Jump to a problem, path or page
Jump to a problem, path or page
01 · Develop the design
Fan out through authenticated source adapters under explicit deadlines, then merge bounded candidate lists using a deterministic rank-fusion policy. Persist a short-lived authorized session and disclose incomplete or truncated source coverage.
Stage 1
Normalize contracts, not arbitrary scores
Adapters return stable document IDs, source rank and authorization context. Fuse source ranks under a pinned policy rather than assuming score comparability.
Stage 2
Budget every branch
Cap providers, candidates per source, total fan-out and wall-clock wait. Record failed, timed-out and truncated sources separately.
Stage 3
Bound pagination claims
Stored candidates provide stable pagination only within that bounded session. They do not prove no omitted provider document could rank higher.
02 · Work through the method
03 · Justify the choices
Can raw relevance scores from unrelated providers be sorted together?
Only after an explicitly justified calibration; otherwise use a declared rank-fusion rule
Native score scales may represent different models and are not automatically comparable.
How should a timed-out source appear?
As an explicit incomplete source in the response coverage
A missing response must not be presented as an exhaustive negative search result.
What makes later result pages stable in the baseline?
A bounded stored fused-result session plus reauthorization on fetch
A session pins the candidate list, while current authorization still controls what can be disclosed.
04 · Plan the evidence