> Source: https://amayakavya.com/projects/quire

# Quire

Document Search
Open Source

Quire answers questions about a body of documents. You ask in plain language and it returns an answer grounded in the corpus, with the source document and the page number sitting beside it so the claim can be checked rather than taken on trust.

Nothing leaves the machine. The embeddings, the reranking and the generation all run locally, against a vector store and a local model on the same box. For a corpus that cannot be uploaded to somebody else's service, which is most corpora worth searching, that constraint decides the architecture rather than being a feature bolted on later.

A question goes through spelling correction, decomposition into atomic queries, a scope check, metadata resolution, dense retrieval, cross-encoder reranking, and finally generation constrained to the passages that were actually retrieved. An answer the corpus cannot support is refused rather than invented, which is the behaviour that makes the citations mean anything.

There are three consoles: one for searching, one for ingesting documents and seeing what is indexed, and a log terminal streaming backend events as they happen. Verified answers are kept and replayed as examples, so the system gets better at the corpus it is actually pointed at.

## Source

- [github.com/amayakavya/quire](https://github.com/amayakavya/quire)

[← Back to Projects](/projects)