarios. It presents model decisions, tool calls, tool results, and final responses as a live-updating notebook view without making IPython a required runtime dependency for the package. ContinuationGuardMiddleware is an optional companion for these scenarios when the harness expects the Research Agent to keep acting until a completed answer and needs provider-safe continuation rather than assistant-final implicit continuation.

Proof Evaluation Harness

This package is also the intended home for a reusable proof evaluation harness used by Development Agents to assess DirectProof outputs produced in experiments.

• The harness SHOULD expose framework-agnostic Pydantic models for evaluation requests and responses.
• The harness SHOULD support a prompt and structured-output contract that can be used without committing to a specific orchestration framework.
• The harness MAY include thin LangChain or LangGraph glue code, but a full evaluator agent is not required for the baseline.
• The baseline notebook SHOULD consume the harness; it SHOULD NOT define the evaluation contract inline.

Initial proof evaluation dimensions

The initial harness is intentionally narrow. It is meant to support the baseline proof notebook before full axiomatization or a complete reasoning profile exists.

Dimension	Purpose
Groundedness	Check whether proof elements are supported by the input document
Hallucination detection	Flag entities or relationships not grounded in the input document
Missing support	Identify support the proposed proof appears to require but does not provide
Baseline verdict	Decide whether the proposed proof is acceptable for the baseline task
Optional rationale	Preserve a short explanation or typed error labels for later analysis

Feature Matrix

Status values:

• Implemented: available in the package today
• Not started: identified feature target with no concrete implementation yet
• Out of scope: intentionally not roadmapped for the current implementation phase

The matrix below reflects the package's current implemented state. Some entries are part of the broader planned 0.3.0 target but are intentionally still in progress.

Dataset middleware

Feature	Status	Notes
Default-graph triple access	Implemented	0.2.0 baseline: list, add, and remove triples in the default graph
Default-graph serialization	Implemented	0.2.0 baseline: serialize current state as RDF text for inspection
Dataset reset	Implemented	0.2.0 baseline escape hatch for clearing the middleware-owned dataset session
Shared dataset runtime injection	Implemented	DatasetRuntime can be injected explicitly when multiple middleware components need the same live dataset boundary
Namespace whitelisting: enforcement	Implemented	0.3.0: reject URIs from non-whitelisted namespaces in add_triples
Namespace whitelisting: enumeration	Implemented	0.3.0: include allowed vocabulary list in middleware prompt when enabled
Namespace whitelisting: remediation	Implemented	0.3.0: suggest nearest valid term via Levenshtein distance for closed-vocabulary near-misses
Named graph management	Not started	Later phase: list graphs, create named graphs, and remove named graphs
Graph-scoped triple access	Not started	Later phase: extend triple tools with an optional graph/context argument while keeping the default graph as the default target
Dataset quad access	Not started	Later phase: explicit quad-level CRUD across graphs
Fully general dataset lifecycle	Not started	Later phase: broader create/load/save/delete and other dataset-wide concerns beyond the initial in-memory baseline
RDF 1.2 triple statement support	Out of scope	Forward-looking placeholder for RDF 1.2 quoted or triple-term support

Tracing and presentation

Feature	Status	Notes
Trace recording callbacks	Implemented	Normalized capture of model and tool lifecycle events for LangChain-based runs
Notebook-rendered tracing of agent activity	Implemented	Optional IPython-backed live rendering of model decisions, tool calls, tool results, and final responses
Serialized trace transcript export	Implemented	Versioned JSON-friendly turn-trace transcript export via turn_traces_to_json_document(...) and LiveNotebookTrace.dump(...)
Continuation guard middleware	Implemented	Optional continuation-control middleware for single-run harnesses with provider-safe re-prompting and deterministic end on valid final Turtle
Additional non-notebook rich trace renderers	Not started	Future console, HTML, or other presentation layers over the same core trace sink

RDF vocabulary middleware

Feature	Status	Notes
List indexed vocabularies	Implemented	list_vocabularies tool: enumerate available vocabulary namespaces with labels and term counts
List vocabulary terms	Implemented	list_terms tool: list indexed terms from a vocabulary, filterable by term type and pageable with offset / limit
Search vocabulary terms	Implemented	search_terms tool: search indexed terms by intended meaning as the primary discovery path
Inspect term	Implemented	inspect_term tool: return a compact normalized description of one indexed term, optionally including native RDF with transitive rdfs:subClassOf / rdfs:subPropertyOf paths
Whitelist-aware vocabulary filtering	Implemented	list_vocabularies, list_terms, and inspect_term honor the injected VocabularyContext policy and index set
Unified vocabulary configuration	Implemented	VocabularyConfiguration.build_context() produces the single runtime VocabularyContext injected into both middleware components
Incremental vocabulary extension	Implemented	VocabularyConfiguration.plus(...), plus_dublin_core(), and plus_vann() return extended immutable configurations
Using VANN annotation metadata	Implemented	list_vocabularies surfaces advisory vann:preferredNamespacePrefix / vann:preferredNamespaceUri metadata for bundled and user-supplied vocabularies without changing namespace policy
Bundled RDF specification	Implemented	Included in the default VocabularyConfiguration.bundled_plus(...) set
Bundled RDFS specification	Implemented	Included in the default VocabularyConfiguration.bundled_plus(...) set
Bundled OWL specification	Implemented	Included in the default VocabularyConfiguration.bundled_plus(...) set
Bundled PROV-O specification	Implemented	Included in the default VocabularyConfiguration.bundled_plus(...) set
Bundled SKOS specification	Implemented	Included in the default VocabularyConfiguration.bundled_plus(...) set
Opt-in bundled VANN specification	Implemented	Available through explicit declaration or VocabularyConfiguration.plus_vann() but not included by default
Opt-in bundled Dublin Core extensions	Implemented	VocabularyConfiguration.plus_dublin_core() adds DCAM, DCMITYPE, and DCTERMS as an explicit grouped extension
Additional bundled vocabularies	Implemented	Optional bundled additions currently include VANN and grouped Dublin Core extensions beyond the default indexed set
User-supplied vocabularies at construction	Implemented	VocabularyDeclaration(user_spec=UserVocabularySource(...)) feeds VocabularyConfiguration, and the resulting VocabularyContext indexes those declarations explicitly

Dataset middleware implementation pattern

• Core dataset behavior SHOULD be implemented in internal middleware methods using RDFLib-native values.
• Research Agent-facing tools SHOULD remain thin adapters over those internal methods.
• Tool descriptions and schema models together form the runtime boundary contract; they SHOULD stay distinct from the middleware's internal implementation API.

Knowledge retrieval middleware

Feature	Status	Notes
Remote entity search	Not started	Resolve labels or names to candidate external entities; initial provider target is DBpedia, with Wikidata open for later
Remote entity describe	Not started	Fetch RDF descriptions for resolved entities from remote providers
External RDF import	Not started	Fetch RDF over HTTP and normalize it into local graph or dataset state
Entity resolution pipeline	Not started	Search, disambiguate, fetch, and attach provenance for external entities
Site metadata extraction	Not started	Fetch HTML and extract embedded JSON-LD or related structured metadata for import

Inference middleware

Feature	Status	Notes
Reasoner-backed dataset interface	Not started	Middleware surface for datasets backed by an inference engine
Inference execution	Not started	Trigger rule materialization against current dataset state
Direct proof generation	Not started	Return a DirectProof for a supplied claim, structural element, or contradiction
Derivation trace access	Not started	Expose rule-level derivation traces produced by the engine
Proof rendering	Not started	Render canonical proof objects into notebook-friendly or other human-facing views without changing the underlying proof data
Consistency / contradiction checks	Not started	Detect and report contradictions or unsatisfiable states

Concrete middleware profiles

Feature	Status	Notes
RDFS middleware profile	Not started	Reasoner-backed dataset middleware configured for RDFS entailment
OWL 2 RL middleware profile	Not started	Reasoner-backed dataset middleware configured for OWL 2 RL entailment