Grounded retrieval
Compare lexical and dense retrieval, combine rankings, and apply an answerability gate.
Research / App case study
LIGTAS-AI
An offline English-Filipino-Taglish disaster-preparedness research system that retrieves grounded guidance, shows traceable citations, and abstains when its local corpus cannot support an answer.
Preparedness research prototype · fixture corpus
- Python
- FastAPI
- SQLite FTS5
- BM25
01 / Overview
What it’s built for.
Provide bounded educational preparedness answers with traceable citations and refuse questions the active corpus cannot support.
Who it serves
English, Filipino, and Taglish readers and researchers evaluating grounded preparedness retrieval.
02 / Features
Inside the app.
Evidence and abstention
Return cited answers, abstention, or safety redirects with active corpus and backend context.
Readiness and research tools
Include a browser-local checklist, printable plan, read-only command center, and evaluation-status views.
03 / Workflow
From task to outcome.
Open the local system and inspect its corpus status.
Ask a preparedness question in a supported language.
Review cited evidence or the explanation for abstention.
Use readiness tools or inspect the research evaluation status.
04 / Technical layers
How the pieces work.
Jinja · Vanilla JavaScript
FastAPI serves the local interface without a Node frontend or user account.
BM25 · Dense retrieval · Rank fusion
A bounded retrieval pipeline applies calibrated gates and exposes the active backend.
SQLite FTS5
A local indexed corpus retains evidence identifiers, source metadata, and citation context.
Python · FastAPI · Optional ONNX
Offline local operation is distinct from the fixture-based hosted demonstration and its temporary index.
05 / Limitations
What to keep in mind.
- The bundled corpus is researcher-authored paraphrase fixture data, not exact agency wording or the final thesis corpus.
- Human bilingual validation and final locked evaluation remain documented research gates.
- The system is not a live warning service, evacuation router, dispatcher, predictor, or medical adviser.
06 / Evidence
Follow the source.
These descriptions were reviewed against the linked source snapshot on October 12, 2026. Repository code and documentation may evolve after this review.