multi_vector_query
Combine results from multiple vector queries (e.g., dense + sparse) using weighted or RRF re-ranking to find relevant documents across different embedding spaces.
Instructions
Perform multi-vector similarity search with score fusion and re-ranking.
This tool searches across multiple vector embeddings simultaneously and combines their results using a re-ranking strategy. This is useful when documents have multiple types of embeddings (e.g., dense + sparse, text + image).
Args: params (MultiVectorQueryInput): Validated input parameters containing: - collection_name (str): Collection identifier - vectors (List[MultiVectorQuerySpec]): List of vector queries (min 2) - topk (int): Candidates to retrieve from each vector field (default: 10) - topn (int): Final documents to return after re-ranking (default: 5) - reranker_type (str): 'weighted' or 'rrf' (default: weighted) - weights (Optional[Dict[str, float]]): Field weights for weighted re-ranker - rank_constant (int): RRF rank constant (default: 60) - metric_type (str): Metric for weighted re-ranker (default: IP) - filter (Optional[str]): Filter expression - response_format (str): Output format
Returns: str: Re-ranked search results or error message
Examples: - Use when: "Search using both dense and sparse embeddings" - Use when: "Combine text and image similarity for multi-modal search"
Re-ranking Strategies: - Weighted: Combines normalized scores with custom weights per field Best when scores are comparable and you know field importance - RRF (Reciprocal Rank Fusion): Combines based on rank positions only Best when scores use different metrics/scales or prefer tuning-free approach
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| params | Yes |
Output Schema
| Name | Required | Description | Default |
|---|---|---|---|
| result | Yes |