Search · Live · 2026 to now
FMC Research Assistant
Search over 4,033 documents of 19th-century French music press, by keywords or by semantic meaning, running entirely on local GPUs.
- Semantic search
- PostgreSQL
- Embeddings
- Reranking
- Local GPUs
What they had
The FMC corpus (France: Musiques, Cultures, 1789–1918, University of Southampton) holds 4,033 transcribed press documents across 16,574 pages. After my talk in Venice, the musicologists behind it asked whether AI could make it searchable.
What I built
A search tool with two ways in:
| By keywords | By semantic meaning | |
|---|---|---|
| Finds | every page containing the words | the 20 passages closest in meaning to a question |
| Query | words, phrases, OR, exclusions, prefixes, “within N words” | a question or description, in French or English |
| Engine | PostgreSQL full-text search | bge-m3 embeddings and French full-text search, fused and reranked |
| Typical time | 10 ms to 1 s | 1.5 to 3 s |
Every result is a page of a source: a short verbatim excerpt, a citation, and a link to the PDF at that page. Results export to Excel, and any search can be shared as a link. It all runs on a local machine with two GPUs; no cloud AI is involved.
A decision worth noting
It started as a RAG system that also wrote grounded essays. In September 2026 the scholars decided they wanted the sources in front of them, not a generated answer, so I retired generation. The tool finds and quotes; it doesn’t write.
Status
Live and open to anyone. A paper on it is planned for a conference in March 2027.