Module: Pikuri::VectorDb::Reranker

Defined in:

lib/pikuri/vector_db/reranker.rb,
lib/pikuri/vector_db/reranker/hit.rb,
lib/pikuri/vector_db/reranker/llama_server.rb

Overview

Namespace for rerankers — the optional quality knob that narrows a top-N candidate set (~50) down to the top-k the LLM sees (~5) by scoring each candidate against the query with a cross-encoder. One ships: LlamaServer (+/v1/rerank+ in Cohere's wire format, so a Reranker::Cohere later is a thin adapter).

Reranker protocol

Duck-typed: #rerank(query:, documents:) → Array<{Hit}> sorted descending by score, one entry per input document, Hit.index the position in documents; empty documents → [] with no HTTP call.

Index, not document text: Tools::Search already holds the candidate texts, so returning bare indices avoids re-shuttling them over HTTP (Cohere's return_documents exists for callers that don't have the text, which isn't us).

nil-skip, not a NOOP default: Extension takes reranker: nil as "skip rerank, retrieve top-k directly." A Reranker::NOOP was rejected — it wouldn't unify the candidate-window choice (still branch on real-vs-fake) and its faux scores would pollute observations and logs.

Defined Under Namespace

Classes: Hit, LlamaServer