Function crossEncoderRerank

src/context/reranker.ts:58–85

(
  query: string,
  candidates: RerankCandidate[],
  opts: RerankOptions = {},
)

 * (caller should then fall back to the bi-encoder order).
 */
export async function crossEncoderRerank(
  query: string,
  candidates: RerankCandidate[],
  opts: RerankOptions = {},
): Promise<RerankResult[] | null> {
  if (candidates.length === 0) return [];
  const baseUrl = opts.baseUrl ?? process.env.QODEX_RERANK_URL ?? process.env.OLLAMA_BASE_URL ?? 'http://localhost:11434';
  const model = opts.model ?? process.env.QODEX_RERANK_MODEL ?? 'bge-reranker-v2-m3';
  const maxChars = opts.maxCharsPerDoc ?? 2000;

  const documents = candidates.map(c =>
    c.text.length > maxChars ? c.text.slice(0, maxChars) : c.text,
  );

  // Try the two common local reranker API shapes in turn. Both take
  // {query, documents[]} and return per-document relevance scores.
  const scores = await tryOllamaRerank(baseUrl, model, query, documents, opts.signal)
    ?? await tryOpenAIRerank(baseUrl, model, query, documents, opts.signal);

  if (!scores) return null; // nothing reachable — caller falls back

  const results: RerankResult[] = candidates.map((c, i) => ({
    id: c.id,
    score: scores[i] ?? 0,
  }));
  results.sort((a, b) => b.score - a.score);
  return typeof opts.topN === 'number' ? results.slice(0, opts.topN) : results;
}

/** Ollama-style reranker: POST /api/rerank {model, query, documents} → {results:[{index,relevance_score}]}. */
async function tryOllamaRerank(