Mnemosyne
liveAsk a question about the risks a big company listed in its annual report, and it answers in plain English while pointing to the exact sentence in the exact filing each fact came from.
View source on GitHubA retrieval-augmented generation (RAG) system: structure-aware chunking, local sentence-embeddings in a Chroma vector store, and grounded generation where citations are resolved in code, not written by the model. Shipped with an LLM-as-judge eval harness scoring retrieval hit rate, faithfulness, and answer relevance.
mnemosyne.demo/results
loading demo…
