I've been exploring the vectorless RAG concept through an experimental project, focused on understanding how graph-based retrieval can complement embedding-driven approaches.
For this experiment, I used the AWS Overview whitepaper as the source corpus and progressively evolved the extraction and retrieval strategy:
Phase 1 — Regex-based Entity Extraction
Started with a simple regex extractor to identify entities and relationships from the document and construct an initial knowledge graph.
Phase 2 — NLP-based Extraction (spaCy)
Introduced spaCy to improve entity recognition using traditional NLP techniques and enrich the graph structure.
Phase 3 — LLM-based Entity Extraction
Added an LLM-based extractor to capture richer, more contextual entities and relationships that are difficult to model with rules or classical NLP alone.
Retrieval & Generation Flow
- User query triggers graph traversal and subgraph retrieval
- The retrieved context + subgraph is passed to the LLM
- The LLM generates answers grounded in explicit graph relationships, not vector similarity
Retrieval architecture matters as much as the LLM itself.
- To deeply understand vectorless RAG trade-offs
- To explore explainable, deterministic retrieval
- To reduce dependency on embeddings and vector stores
This is an experimental learning project, but it reinforced an important insight: retrieval architecture matters as much as the LLM itself.