Grounding

Grounding mandates that every factual claim generated by a model is explicitly linked to a canonical source document or structured knowledge base entry. This transforms a probabilistic text generation system into an auditable information retrieval pipeline.

• Citation Integrity: Outputs include direct pointers to source paragraphs, page numbers, or database rows.
• Provenance Tracking: The lineage of a fact—from raw document ingestion to final output—is fully traceable.
• Example: In legal AI, a grounded statement about a contract clause must cite the specific PDF filename, page, and paragraph from which the obligation was extracted.

Grounding resolves the symbol grounding problem by connecting abstract tokens in a model's vocabulary to unique, disambiguated identifiers in a knowledge graph or database.

• Named Entity Linking (NEL): Textual mentions like 'Apple' are resolved to a specific QID in Wikidata or a legal entity identifier in a corporate registry.
• Canonicalization: Synonymous terms and aliases are mapped to a single authoritative node, preventing semantic fragmentation.
• Example: The phrase 'the landlord' in a lease agreement is grounded to the specific legal person entity 'John Doe, LLC' with a unique URI in the contract knowledge graph.

RAG is the primary architectural pattern for implementing grounding. Instead of relying on parametric memory, the system retrieves relevant chunks of text from a vector database or triplestore and conditions generation strictly on that evidence.

• Hybrid Retrieval: Combines dense vector search with sparse keyword retrieval (BM25) to maximize recall of relevant source material.
• Closed-Book vs. Open-Book: Grounded systems operate in an 'open-book' mode, refusing to answer from internal weights alone.
• Example: A legal RAG system answers a question about a merger clause by first retrieving the exact clause text from the contract repository, then synthesizing an answer that quotes the source directly.

Grounding extends beyond text retrieval to enforce structural and semantic consistency with a formal knowledge model. Outputs are validated against an ontology or schema before presentation.

• SHACL Validation: RDF graphs generated by the AI are checked against shape constraints to ensure they conform to the expected legal ontology.
• Deontic Logic Consistency: Normative statements about obligations and permissions are checked for logical contradictions against the existing rule base.
• Example: If a model extracts a contract termination date, a grounding layer validates that the date conforms to xsd:date format and that the termination event does not precede the effective date in the temporal reasoning graph.

The primary operational goal of grounding is the systematic elimination of factual hallucination. A grounded system is designed to fail gracefully by expressing uncertainty rather than fabricating a plausible-sounding but false answer.

• Abstention Triggers: If no source document supports a claim above a confidence threshold, the system outputs 'Information Not Found' instead of guessing.
• Citation Precision: Grounded outputs are evaluated using metrics like citation recall and citation precision, measuring whether every sentence is backed by a source.
• Example: A grounded legal summarizer will state 'The contract does not specify a governing law clause' rather than hallucinating a common jurisdiction like Delaware.

Grounding is not limited to text. In legal and enterprise contexts, it connects textual analysis to non-textual evidence such as tabular data, images, and structured metadata.

• Table Grounding: A statement about a financial penalty is grounded to a specific cell in a structured damages schedule.
• Image Grounding: A description of a signature block is grounded to the bounding box coordinates on the original scanned PDF page.
• Example: A due diligence AI grounds its claim that 'the signature is missing on page 12' by linking the text output to the visual region of the document image where the anomaly was detected.