Vector Space Model
The Vector Space Model (VSM) is a statistical approach to information retrieval and text analysis that represents texts and other objects as vectors of identifier occurrences, often weighted, such as indexed terms, suitably displayed in a vector space.
Core Principles
- Representation: Documents are represented as sparse vectors where each dimension corresponds to a term in the vocabulary.
- Similarity: Similarity between documents is typically measured using Cosine Similarity, which calculates the cosine of the angle between two vectors.
- Weighting: Term frequencies are often weighted using TF-IDF (Term Frequency-Inverse Document Frequency) to reduce the impact of common, less informative words.
Evolution and Relation to Embeddings
While traditional VSM relies on discrete term counts, modern approaches utilize dense embeddings to capture semantic meaning beyond lexical overlap. Recent advancements in Retrieval-Augmented Generation (RAG) have expanded beyond pure text processing to handle multimodal inputs.
- Multimodal Retrieval: Traditional text-based RAG often fails with visually complex documents (e.g., scientific papers, financial reports) where layout and visual structure are critical.
- PixelRAG: A novel approach introduced in PixelRAG: Screenshot-Based RAG for Complex Document Comprehension that utilizes screenshots rather than raw text extraction. This method preserves visual context and layout information, improving comprehension for documents where spatial arrangement carries semantic weight.