Definition
The application of hard constraints based on structured attributes (e.g., date, source, category) to narrow the search space within a vector database during a RAG retrieval step. This process combines Boolean filtering with similarity searching to ensure retrieved context is both semantically relevant and logically valid.
Distinct from semantic similarity; it handles 'hard' requirements like 'only documents from 2023' rather than 'concepts similar to 2023'.
"A colander that only lets through pasta of a specific shape before a chef selects the best-cooked pieces for a dish."
- Vector Search(Component)
- Pre-filtering(Implementation Strategy)
- Post-filtering(Implementation Strategy)
- Self-Querying Retriever(Advanced Extension)
Conceptual Overview
The application of hard constraints based on structured attributes (e.g., date, source, category) to narrow the search space within a vector database during a RAG retrieval step. This process combines Boolean filtering with similarity searching to ensure retrieved context is both semantically relevant and logically valid.
Disambiguation
Distinct from semantic similarity; it handles 'hard' requirements like 'only documents from 2023' rather than 'concepts similar to 2023'.
Visual Analog
A colander that only lets through pasta of a specific shape before a chef selects the best-cooked pieces for a dish.