Retrieve
Permitted PDFs are indexed with page-aware text extraction, local embeddings, and a searchable collection.
Case study · verified local demonstration
AI document intelligence,
with evidence attached.
A local research workspace designed to help people ask useful questions of permitted document collections without losing the source trail.
The problem
Document-heavy work becomes difficult when answers need to be traceable, current within a defined collection, and honest about uncertainty. This demonstration explores a source-first workflow for research teams and operations groups.
This is a local, single-user research-engine demonstration. It is not a deployed commercial SaaS product.
Check
Answers include source title, page, section, and publisher link. Citation markers are checked against retrieved evidence.
Review
Candidate observations can be reviewed and revised. Memory remains advisory; fresh citations are still required.
Architecture
Local inference.
Visible boundaries.
Next.js and TypeScript provide the workspace interface. FastAPI and SQLite support the local API and persistence. PyMuPDF, MiniLM embeddings, FAISS, and an Ollama-compatible Qwen-class model support the research loop.
When available evidence is insufficient, the system abstains rather than presenting an unsupported answer.
Demonstration
Evidence
Source: public research-engine repository, including README architecture, safeguards, verification reports, screenshots, event timelines, and demonstration videos.
What this means for a business
Useful when the
source trail matters.
The pattern is relevant to policy lookup, internal research, and document-heavy decisions where a helpful answer must remain reviewable.
Discuss a workflow ↗