Summary
When feasibility analysis completes, generate 3-5 natural-language queries that would find this study useful and store them alongside findings for FTS5 indexing.
Motivation
The vocabulary gap between a deep research query and stored feasibility findings causes missed matches even with FTS5 stemming. Sift (Edmonds, 2026) demonstrates that LLM-generated anticipated queries at write time bridge this gap by capturing the searcher's perspective rather than the author's perspective, at zero marginal cost since the LLM is already in-context during feasibility synthesis.
Proposed Change
- After feasibility analysis saves results, call the reasoning model to generate 3-5 anticipated queries per finding
- Store in a new
anticipated_queries column on the feasibility results table
- FTS5-index the anticipated queries alongside
key_finding, conclusion, and tab text
- Weight anticipated queries in BM25 ranking
Acceptance Criteria
- Feasibility results include
anticipated_queries populated by LLM
- FTS5 index covers anticipated queries
- Deep research queries that share no surface tokens with findings but match anticipated queries return relevant scouting sources
Dependencies
- Requires SEP-065 (FTS5 index) to be implemented first
Summary
When feasibility analysis completes, generate 3-5 natural-language queries that would find this study useful and store them alongside findings for FTS5 indexing.
Motivation
The vocabulary gap between a deep research query and stored feasibility findings causes missed matches even with FTS5 stemming. Sift (Edmonds, 2026) demonstrates that LLM-generated anticipated queries at write time bridge this gap by capturing the searcher's perspective rather than the author's perspective, at zero marginal cost since the LLM is already in-context during feasibility synthesis.
Proposed Change
anticipated_queriescolumn on the feasibility results tablekey_finding,conclusion, andtabtextAcceptance Criteria
anticipated_queriespopulated by LLMDependencies