Hybrid Retrieval (RRF)
Standard vector similarity searches fail when exact keywords, temporal recency, or complex multi-hop entity relationships are required. Memron fuses four orthogonal signals using Reciprocal Rank Fusion (RRF).
The 4 Retrieval Signals
The hybrid retrieval orchestrator executes four searches concurrently:
- Vector Cosine Similarity (Weight: 1.0): HNSW vector search on 1536-dimensional embeddings generated by OpenAI text-embedding-3-small.
- BM25 Full-Text (Weight: 0.8): PostgreSQL tsvector and ts_rank_cd for exact keyword matches, symbol names, and acronyms.
- Knowledge Graph Traversal (Weight: 1.2): Blind-hash anchor discovery followed by recursive CTE N-hop expansion.
- Ebbinghaus Memory Decay (Weight: 0.6): Exponential time decay with a configurable 7-day half-life and frequency reinforcement.
RRF Formula
Score(d) = Σ [ w_i / (k + rank_i(d)) ], where k = 60. By focusing on rank positions rather than raw incompatible score distributions, RRF ensures balanced fusion across all signals.