Chunk Graph

Each node in a chunk graph represents a discrete text segment—typically a paragraph, section, or atomic fact. Nodes carry metadata (source document, position, heading level) and the raw text payload. Unlike flat vector indexes where chunks float in isolation, graph nodes maintain explicit adjacency relationships to their neighbors, preserving the document's original narrative structure alongside any semantic cross-references.

Edges define the typed connections between chunks. Common edge types include:

• Sequential: Links to the next and previous chunk in the source document
• Referential: Links between chunks that cite the same entity, definition, or figure
• Hierarchical: Parent-child links reflecting heading structure (H2 contains H3)
• Semantic: Links derived from high cosine similarity in embedding space These explicit relationships allow traversal algorithms to follow precise paths rather than relying solely on vector proximity.

The defining capability of a chunk graph is multi-hop retrieval. When a query requires synthesizing information spread across disconnected sections—such as 'How does the pricing model affect the SLA terms?'—the graph enables traversal from a chunk about pricing, across a referential edge to a definitions chunk, and onward to the SLA section. This chain-of-chunks traversal mirrors how a human analyst navigates a document, dramatically improving recall for complex, compound questions.

Building a chunk graph involves a multi-stage pipeline:

1. Chunking: Segment documents using semantic or structural strategies
2. Entity Extraction: Identify named entities, key terms, and citations within each chunk
3. Relationship Extraction: Detect co-references, shared entities, and cross-references between chunks
4. Edge Weighting: Assign confidence scores to edges based on relationship strength
5. Indexing: Store both the vector embeddings and the graph adjacency list for hybrid retrieval

Chunk graphs serve as the retrieval backbone for GraphRAG (Graph Retrieval-Augmented Generation) architectures. Unlike standard RAG that retrieves the top-k most similar chunks, GraphRAG uses the chunk graph to perform community detection and subgraph extraction around a query's seed nodes. This produces a context package that includes not just the most relevant chunks but also their structural neighbors, definitions, and related concepts—providing the LLM with a comprehensive, coherent context window.

The foundational process for building a Chunk Graph. This strategy extracts entities and relationships from text to create graph-native chunks, rather than flat text segments. Each chunk becomes a node with typed edges (e.g., works_for, located_in) connecting it to other chunks, enabling retrieval via structured semantic queries like Cypher or SPARQL alongside vector search.

The primary query pattern enabled by the Chunk Graph. Unlike single-shot retrieval, multi-hop retrieval traverses the graph across multiple edges to answer complex questions. For example, answering 'What is the funding round of the company that acquired the startup that developed the chunking algorithm?' requires traversing developed_by → acquired_by → funding_round relationships between distinct chunk nodes.

The ideal node unit within a Chunk Graph. An atomic chunk represents a single, indivisible fact that cannot be further decomposed without semantic loss. By decomposing documents into atomic propositions (e.g., 'Acme Corp was founded in 2011') and linking them via shared entities, the graph achieves maximum precision and avoids the chunk contamination problem where multiple facts blur together in a single vector embedding.

A search paradigm that combines dense vector similarity with graph traversal to power Chunk Graph queries. Initial retrieval often uses vector search to find entry-point chunks, followed by graph expansion to pull in connected context. This is frequently augmented with sparse keyword matching (BM25) for exact entity name matching, ensuring both high semantic recall and precise, deterministic graph-based context enrichment.

The mechanism for tracing a generated response back to its source nodes in the Chunk Graph. Because the graph explicitly models provenance via relationships like cited_by or extracted_from, attribution becomes a deterministic path traversal rather than a probabilistic guess. This provides verifiable citations and allows for fine-grained auditing of which specific facts grounded the LLM's output.

A complementary strategy where the Chunk Graph serves as the precise index, but retrieval returns the larger parent document for final context. Small, atomic graph nodes enable pinpoint retrieval accuracy, while the linked parent document provides the full narrative and formatting context the LLM needs to synthesize a coherent, well-structured answer without hallucinating surrounding details.