Skip to content

Perspectival Inquiry — STORM made native

Concepts: AU-KG.research.perspectival-inquiry (engine), AU-KG.research.contradiction-agreement-blind-spot (contradiction/agreement/blind-spot structures), AU-KG.research.peer-review-self-critique (peer-review self-critique). Code: knowledge_graph/research/perspective.py, wired into knowledge_graph/research/search.py + research/loop_controller.py, surfaced via research/ara/service.py (action=inquire).

Stanford's STORM (NAACL 2024) showed that researching a topic from several distinct expert lenses — each asking different questions — then mapping where they disagree, produces markedly more organized and broader coverage than a single prompt. Its one known weakness is the lack of self-critique.

We make that pattern the default behaviour of the research fan-out, not a separate tool. Where the loop used to take one semantic probe of the topic name (acquire_for_topic), it now fans the same probe across questions asked from multiple perspectives (acquire_for_topic_perspectival), derives a contradiction/agreement/ blind-spot map, and runs a peer-review whose frontier question is submitted back as the next research loop — closing the loop STORM left open. The whole engine is deterministic and KG-grounded (an llm_fn is optional, only enriching question phrasing), so it runs on the cheap zero-infra cycle.

Flow

flowchart TD
    T["Research topic in a Loop"] --> D["derive perspectives (ontology-flavoured by KG neighbours)"]
    D --> Q["each lens asks distinct questions"]
    Q --> P["acquire_for_topic per question (reuses the single-lens KG probe)"]
    P --> CM["Contradiction map"]
    CM --> AG["Agreements: 2+ lenses, likely true"]
    CM --> DV["Divergences: lenses with disjoint evidence"]
    CM --> BS["Blind spot: KG neighbours no lens covered"]
    AG --> PR["Peer review: confidence, bias, missing lens"]
    DV --> PR
    BS --> PR
    PR --> M["materialize typed KG nodes (Perspective/Agreement/Contradiction/BlindSpot/PeerReview)"]
    PR --> F["frontier question, submit_loop"]
    F -->|next cycle| T
    P --> U["union of sources, mark_addressed, topic converges"]

Phases

  1. Perspectives — PerspectiveEngine.derive_perspectives returns distinct lenses (practitioner / academic / skeptic / economist / historian), with their rationale annotated by the topic's KG neighbour types (ontology-flavoured grounding).
  2. Fan-out — each lens's questions are answered by acquire_for_topic (the existing single-lens probe, reused per question), giving each lens a source set.
  3. Contradiction map — sources ≥2 lenses share are agreements ("likely true"); lenses with disjoint sets are divergences; the topic's KG-neighbour types no source covers are the blind spot.
  4. Peer review — per-source confidence from corroboration (1–10), the dominant lens (bias check), the missing lens, and a frontier question (about the blind spot or missing lens) submitted as the next research loop.

The inquiry materializes as typed KG nodes (research_inquiry, perspective, agreement, contradiction, blind_spot, peer_review with asks_from / agrees_with / reviews edges; ontology classes ⊑ :Concept), so it is graph-queryable next to the topic it addresses.

Surfaces (two by default)

  • Native (default-on): LoopController research cycle + _advance_research call acquire_for_topic_perspectival — every research run is multi-perspective, no flag.
  • MCP: research_artifact tool, action=inquire (topic=…).
  • REST: POST /api/research/inquire (topic, materialize).
  • On-demand skill: the multi_perspective_inquiry workflow-skill (delegates to the engine via research_artifact action=inquire — never re-implements the prompts).

Why deterministic

The single-lens path is subsumed, not kept beside it (no legacy). The fan-out reuses the existing bounded embed + semantic search per question, so an unreachable embedding endpoint degrades in seconds and the path falls back to the direct single-lens probe — behaviour never regresses. Lens count and questions are bounded, keeping the added cost a small multiple of the prior single probe.