Relevance Feedback Inside the Search Engine
summary
A new API injects model-provided relevance scores directly into the vector index so the HNSW traversal picks next hops by a weighted mix of cosine similarity and feedback, improving recall and discoverability without reranking the whole corpus. Implementation lives in Qdrant, is open source, and evaluates on BEIR showing measurable gains. The formula uses three learned parameters, trains with ~150 training examples, and implements HNSW traversal modification at roughly dense-search cost while paying separately for feedback scoring.
AI-written summary · about
chapters (5)
- 0:00Introduction and motivation
- 3:57What is relevance feedback
- 7:29Vector index native feedback
- 13:29Summary and future
- 14:31Q&A session
related talks
44mRediscover your keyword search: Expand, Enrich and Rewrite
37mEnhancing AI and LLM Systems with Hybrid Search
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
47mBoosting AI With OpenSearch: Vector Stores, Similarity Search, and Retrieval-Augme...
45mMaking semantic search & RAG real: How to build a prod-ready app
13mThe State of AI Powered Search and Retrieval
49mHybrid search on hybrid models, at scale
33mThe Invasion of Transformers - Boosting Search with Latest NLP