—

—

—

Chunking

600 words/passage, 75-word overlap, heading-aware

Embedding model

nomic-embed-text (768-dim) via LiteLLM on Colin

Vector DB

bnl_client_chunks in Qdrant (Colin Docker)

Search method

Hybrid (vector + keyword), reciprocal rank fusion, private boost 1.5×

Graph DB

dbn_client_graph in FalkorDB (Colin) — citation edges

.