Contextual-Retrieval Enrichment (CONCEPT:AU-KG.enrichment.contextual-retrieval-enrichment)¶
Overview¶
Before a chunk is embedded, a short context string situating it within the whole document is
computed and prepended to the embedding input (Anthropic "Contextual Retrieval"), markedly
improving recall. LLM path uses the lite chat model; a deterministic heuristic (title + nearest
heading + part i/n + opening sentence) keeps it working offline. Default OFF in the KG-2.48
DocumentProcessor; the connector ingestion path turns it ON.
Implementation Details¶
- Source Code:
agent_utilities/knowledge_graph/ontology/contextual_enrichment.py; wired inontology/document_processing.py(DocumentProcessor(contextual=True)) - Pillar: KG