All Products
Search

This Product

    DashVector:Keyword-aware search

    Document Center

    DashVector:Keyword-aware search

    Last Updated:Apr 01, 2026

    Keyword-aware semantic search combines keyword matching and semantic understanding in a single DashVector query. By storing both sparse vectors (for term frequency) and dense vectors (for semantic meaning) in one collection, you get stronger recall without running two separate search systems.

    Key concepts

    Sparse vectors represent keywords and term frequency. They contain mostly zero values with a few non-zero values, one for each keyword in the document. In DashVector, the sparse vector {1:0.4, 10000:0.6, 222222:0.8} has three keywords (elements 1, 10000, and 222222), where the values are their weights.

    Dense vectors represent the semantic meaning of content. An embedding model converts text into a fixed-length array of floating-point numbers, capturing conceptual relationships that keyword matching misses.

    Sparse vectorsDense vectors
    RepresentsKeywords and term frequencySemantic meaning
    StructureMostly zeros, few non-zero valuesMostly non-zero values
    Built withInverted indexes, TF-IDF, BM25Embedding models
    StrengthExact term matchingConceptual similarity
    Example gapCannot match "raining heavily" to "rain cats and dogs"Cannot match "rice irrigating" to "irrigating rice" if untrained

    Keyword-aware semantic search closes both gaps at once. The final score for each result is the dot product of the query vectors and the document vectors — dense and sparse scores are summed together.

    When to use keyword-aware semantic search

    Use keyword-aware semantic search when:

    • Queries mix natural language with specific terms (product names, error codes, IDs)

    • Users search for idioms, synonyms, or paraphrased concepts

    • Your corpus requires both precise keyword recall and contextual understanding

    Prerequisites

    Before you begin, ensure that you have:

    Run keyword-aware semantic search

    The following steps create a collection, insert a document with both vector types, and run a hybrid query.

    Important

    Only collections that use metric='dotproduct' support sparse vectors.

    Step 1: Create a collection

    import dashvector

# Replace YOUR_API_KEY with your API key (see Manage API keys).
# Replace YOUR_CLUSTER_ENDPOINT with your cluster endpoint (see the cluster details page).
client = dashvector.Client(
    api_key='YOUR_API_KEY',
    endpoint='YOUR_CLUSTER_ENDPOINT'
)

# dimension=4 is used here for simplicity.
# Set this to match your actual embedding model's output dimension.
ret = client.create('hybrid_collection', dimension=4, metric='dotproduct')

collection = client.get('hybrid_collection')
assert collection

    Step 2: Insert a document with a sparse vector

    from dashvector import Doc

collection.insert(Doc(
    id='A',
    vector=[0.1, 0.2, 0.3, 0.4],
    sparse_vector={1: 0.3, 10: 0.4, 100: 0.3}
))

    The sparse_vector field maps keyword IDs to their weights. Use a sparse encoder like DashText to generate these from raw text — see Sparse vector encoder.

    Step 3: Query with both vector types

    Pass both a vector and a sparse_vector in the same query. DashVector scores each result by summing the dot products from both vector types.

    docs = collection.query(
    vector=[0.1, 0.1, 0.1, 0.1],
    sparse_vector={1: 0.3, 20: 0.7}
)

    Results are ranked by the combined dot product score — higher is more relevant.

    Sparse vector encoder

    DashText is the recommended sparse vector encoder for DashVector. It converts raw text into the sparse_vector format shown above. See DashText quick start for setup instructions.

    What's next