Open HUB-024: adaptive query planning (prior-art-grounded methodology) - #10
Merged
Conversation
…ology The revisit trigger tripped: a job issues exactly one SearXNG query, so a report is built from whatever that phrasing surfaced. The narrowness is on the acquisition side (hybrid retrieval already measures hit@4 1.0), and it caps HUB-034 cross-source pair drafting, which can only find disagreements between sources that were actually crawled. Design (PRDs/hub-024-query-planning.md), grounded in 16 arXiv abstracts fetched and read 2026-08-13: - Breadth is emergent, not a fixed count: candidate facets are admitted by marginal distinctness (embedding cosine below a threshold against the admitted set), so a narrow topic admits one facet and behaves exactly as today while a broad topic admits as many angles as it has. This lifts ScoreGate's threshold-not-top-K cardinality principle from chunk selection to query planning; Adaptive-RAG-style complexity routing falls out of single-facet collapse rather than needing a trained classifier. - Rounds are gap-driven (KiRAG): each round names what is still uncovered from a per-facet coverage summary, and only those gaps become queries. - Stopping is saturation-first (KAIR, instanced on canonical-URL identity), coverage-second, with hard budget caps as backstop only (AdaRankLLM argues learned necessity over fixed iteration budgets). - Traps avoided, each from a paper's own stated limitation: fixed depth x breadth (Static-DRA), expecting breadth to fix report quality (DeepWeb-Bench: retrieval is 12-14% of errors, derivation/calibration exceed 70%), redundant tool calls (HotelQuEST). Acceptance measures corpus breadth (distinct domains, represented sources) against the single-query baseline — not report quality, which stays the judge gate's job. Ships behind REPORT_QUERY_PLANNING=false; judge calls per report stay bounded by existing drafting caps. Implementation not started. Also refreshes the staged next-session prompt, which said P3 was fully closed. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Docs only — opens HUB-024 (revisit trigger tripped) with a design grounded in 16 arXiv abstracts fetched and read 2026-08-13.
Core choice: breadth is emergent, not a fixed count. Facets are admitted by marginal distinctness (embedding cosine threshold against the admitted set), so a narrow topic admits one facet and behaves exactly as today. Rounds are gap-driven (KiRAG); stopping is novelty-saturation first (KAIR), coverage second, hard budget caps as backstop only.
Traps avoided, each from a paper''s own stated limitation: fixed depth x breadth (Static-DRA), assuming breadth fixes report quality (DeepWeb-Bench), redundant tool calls (HotelQuEST).
Implementation not started; ships behind
REPORT_QUERY_PLANNING=falsewhen built.🤖 Generated with Claude Code