Grounding & Citation Enforcement

What is Grounding?

In enterprise AI applications (legal discovery, medical diagnosis, financial auditing), generating an accurate answer is not enough.

Systems must be Factually Grounded:

Every single claim, number, and statement in the generated output must be traceable back to an explicit source document.

  UN-GROUNDED RESPONSE (High Risk / Failure):
  "Company revenue grew 15% in Q3." (Where did this number come from? No source!)

  GROUNDED RESPONSE WITH CITATIONS (Enterprise Compliant):
  "Company revenue grew 15% in Q3 [Doc 2, Page 4], driven by enterprise cloud sales [Doc 5]."

System Prompt Citation Engineering

To enforce citations, structure your system prompt with explicit document tagging and Few-Shot Exemplars:

You are a factual research assistant. 
Answer the user query using ONLY provided source passages inside <context> tags.

CRITICAL CITATION RULES:
1. Every factual statement MUST be followed by its source document ID in brackets [Doc X].
2. If context does not contain enough information to answer, state "Information not available in sources."
3. Do NOT use outside knowledge or un-cited assumptions.

<context>
<doc id="1">Acme Corp acquired WidgetTech for $50M in 2023.</doc>
<doc id="2">WidgetTech specializes in industrial IoT sensors.</doc>
</context>

User Query: What does WidgetTech do and when was it acquired?
Answer: WidgetTech specializes in industrial IoT sensors [Doc 2] and was acquired for $50M in 2023 [Doc 1].

Post-Generation Attribution Verification (NLI)

Even with strict system prompts, LLMs occasionally append fake citations (pointing to [Doc 1] for a fact that actually appears in [Doc 3]).

Production systems add an Automated Attribution Checker using Natural Language Inference (NLI):

  Generated Statement: "Acme Corp acquired WidgetTech for $50M [Doc 1]."
                             │
                             ▼
  NLI Model Checks:  Premise:   Content of Doc 1
                     Hypothesis: "Acme Corp acquired WidgetTech for $50M."
                             │
                             ▼
  NLI Classification Output:
  - ENTAILMENT  ──► Citation Verified True!
  - CONTRADICTION ──► Hallucinated Citation! Strip or Flag Response!

NLI models (like roberta-large-mnli) classify the relationship between source passage (Premise) and response claim (Hypothesis) as Entailment, Neutral, or Contradiction.

Citation & Grounding Metrics

1. Citation Precision: Percentage of generated citations that accurately contain the supporting fact ($|\text{Valid Citations}| / |\text{Total Citations}|$).
2. Citation Recall: Percentage of factual claims in the answer that have an accompanying citation ($|\text{Cited Claims}| / |\text{Total Claims}|$).
3. Groundedness Score (Ragas / TruLens): Fraction of response statements directly entailed by retrieved context passages.

Say this out loud

Factual Grounding ensures all generated claims are verifiable by retrieved context passages. System prompts enforce in text bracketed citations [Doc X] using explicit document ID tags and few shot exemplars. Production systems add NLI attribution checkers to verify entailment between source passages and response claims, flagging fake citations.

Followups to expect

1. What is Citation Hallucination? When an LLM generates a response with correct looking citation brackets [Doc 2], but the cited document Doc 2 does not contain the supporting evidence.
2. How does Constrained Decoding help Grounding? Using grammar-guided decoding (Jsonformer, Outlines) to force the model output structure to follow a rigid schema requiring { "claim": "...", "citation_doc_id": "..." } pairs.