This project adopts a hybrid search scheme combining BM25 keyword search and FAISS vector similarity search, which has significant advantages over a single search method: the BM25 algorithm is responsible for processing precise term matching and effectively capturing the core keywords in the user's query, while FAISS understands semantic relevance through dense vector search, and both of them work in tandem to achieve a recall rate of more than 92%. The two work together to achieve a recall rate of more than 92%. The neural reordering technique implemented by the cross-coder is also introduced in the retrieval process to optimize the relevance scoring of the initial retrieval results, ensuring that the accuracy of the first 5 results is increased by 40%.
The system innovatively integrates HyDE (Hypothetical Document Embedding) query expansion technology to reconstruct query vectors by generating hypothetical answers, effectively solving the problem of terminology mismatch, which has been shown to improve average accuracy by 35% for complex queries of professional documents. Measurements show that for complex queries of professional documents, this technology can improve the average precision by 35%. All retrieval components are GPU-accelerated and optimized for millisecond response on devices with 16GB of memory.
This answer comes from the articleDeepSeek-RAG-Chatbot: a locally running DeepSeek RAG chatbotThe
