Enterprise Generative Search — Azure and Open Source  ·  View 07 of 41  ·  People and journeys

Journey — Multi-Hop Under Review

Where an averaged answer is worse than no answer at all.

Editable source SVG draw.io All views
Research Analyst regulated deliverable, named reviewer Goal — Assemble a defensible position from sources that do not agree with each other Trigger — A regulatory change, or a question from counsel Done when — A reviewer accepts the evidence pack without re-doing the search 1 · Frame 2 · Decompose 3 · Gather 4 · Reconcile ◆ moment of truth 5 · Cite ◆ moment of truth 6 · Deliver What they do States the real question Accepts or edits the plan Scans evidence, adds a source Weighs conflicting passages Checks every citation Exports the pack Where Research workspace Visible query plan Evidence tray Comparison table Citation panel Signed export Behind the glass Intent and scope model Planner, 3 sub-queries Graph expansion Per-hop retrieval Diversity selection Contradiction detector Authority and recency Claim to evidence binding Provenance record persisted How it feels In control Working Losing trust Where it hurts Plan is invisible, cannot steer Six near-identical passages Draft averaged two positions False consensus reads as fact Citation points at the document What answers it Editable plan before it runs Redundancy collapsed by MMR Disagreement is a stated state Both positions kept, dated Cite anchors to the passage Pack replays from evidence ids Journey — Answering a Multi-Hop Question That Has to Survive Review An averaged answer is worse than no answer here. Reconciliation is a first-class output, which is why contradiction is detected rather than smoothed. v 1.0 · owner Data and AI Global Practice

The trough, and what it demands

  • The low point is Reconcile: two authoritative sources disagree and the first draft blends them into a consensus that never existed. The analyst cannot see that it happened, and neither can their reviewer.
  • This is why contradiction detection is a stage in view 26 rather than a prompt instruction, and why disagreement is a state an answer can be in rather than a tone it can take.
  • It also decides that both positions are kept, dated and attributed, which requires the effective dates in the data model in view 19.

What the platform owes each phase

  • Decompose: the plan is visible and editable before it runs. An analyst who cannot steer the search will not trust its output.
  • Gather: diversity-aware selection, because six near-identical passages from one document look like corroboration and are not.
  • Deliver: an evidence pack that replays from evidence ids, so a reviewer checks the work rather than redoing it.

Numbers and risks

  • 9,000 multi-hop queries a month, average 2.4 hops, 18 s P95 end to end, 45k token ceiling.
  • Risk: the agentic path is 45 times the cost of a lookup. It is gated by classification and by a hard budget rather than by asking users to be sparing.
◀  Journey — Ask and Act All views Journey — Why Is My Document Not Found  ▶
Solution Architecture v1.0 · Data & AI Global Practice · 2026-09 · 41 views · Microsoft Azure and open source07-journey-research-analyst