API Reference
This is a lightweight overview of the public API. Full API documentation will expand as modules stabilize.
Core
ArgumentGraph: graph container and helpers (arglib/core/graph.py)ArgumentUnit,Relation,TextSpan,EvidenceItem: core data modelsArgumentGraph.save(path)writes JSON via the IO layer.ArgumentGraph.render(path, engine="graphviz")writes DOT output.
EvidenceCardandSupportingDocumentfor evidence pipelines.ArgumentBundleandArgumentBundleGraphfor argument-as-subgraph abstraction.ArgumentGraph.define_argument(...)andto_argument_graph()project bundles.
Reasoning
Reasoner: unified reasoning entrypoint (arglib/reasoning/reasoner.py)compute_credibility: warrant-gated credibility propagation (arglib/reasoning/credibility.py)explain_credibility: break down evidence and gated influences (arglib/reasoning/explain.py)
AI Evaluation
score_evidenceandvalidate_edgesprovide deterministic evaluation helpers.HeuristicEvaluatoris a baseline; LLM adapters can swap in later.
Critique
detect_patternsandapply_gate_actionsfor flaw detection and gate invalidation.analyze_warrant_fragilityto identify critical warrants per edge.
Mining
ArgumentMiner,SimpleArgumentMiner, andLongDocumentMinerfor extraction pipelines.ClaimRelationTaggerfor fast span-based claim typing and support/attack tagging.HybridClaimRelationTaggerfor small-model inference with deterministic fallback.HybridClaimRelationTaggersupports optional neural backend vianeural_model_dir.HybridClaimRelationTaggersupports optional Hugging Face artifact auto-download (auto_download_artifacts,artifact_manifest_path,artifact_cache_dir).HybridClaimRelationTagger.predict_evidence_links(...)for claim-evidence retrieval + stance inference.ConversationMemoryfor incremental chat-turn ingestion, claim dedup, and contradiction links.Splitter,ParagraphSplitter,FixedWindowSplitterfor long-document chunking.MergePolicyandSimpleGraphReconcilerfor merging per-segment graphs.token_jaccard_similarityas a simple dedup/coreference heuristic.LLMClient/LLMHookand async variants for model-backed parsing or splitting.OllamaClientfor local model inference.AsyncArgumentMinerAdapterto wrap sync miners for async pipelines.
I/O
dumps/loads: JSON serialization (arglib/io/json.py)validate_graph_payload: minimal validation (arglib/io/schema.py)validate_chat_grounded_payloadandvalidate_chat_exploration_payload: dataset row validation (arglib/io/datasets.py)validate_chat_deferred_relevance_payload: deferred-relevance dataset validation (arglib/io/datasets.py)validate_chat_grounded_evidence_payload: grounded-evidence dataset validation (arglib/io/datasets.py)
Visualization
to_dot: Graphviz DOT export (arglib/viz/graphviz.py)
Integrations
run_serverinarglib/integrations/demo_ui.pypowers the local demo UI (arglib demo-ui).scripts/publish_hf_artifacts.pyuploads heavy artifacts to Hugging Face and writesarglib/data/hf_artifacts.json.scripts/download_hf_artifacts.pydownloads artifacts from the manifest.
Training
TASK_REGISTRY,TaskConfig,TrainingConfig, andtrain_multitaskinarglib/ai/training.JsonlTaskDatasetandMultiTaskRoundRobinLoaderfor task shard loading and round-robin batch scheduling.scripts/train_neural_tagger.pytrains transformer models for sequence + token tasks:span_tagging,claim_type,relation_class,coref_linkdeferred_pending,deferred_resolution,discourse_function,contradiction_nliclaim_evidence_stance,evidence_retrieval