Hologres:Search SDK

Set a vector index

Configure the vector index before running vector searches. The following example uses Cosine distance with RaBitQ quantization:

table.set_vector_index(
    column="vector_column",
    distance_method="Cosine",
    base_quantization_type="rabitq",
    use_reorder=True,
    max_degree=64,
    ef_construction=400
)

For guidance on choosing index parameters, see HGraph Index User Guide.

Set a full-text index

# Create a full-text index with the jieba tokenizer
table.create_text_index(
    index_name="ft_idx_content",
    column="content",
    tokenizer="jieba"
)

# Switch to a different tokenizer
table.set_text_index(
    index_name="ft_idx_content",
    tokenizer="ik"
)

# Delete the index when no longer needed
table.drop_text_index(index_name="ft_idx_content")

Vector search

All vector search examples use search_vector and chain modifiers before calling fetchall.

query_vector = [0.2, 0.3, 0.4]

# Return the top 10 nearest vectors
results = table.search_vector(
    vector=query_vector,
    column="vector_column",
    distance_method="Cosine"
).limit(10).fetchall()

# Filter out low-similarity results — use min_distance when you want to exclude
# results below a relevance threshold rather than capping by count
results = table.search_vector(
    vector=query_vector,
    column="vector_column",
    distance_method="Cosine"
).min_distance(0.5).fetchall()

# Rename the distance output column — use output_name when downstream code
# expects a specific column name (for example, when combining with text search scores)
results = table.search_vector(
    vector=query_vector,
    column="vector_column",
    output_name="similarity_score",
    distance_method="Cosine"
).fetchall()

Full-text search

All full-text search examples use search_text. The mode parameter controls how the expression is interpreted.

# Basic full-text search — returns all columns
results = table.search_text(
    column="content",
    expression="machine learning",
    return_all_columns=True
).fetchall()

# Include a BM25 relevance score in the results
results = table.search_text(
    column="content",
    expression="deep learning",
    return_score=True,
    return_score_name="relevance_score"
).select(["id", "vector_column", "content"]).fetchall()

# Keyword mode (default) — all keywords must be present
results = table.search_text(
    column="content",
    expression="python programming",
    mode="match",
    operator="AND"
).fetchall()

# Phrase mode — exact phrase match
results = table.search_text(
    column="content",
    expression="machine learning",
    mode="phrase"
).fetchall()

# Natural language mode — use + and - operators to include/exclude terms
results = table.search_text(
    column="content",
    expression="+python -java",
    mode="natural_language"
).fetchall()

# Term mode — exact match with no tokenization
results = table.search_text(
    column="content",
    expression="python",
    mode="term"
).fetchall()

Hybrid search

Combine a text or vector search with scalar filters by chaining .where(), .order_by(), and .limit().

# Full-text search filtered by date, sorted by relevance
results = (
    table.search_text(
        column="content",
        expression="artificial intelligence",
        return_score=True,
        return_score_name="score"
    )
    .where("publish_date > '2023-01-01'")
    .order_by("score", "desc")
    .limit(10)
    .fetchall()
)

# Vector search filtered by date, sorted by similarity
results = (
    table.search_vector(
        vector=query_vector,
        column="vector_column",
        output_name="similarity_score",
        distance_method="Cosine"
    )
    .where("publish_date > '2023-01-01'")
    .order_by("similarity_score", "desc")
    .limit(10)
    .fetchall()
)

Primary key point query

# Fetch a single record by primary key
result = table.get_by_key(
    key_column="id",
    key_value=1,
    return_columns=["id", "content", "vector_column"]  # Omit to return all columns
).fetchone()

# Fetch multiple records by primary key list
results = table.get_multi_by_keys(
    key_column="id",
    key_values=[1, 2, 3],
    return_columns=["id", "content"]  # Omit to return all columns
).fetchall()