Hybrid Search

Hybrid Search is an enhanced retrieval mode for vector libraries that combines vector semantic retrieval and keyword matching retrieval, leveraging the strengths of both to improve RAG accuracy and recall.

What is Hybrid Search

In pure vector retrieval mode, user queries are vectorized and matched against semantically similar text chunks in the vector space. This works well for semantically similar but differently worded questions but may underperform for scenarios requiring literal matching, such as precise terminology, serial numbers, or code.

Hybrid Search simultaneously performs keyword retrieval alongside vector retrieval, fusing and ranking results from both methods to return the top-ranked text chunks. This achieves a balance between:

• Semantic Understanding: Capturing the deeper meaning of user queries through vector retrieval
• Exact Matching: Ensuring content containing specific terms is not missed through keyword retrieval

Enabling Hybrid Search

Hybrid Search is configured at the knowledge base level. When creating or editing a vector library, you can find the Hybrid Search toggle in the vectorization strategy settings step.

Enable the toggle and save. The knowledge base will then use hybrid search mode.

Notes

• Hybrid Search is only available for vector-type knowledge bases. File libraries do not support this feature.
• Enabling or disabling Hybrid Search triggers re-vectorization (similar to switching embedding models). Confirm before saving.
• Hybrid Search does not require additional model configuration. The platform automatically fuses results from both retrieval methods at query time.
• Hybrid Search may slightly increase retrieval latency, as two retrieval methods run simultaneously and results must be fused and ranked. This latency is typically within acceptable limits.