Summary

Add Reciprocal Rank Fusion (RRF) as a rank aggregation option in memory/ranking.py for merging ranked lists from multiple retrieval rounds or sources.

Motivation

Two independent sources (NVIDIA NeMo Retriever, Agentic RAG/TDS) recommend RRF over linear weighted combination for fusing results from different retrieval strategies. Qdrant's own analysis explicitly rejects linear fusion (H = (1-a)*K + a*V) because "relevant and non-relevant items are mixed" in 2D score space. RRF is the de facto standard for hybrid search fusion.

Current ranking uses a linear combination (0.7*relevance + 0.3*recency). This works for single-signal ranking but is insufficient when:

• BM25/sparse search is added alongside vector search
• Multiple retrieval rounds produce separate ranked lists (tool-based injection)

Design

~30-line algorithm: RRF_score(doc) = sum(1 / (k + rank_i)) across all ranking sources.

Add as an alternative to the current compute_combined_score in memory/ranking.py. Use RRF when multiple ranked lists need merging; keep linear combination for single-source relevance+recency scoring.

Affected Files

• src/ai_company/memory/ranking.py (add RRF function)
• src/ai_company/memory/retrieval_config.py (config for fusion strategy)

Research

• NeMo Retriever Agentic Retrieval (NVIDIA)
• Agentic RAG with Hybrid Search (TDS)
• Qdrant: Hybrid Search Revamped — rejects linear fusion