Search methodology
Build a search strategy you can inspect and test
A valid Boolean query can still miss relevant studies. Prepare the concepts first, check what the database actually executes, and test the strategy against articles you expect to find.
Prepare your search strategyUpdated October 7, 2026 · An account is required to use the assistant.
From question to search
- Write the research scope clearly. For clinical questions PICO can help; other disciplines may need different concept groups.
- Identify essential concepts and collect synonyms from relevant literature. Combine alternatives within a concept with OR.
- Choose which concepts must appear together using AND. Avoid requiring every comparison, outcome or study-design detail just because it is part of the review question.
- Inspect syntax and source-specific interpretation. A database may change phrases, wildcards, field tags or controlled vocabulary.
- Test known relevant articles and compare the original strategy with any proposed revision. Keep the original and document accepted changes.
What the research contributes
AutoBool: Reinforcement-Learned LLM for Effective Automatic Systematic Reviews Boolean Query Generation
Wang et al. · 2026 · EACL
Evaluate executed queries against relevant studies; measure coverage alongside screening workload.
Chained Prompting for Better Systematic Review Search Strategies
Nasser et al. · 2025 · FLLM; arXiv manuscript
Separate question interpretation, concept identification, terminology expansion and query construction into reviewable stages.
Reassessing Large Language Model Boolean Query Generation for Systematic Reviews
Wang et al. · 2025 · SIGIR
Validate generated queries and record the model, prompt and examples used; performance depends on these choices.
These studies primarily concern medical systematic reviews and PubMed. Their findings are useful design evidence; they do not establish performance across every discipline or database.
Read performance claims carefully
AutoBool optimizes a model using retrieval feedback. Lumina currently applies the lesson of testing queries; it does not run that trained model. The paper reports substantial improvements, but high coverage is not achieved for every topic.
The chained-prompting study evaluates 81 selected reviews and excludes searches returning more than 1,000 citations. Its full pipeline reports average recall of 0.87; the 0.90 figure applies after excluding low-recall results. Those results cannot be treated as a guarantee for large or unfamiliar reviews.
The reassessment study highlights sensitivity to model choice, prompt design, validation and seed studies. A successful example is insufficient to establish a dependable tool.
What the Lumina assistant checks
You can review AI suggestions or enter concepts manually. Lumina builds the Boolean expression from the concepts you include and checks its structure. AI suggestions remain unverified; no MeSH descriptor is automatically labeled as validated.
The optional PubMed reference check first verifies which supplied PMIDs are indexed, then checks which match the proposed and original strategies. It uses the same default publication-type exclusions as Lumina’s PubMed search and no date restriction. It intersects identifiers directly rather than relying on the first page of search results.
Finding all supplied references shows coverage of those examples only. It does not demonstrate complete recall. Articles used to suggest terms should be distinguished from independent articles used for testing. API failures are reported as unavailable checks, rather than missing studies.
A search returning fewer results is not automatically better. Evaluate coverage and the amount of screening work together. This implementation has not yet been independently benchmarked against the studies above.
Keep a reproducible record
Download the reviewed strategy and check results as JSON. This record includes your question, original and proposed queries, PubMed expression, date, filters, reference-check results and any AI generation settings and input text. The assistant saves your current draft to your account so you can return on another device. This export documents search preparation and reference checks, rather than an executed database search. Keep the dates, filters and result counts of your executed searches with your review.
Start with your research question →