Vector Search Architecture Consultant

We’re looking for a Retrieval & Ranking consultant to redesign the first stage of our AI matchmaking system. Profiles are rich but inconsistently structured - varying questionnaires, short yes/no answers mixed with long free text, plus notes that range from useful preference insight to operational noise. Today we compress each profile into a small set of LLM-written summaries, embed those, and use vector search to shortlist candidates before an LLM reranks them. We want a more accurate retrieval architecture for this kind of data: what should be filtered vs embedded, how to handle messy non-standard Q&A, how many vectors/chunks to use, and how to keep rankings stable when profiles are updated. We need a clear recommendation for the best production approach - hybrid search, multi-vector design, scoring, migration, cost, and how to measure whether match quality actually improves - without overcomplicating the system. Ideal background: information retrieval, embeddings, hybrid search, and ranking/evaluation in production, ideally for matching or recommendation products.