Atomic Fact Generation

An atomic fact must be interpretable without relying on surrounding context or external references. It cannot contain pronouns with ambiguous antecedents, unresolved anaphora, or implied subjects that require prior sentences for resolution.

• Bad: 'It was released in 2017.' (What is 'it'?)
• Good: 'The Transformer architecture was introduced in the paper "Attention Is All You Need" in 2017.'

This property ensures that when a fact is retrieved in isolation by a vector store, the language model can ground on it without hallucinating missing context.

Each atomic fact expresses exactly one proposition that can be evaluated as true or false against a source document. Compound sentences containing conjunctions, disjunctions, or multiple clauses must be decomposed.

• Compound: 'BERT uses bidirectional attention and is pre-trained on masked language modeling and next sentence prediction.'
• Atomic 1: 'BERT uses bidirectional attention.'
• Atomic 2: 'BERT is pre-trained on masked language modeling.'
• Atomic 3: 'BERT is pre-trained on next sentence prediction.'

This granularity enables contradiction detection at the claim level rather than the document level.

Every atomic fact must explicitly name the subject entity to which the claim pertains. Implicit subjects or passive constructions that obscure agency are invalid.

• Vague: 'Training was performed on 16 TPUv3 pods.'
• Atomic: 'Google trained the T5 model on 16 TPUv3 pods.'

Entity-bound facts enable provenance tracking by linking each claim to a specific entity in a knowledge graph, supporting downstream tasks like entity salience scoring and citation graph construction.

Where applicable, atomic facts must include explicit temporal markers indicating when the stated fact was true. This prevents stale information from being treated as current during retrieval.

• Ungrounded: 'GPT-3 is the largest language model available.'
• Grounded: 'As of June 2020, GPT-3 was the largest publicly known language model with 175 billion parameters.'

Temporal grounding supports content freshness filtering in retrieval pipelines, ensuring time-sensitive queries receive chronologically appropriate facts.

Atomic facts should minimize lexical dependencies on the original document's specific phrasing. While preserving factual accuracy, the fact should be expressed in canonical, normalized language to improve embedding alignment with diverse query formulations.

• Original: 'The darn thing achieved SOTA on a boatload of NLP benchmarks.'
• Atomic: 'The T5 model achieved state-of-the-art results on multiple NLP benchmarks including GLUE, SuperGLUE, and SQuAD.'

This normalization increases the cosine similarity between the fact embedding and a wider range of user queries, improving recall in vector search.

Each atomic fact must retain a pointer to its originating source document and, ideally, the specific passage or chunk from which it was extracted. This metadata enables citation accuracy and attribution fidelity in downstream generation.

• Metadata: { source_doc: 'attention_is_all_you_need.pdf', chunk_id: 'chunk_42', span: [120, 245] }

Provenance binding transforms the RAG pipeline from a black-box generator into an auditable system where every output claim can be traced back to its evidentiary basis.