Most enterprise AI teams treat retrieval-augmented generation as a search feature. Build an index, embed your documents, retrieve the top-k results, append to the prompt. Ship it.
This works well enough to pass a demo. It does not work well enough to run financial analysis, answer regulatory questions, or drive autonomous agents that need accurate context every time.
The failure mode is silent and expensive. Vector-only retrieval achieves recall@10 between 65% and 78% on enterprise knowledge bases. That means in 22–35% of queries, the relevant document is not in the retrieved set. The model answers confidently from incomplete information.
The hybrid approach
Hybrid RAG combines two retrieval strategies:
Dense retrieval (vector search): Documents are embedded into a high-dimensional vector space. Query time: embed the question, find nearest neighbors by cosine similarity. Excellent at semantic understanding - finds relevant content even when exact terms differ.
Sparse retrieval (BM25): Classical keyword-based ranking. Excellent at precise term matching - finds exact product names, regulation numbers, account IDs, error codes. Does not suffer from embedding drift.
Neither approach alone is sufficient. Together, they cover each other's failure modes.
Reciprocal Rank Fusion
The combination method matters. Naive approaches - score averaging, threshold filtering - do not work well in practice. Reciprocal Rank Fusion (RRF) is the correct algorithm.
RRF assigns each document a score based on its rank in each retrieval list, not its raw score. The formula is simple:
where k is a constant (typically 60) and rank_i is the document's rank in retrieval system i.
The result: documents that appear in the top results of both retrieval systems get a strong combined score. Documents that excel in one system but are absent in the other are appropriately discounted.
Benchmark results from 2026 RAG evaluations: recall@10 jumps from 65–78% (vector-only) to 91%+ (hybrid with RRF).
Why this matters for agents
Autonomous agents that drive business processes - financial close, procurement analysis, customer retention - cannot tolerate 22–35% retrieval failure rates. A missed document in a financial reconciliation means a wrong answer. A missed contract term means a compliance gap.
At 91%+ recall, retrieval becomes a reliable substrate for reasoning, not a best-effort lookup.
NXπ implements hybrid RAG with RRF as the default retrieval strategy. Every RAG pipeline benefits from it automatically. Sources are attributed per chunk. Every answer is grounded.