For the Markdown index, I propose testing BM25 plus multilingual embeddings, rank fusion, and reranking. Query expansion and ColBERT are optional experiments, not automatic upgrades; judge them on the actual corpus, including cross-language queries and latency.

Proposed retrieval baseline and comparison

Build this on the size-aware Markdown index. The following is a test plan, not a reported benchmark result.

Use a lexical route for names, quoted phrases, paths, and identifiers, alongside an embedding route for paraphrases and conceptual queries. Merge candidate lists, then test whether a reranker improves the final evidence set. Cloudflare's hybrid-search documentation describes rank fusion followed by an optional cross-encoder stage; that documents an available mechanism, not a guaranteed gain on this corpus.

For a first experiment, retrieve roughly 30-50 candidates and return a smaller, deduplicated set within a fixed context budget. These are tunable starting values. Preserve an inexpensive lexical-only path for exact lookups rather than requiring query generation and reranking for every request.

For locally operated models, include the Qwen3 embedding and reranking families among the candidates. Their documented multilingual coverage includes Russian and English. That is a reason to test Russian-query/English-document retrieval where relevant, not proof of its quality on these repositories.

Qdrant documents ColBERT late-interaction reranking as another option. Compare it with the selected reranker rather than automatically stacking both. Similarly, query rewrites or hypothetical-document expansion should earn their latency and complexity through recovered evidence.

Use a labeled set of real searches covering exact identifiers, paraphrases, different languages, evidence spread across repositories, long-file sections, version conflicts, and questions with no supporting material. Compare lexical-only, hybrid, and hybrid-plus-reranker results. Measure whether the needed passages appear in the result budget, citation fidelity, cold and warm latency, and update/deletion lag. No candidate has been benchmarked on the Operator's repositories in this exchange.