qdrant-relevance-feedback

Python package that fits the three parameters (a, b, c) of Qdrant’s index-native Relevance Feedback query — the scoring formula that guides HNSW traversal toward what a feedback model judged relevant.

You do not train per query. Fitting happens once per (feedback model, collection, retriever) and the resulting weights are then passed with every relevance-feedback query. The package exists so that using the feature doesn’t require an ML background — “nobody needs a machine learning degree to use a search engine.”

What it does

• Scores a sample of domain queries with your chosen feedback model — a Cross-Encoder, a bi-encoder, a late-interaction model like ColBERT, a custom Learning to Rank model, or an LLM
• Mines context pairs from the top results
• Fits a, b, c by pairwise ranking loss against that signal — effectively Knowledge Distillation of the feedback model into the vector index
• Emits weights for the RelevanceFeedbackQuery API

Reference experiments use 50–6,000 queries depending on collection size, with a 50/50 train/validation split.

Related

• Shipped alongside the relevance feedback query in Qdrant 1.17.0 (February 2026)
• Qdrant Vector DB — the engine holding the index
• Article: Relevance Feedback in Qdrant — the formula it fits, and the BEIR evaluation
• Tutorial: Relevance Feedback Retrieval in Qdrant
• Talk: Evgeniya Sukhodolskaya - Relevance Feedback Inside the Search Engine — Berlin Buzzwords 2026
• Compare: qdrant-sparse-finetune — the same “make ML approachable” framing on the sparse side