SAP HANA Cloud 向量引擎集成 - AI 开发中文文档

SAP HANA Cloud Vector Engine 是一个完全集成在 SAP HANA Cloud 数据库中的向量存储。

设置

安装 langchain-hana 外部集成包以及本笔记本中使用的其他包。

pip install langchain-hana

凭证

确保您的SAP HANA实例正在运行。从环境变量加载数据库凭据并创建连接：

import os

from dotenv import load_dotenv
from hdbcli import dbapi

load_dotenv()
# Use connection settings from the environment
connection = dbapi.connect(
    address=os.environ.get("HANA_DB_ADDRESS"),
    port=os.environ.get("HANA_DB_PORT"),
    user=os.environ.get("HANA_DB_USER"),
    password=os.environ.get("HANA_DB_PASSWORD")
)

Learn more about SAP HANA in What is SAP HANA?.

初始化

初始化 HanaDB 向量存储需要数据库连接和嵌入实例。SAP HANA Cloud Vector Engine支持外部和内部嵌入。

使用外部嵌入

# | output: false
# | echo: false
from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(model="text-embedding-3-large")

使用内部嵌入

Alternatively，您可以直接在SAP HANA中通过其原生 VECTOR_EMBEDDING() 函数计算嵌入。要启用此功能，创建一个使用 HanaInternalEmbeddings 内部模型ID的实例，并将其传递给 HanaDB。请注意， HanaInternalEmbeddings 实例专门设计用于与 HanaDB 配合使用，不适用于其他向量存储实现。如需了解更多关于内部嵌入的信息，请参考SAP HANA VECTOR_EMBEDDING Function。

Caution: 请确保您的SAP HANA Cloud实例已启用NLP功能。

from langchain_hana import HanaInternalEmbeddings

embeddings = HanaInternalEmbeddings(internal_embedding_model_id="SAP_NEB.20240715")

# optionally, you can specify a remote source to use models from your deployed SAP AI CORE instance

# embeddings = HanaInternalEmbeddings(
#     internal_embedding_model_id="your-embedding-model-id",
#     remote_source_schema="your-remote-source-schema-name",
#     remote_source="your-remote-source-name",
# )

在获得连接和嵌入实例后，通过 HanaDB 创建向量存储，并传入存储向量的表名：

from langchain_hana import HanaDB

db = HanaDB(
    embedding=embeddings, connection=connection, table_name="MY_TABLE"
)

管理向量存储

在创建向量存储后，我们可以通过添加和删除不同项目与之交互。

向向量存储添加项

我们可通过 add_documents 函数向向量存储添加项目。

from langchain_core.documents import Document

docs = [Document(page_content="Some text"), Document(page_content="Other docs")]
db.add_documents(docs)

添加带元数据的文档。

docs = [
    Document(
        page_content="foo",
        metadata={"start": 100, "end": 150, "doc_name": "foo.txt", "quality": "bad"},
    ),
    Document(
        page_content="bar",
        metadata={"start": 200, "end": 250, "doc_name": "bar.txt", "quality": "good"},
    ),
]
db.add_documents(docs)

添加带map merge插入的文档。

Note use_map_merge 仅作为关键字参数接受，且仅适用于 HanaInternalEmbeddings 实例。

# a use_map_merge flag can be supplied for faster insertion
db.add_documents(docs, use_map_merge=True)

从向量存储删除项

db.delete(filter={"quality": "bad"})

查询向量存储

直接查询

相似度搜索

可通过以下方式执行带元数据过滤的简单相似度搜索：

docs = db.similarity_search("foobar", k=2, filter={"quality": "bad"})
# With filtering on "quality"=="bad", only one document should be returned
for doc in docs:
    print("-" * 80)
    print(doc.page_content)
    print(doc.metadata)

--------------------------------------------------------------------------------
foo
{'start': 100, 'end': 150, 'doc_name': 'foo.txt', 'quality': 'bad'}

MMR 搜索

可通过以下方式执行带元数据过滤的Maximal Marginal Relevance（MMR）搜索：

docs = db.max_marginal_relevance_search("foobar", k=2, fetch_k=5, filter={"quality": "bad"})
for doc in docs:
    print("-" * 80)
    print(doc.page_content)
    print(doc.metadata)

--------------------------------------------------------------------------------
foo
{'start': 100, 'end': 150, 'doc_name': 'foo.txt', 'quality': 'bad'}

将向量存储转换为检索器

您可以将向量存储转换为检索器以更轻松地将其集成到您的链中。

retriever = db.as_retriever()
docs = retriever.invoke("foobar", filter={"quality": "good"})
for doc in docs:
    print("-" * 80)
    print(doc.page_content)
    print(doc.metadata)

--------------------------------------------------------------------------------
bar
{'start': 200, 'end': 250, 'doc_name': 'bar.txt', 'quality': 'good'}

距离相似性算法

HanaDB 支持以下距离相似性算法：

余弦相似度（默认）
欧氏距离（L2）

您可以通过使用 distance_strategy 参数在初始化 HanaDB 实例时指定距离策略。

from langchain_hana.utils import DistanceStrategy
db = HanaDB(
    embedding=embeddings,
    connection=connection,
    distance_strategy=DistanceStrategy.EUCLIDEAN_DISTANCE,
    # distance_strategy=DistanceStrategy.COSINE_SIMILARITY,  # (default)
    table_name="MY_TABLE",
)

创建 HNSW 索引

向量索引可显著加速向量 top-k 最近邻查询。用户可通过 create_hnsw_index 函数创建分层导航小世界（HNSW）向量索引。对于数据库层面的索引创建说明，请参考官方文档。

db = HanaDB(
    embedding=embeddings, connection=connection, table_name="MY_TABLE"
)
db.create_hnsw_index(
    index_name="MY_TABLE_index",
    m=100,  # Max number of neighbors per graph node (valid range: 4 to 1000)
    ef_construction=200,  # Max number of candidates during graph construction (valid range: 1 to 100000)
    ef_search=500,  # Min number of candidates during the search (valid range: 1 to 100000)
)

若未指定其他参数，将使用默认值默认值：m=64, ef_construction=128, ef_search=200 默认索引名称为：<TABLE_NAME>_idx

高级过滤

除基础值过滤功能外，还支持更复杂的过滤操作。下表列出了可用过滤运算符。

运算符	语义解释	支持的操作数数据类型
`$eq`	等于（==）	int, float, str, bool, date 对象, None
`$ne`	不等于（!=）	int, float, str, bool, date 对象, None
`$lt`	小于（<）	int, float, str, date 对象
`$lte`	小于等于（<=）	int, float, str, date 对象
`$gt`	大于（>）	int, float, str, date 对象
`$gte`	大于等于（>=）	int, float, str, date 对象
`$in`	包含在给定值集合中（in）	同类型值列表（int, float, str, bool, date 对象）
`$nin`	不包含在给定值集合中（not in）	同类型值列表（int, float, str, bool, date 对象）
`$between`	位于两个边界值范围内	两个值列表（int, float, str, date 对象）
`$like`	基于SQL LIKE语义的文本相等（使用%作为通配符）	str
`$contains`	过滤包含特定关键词的文档	非空str
`$and`	逻辑”且”（支持多个操作数）	过滤器列表
`$or`	逻辑”或”（支持多个操作数）	过滤器列表

# Prepare some test documents
docs = [
    Document(
        page_content="First",
        metadata={"name": "Adam Smith", "is_active": True, "id": 1, "height": 10.0},
    ),
    Document(
        page_content="Second",
        metadata={"name": "Bob Johnson", "is_active": False, "id": 2, "height": 5.7},
    ),
    Document(
        page_content="Third",
        metadata={"name": "Jane Doe", "is_active": True, "id": 3, "height": 2.4},
    ),
]

db = HanaDB(
    connection=connection,
    embedding=embeddings,
    table_name="LANGCHAIN_DEMO_ADVANCED_FILTER",
)

# Delete already existing documents from the table
db.delete(filter={})
db.add_documents(docs)


# Helper function for printing filter results
def print_filter_result(result):
    if len(result) == 0:
        print("<empty result>")
    for doc in result:
        print(doc.metadata)

使用 $ne, $gt, $gte, $lt, $lte 进行过滤

advanced_filter = {"id": {"$ne": 1}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"id": {"$gt": 1}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"id": {"$gte": 1}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"id": {"$lt": 1}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"id": {"$lte": 1}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

Filter: {'id': {'$ne': 1}}
{'name': 'Jane Doe', 'is_active': True, 'id': 3, 'height': 2.4}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}
Filter: {'id': {'$gt': 1}}
{'name': 'Jane Doe', 'is_active': True, 'id': 3, 'height': 2.4}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}
Filter: {'id': {'$gte': 1}}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}
{'name': 'Jane Doe', 'is_active': True, 'id': 3, 'height': 2.4}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}
Filter: {'id': {'$lt': 1}}
<empty result>
Filter: {'id': {'$lte': 1}}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}

使用 $between, $in, $nin 进行过滤

advanced_filter = {"id": {"$between": (1, 2)}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"name": {"$in": ["Adam Smith", "Bob Johnson"]}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"name": {"$nin": ["Adam Smith", "Bob Johnson"]}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

Filter: {'id': {'$between': (1, 2)}}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}
Filter: {'name': {'$in': ['Adam Smith', 'Bob Johnson']}}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}
Filter: {'name': {'$nin': ['Adam Smith', 'Bob Johnson']}}
{'name': 'Jane Doe', 'is_active': True, 'id': 3, 'height': 2.4}

使用 $like 进行文本过滤

advanced_filter = {"name": {"$like": "a%"}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"name": {"$like": "%a%"}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

Filter: {'name': {'$like': 'a%'}}
<empty result>
Filter: {'name': {'$like': '%a%'}}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}
{'name': 'Jane Doe', 'is_active': True, 'id': 3, 'height': 2.4}

使用 $contains 进行文本过滤

advanced_filter = {"name": {"$contains": "bob"}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"name": {"$contains": "bo"}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"name": {"$contains": "Adam Johnson"}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"name": {"$contains": "Adam Smith"}}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

Filter: {'name': {'$contains': 'bob'}}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}
Filter: {'name': {'$contains': 'bo'}}
<empty result>
Filter: {'name': {'$contains': 'Adam Johnson'}}
<empty result>
Filter: {'name': {'$contains': 'Adam Smith'}}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}

使用 $and, $or 进行组合过滤

advanced_filter = {"$or": [{"id": 1}, {"name": "bob"}]}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"$and": [{"id": 1}, {"id": 2}]}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {"$or": [{"id": 1}, {"id": 2}, {"id": 3}]}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

advanced_filter = {
    "$and": [{"name": {"$contains": "bob"}}, {"name": {"$contains": "johnson"}}]
}
print(f"Filter: {advanced_filter}")
print_filter_result(db.similarity_search("just testing", k=5, filter=advanced_filter))

Filter: {'$or': [{'id': 1}, {'name': 'bob'}]}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}
Filter: {'$and': [{'id': 1}, {'id': 2}]}
<empty result>
Filter: {'$or': [{'id': 1}, {'id': 2}, {'id': 3}]}
{'name': 'Adam Smith', 'is_active': True, 'id': 1, 'height': 10.0}
{'name': 'Jane Doe', 'is_active': True, 'id': 3, 'height': 2.4}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}
Filter: {'$and': [{'name': {'$contains': 'bob'}}, {'name': {'$contains': 'johnson'}}]}
{'name': 'Bob Johnson', 'is_active': False, 'id': 2, 'height': 5.7}

检索增强生成（RAG）应用

关于如何使用此向量存储实现检索增强生成（RAG），请参考以下章节：

结果重排序

可通过向 similarity_search 方法传递 rerank_config 字典对相似度搜索结果进行重排序。重排序使用 SAP HANA 的 CROSS_ENCODE 函数及跨编码模型实现。 rerank_config 字典支持以下字段：

query (str, 必填) — 用于重排序的查询文本。未提供时使用原始相似度搜索查询。
model_id (str, 必填) — 跨编码模型ID
top_n (int, 可选) — 重排序后返回的顶部结果数（默认3）
rank_fields (list[str], 可选) — 需包含在重排序中的元数据字段列表

注意：确保您的SAP HANA Cloud实例已启用NLP功能以使用重排序功能。更多详情请见 SAP HANA数据库附加功能。

# Prepare sample documents with metadata
docs = [
    Document(
        page_content="Python is a programming language",
        metadata={"category": "programming", "difficulty": "beginner"},
    ),
    Document(
        page_content="Machine learning uses algorithms to learn patterns",
        metadata={"category": "AI", "difficulty": "intermediate"},
    ),
    Document(
        page_content="Neural networks are inspired by the human brain",
        metadata={"category": "AI", "difficulty": "advanced"},
    ),
]

db = HanaDB(
    connection=connection,
    embedding=embeddings,
    table_name="LANGCHAIN_DEMO_RERANKING",
)

# Delete existing entries and add new documents
db.delete(filter={})
db.add_documents(docs)

基础重排序

使用 rerank_config 参数执行相似度搜索并重排序：

# Basic reranking - rerank similarity search results using a cross-encoding model
rerank_config = {
    "model_id": os.environ.get("HANA_DB_RERANK_MODEL_ID"),  # e.g., "SAP_CER.20250701"
    "top_n": 2,
}

results = db.similarity_search(
    "AI technology",
    k=3,
    rerank_config=rerank_config
)

print("Reranked results:")
for doc in results:
    print("-" * 80)
    print(f"Content: {doc.page_content}")
    print(f"Metadata: {doc.metadata}")

Reranked results:
--------------------------------------------------------------------------------
Content: Machine learning uses algorithms to learn patterns
Metadata: {'category': 'AI', 'difficulty': 'intermediate'}
--------------------------------------------------------------------------------
Content: Python is a programming language
Metadata: {'category': 'programming', 'difficulty': 'beginner'}

带元数据的重排序

通过 rank_fields 在重排序过程中包含元数据。当元数据包含应影响排名的相关信息时，此功能非常有用：

# Reranking with metadata fields included in the scoring
rerank_config_with_fields = {
    "query": "beginner AI concepts",
    "model_id": os.environ.get("HANA_DB_RERANK_MODEL_ID"),
    "top_n": 2,
    "rank_fields": ["category", "difficulty"],  # Include these metadata fields in reranking
}

results = db.similarity_search(
    "learning algorithms",
    k=3,
    rerank_config=rerank_config_with_fields
)

print("Reranked results with metadata fields:")
for doc in results:
    print("-" * 80)
    print(f"Content: {doc.page_content}")
    print(f"Metadata: {doc.metadata}")

Reranked results with metadata fields:
--------------------------------------------------------------------------------
Content: Neural networks are inspired by the human brain
Metadata: {'category': 'AI', 'difficulty': 'advanced'}
--------------------------------------------------------------------------------
Content: Machine learning uses algorithms to learn patterns
Metadata: {'category': 'AI', 'difficulty': 'intermediate'}

带分数的重排序

使用 similarity_search_with_score 获取与文档一同返回的重排序分数：

# Get reranking scores with the results
rerank_config = {
    "model_id": os.environ.get("HANA_DB_RERANK_MODEL_ID"),
    "top_n": 3,
}

results_with_scores = db.similarity_search_with_score(
    "neural network architecture",
    k=3,
    rerank_config=rerank_config
)

print("Reranked results with scores:")
for doc, score in results_with_scores:
    print("-" * 80)
    print(f"Score: {score:.4f}")
    print(f"Content: {doc.page_content}")

Reranked results with scores:
--------------------------------------------------------------------------------
Score: 0.0435
Content: Neural networks are inspired by the human brain
--------------------------------------------------------------------------------
Score: 0.0146
Content: Python is a programming language
--------------------------------------------------------------------------------
Score: 0.0145
Content: Machine learning uses algorithms to learn patterns

独立重排序使用 HanaReranker

若需对重排序过程进行更精细控制，可直接使用 HanaReranker 类作为文档压缩器。此方法适用于从任意来源而非仅相似度搜索结果中重排序文档：

from langchain_hana import HanaReranker

# Create a standalone reranker
reranker = HanaReranker(
    connection=connection,
    model_id=os.environ.get("HANA_DB_RERANK_MODEL_ID"),
)

# Documents to rerank (can come from any source)
documents = [
    Document(page_content="Python programming basics"),
    Document(page_content="Advanced machine learning techniques"),
    Document(page_content="Introduction to neural networks"),
    Document(page_content="Deep learning applications"),
    Document(page_content="Reinforcement learning strategies"),
    Document(page_content="Natural language processing techniques"),
]

# Rerank documents based on a query using compress_documents (returns top min(5, len(documents)))
# Reranking scores will be added to the metadata of each document under the key "relevance_score"
reranked_docs = reranker.compress_documents(documents, query="AI and deep learning")

print("Compressed documents:")
for doc in reranked_docs:
    print("-" * 80)
    print(f"Content: {doc.page_content}")
    print(f"Relevance score: {doc.metadata.get('relevance_score'):.4f}")

# Or use the rerank method for more control over top_n
results = reranker.rerank(documents, query="machine learning", top_n=2)
print("\nTop 2 reranked results:")
for idx, score, doc in results:
    print(f"  [{idx}] Score: {score:.4f} - {doc.page_content}")

Compressed documents:
--------------------------------------------------------------------------------
Content: Deep learning applications
Relevance score: 0.2188
--------------------------------------------------------------------------------
Content: Advanced machine learning techniques
Relevance score: 0.0748
--------------------------------------------------------------------------------
Content: Introduction to neural networks
Relevance score: 0.0079
--------------------------------------------------------------------------------
Content: Natural language processing techniques
Relevance score: 0.0040
--------------------------------------------------------------------------------
Content: Reinforcement learning strategies
Relevance score: 0.0036

Top 2 reranked results:
  [1] Score: 0.4351 - Advanced machine learning techniques
  [3] Score: 0.0342 - Deep learning applications

标准表与”自定义”向量数据表

默认情况下，嵌入表包含以下三列：

一列 VEC_TEXT，存储文档的文本内容
一列 VEC_META，存储文档的元数据
一列 VEC_VECTOR，存储文档文本的嵌入向量

# Access the vector DB with a new table
db = HanaDB(
    connection=connection, embedding=embeddings, table_name="LANGCHAIN_DEMO_NEW_TABLE"
)

# Delete already existing entries from the table
db.delete(filter={})

# Add a simple document with some metadata
docs = [
    Document(
        page_content="A simple document",
        metadata={"start": 100, "end": 150, "doc_name": "simple.txt"},
    )
]
db.add_documents(docs)

显示 ‘LANGCHAIN_DEMO_NEW_TABLE’ 表中的列

cur = connection.cursor()
cur.execute(
    "SELECT COLUMN_NAME, DATA_TYPE_NAME FROM SYS.TABLE_COLUMNS WHERE SCHEMA_NAME = CURRENT_SCHEMA AND TABLE_NAME = 'LANGCHAIN_DEMO_NEW_TABLE'"
)
rows = cur.fetchall()
for row in rows:
    print(row)
cur.close()

('VEC_META', 'NCLOB')
('VEC_TEXT', 'NCLOB')
('VEC_VECTOR', 'REAL_VECTOR')

显示插入文档在三个列中的值由于 HANA 的数据库 API 驱动默认将向量列输出为 fvecs 字节对象，我们将创建辅助函数将其转换为数值列表。

import struct
# Helper function to parse fvecs format for REAL_VECTOR
def parseFvecs(fvecs):
    dim = struct.unpack_from("<I", fvecs, 0)[0]
    return list(struct.unpack_from("<%sf" % str(dim), fvecs, 4))

cur = connection.cursor()
cur.execute(
    "SELECT * FROM LANGCHAIN_DEMO_NEW_TABLE LIMIT 1"
)
rows = cur.fetchall()
print(rows[0][0])  # The text
print(rows[0][1])  # The metadata
embedding = parseFvecs(rows[0][2])
print(len(embedding), embedding[:3] + ['...'] + embedding[-3:])  # The vector
cur.close()

A simple document
{"start": 100, "end": 150, "doc_name": "simple.txt"}
768 [-0.01989901065826416, 0.02785174734890461, 0.0020877711940556765, '...', 0.0183248370885849, 0.009469633921980858, 0.04312701150774956]

自定义表必须至少包含三列以匹配标准表的语义：

一列用于嵌入的文本/上下文（类型 NCLOB 或 NVARCHAR）
一列用于元数据（类型 NCLOB 或 NVARCHAR）
一列用于嵌入向量（类型 REAL_VECTOR 或 HALF_VECTOR）

表可包含额外列。当新文档插入表时，这些额外列必须允许 NULL 值。

# Create a new table "MY_OWN_TABLE_ADD" with three "standard" columns and one additional column
my_own_table_name = "MY_OWN_TABLE_ADD"
cur = connection.cursor()
cur.execute(
    (
        f"CREATE TABLE {my_own_table_name} ("
        "SOME_OTHER_COLUMN NVARCHAR(42), "
        "MY_TEXT NVARCHAR(2048), "
        "MY_METADATA NVARCHAR(1024), "
        "MY_VECTOR REAL_VECTOR )"
    )
)

# Create a HanaDB instance with the own table
db = HanaDB(
    connection=connection,
    embedding=embeddings,
    table_name=my_own_table_name,
    content_column="MY_TEXT",
    metadata_column="MY_METADATA",
    vector_column="MY_VECTOR",
)

# Add a simple document with some metadata
docs = [
    Document(
        page_content="Some other text",
        metadata={"start": 400, "end": 450, "doc_name": "other.txt"},
    )
]
db.add_documents(docs)

# Check if data has been inserted into our own table
cur.execute(f"SELECT SOME_OTHER_COLUMN, MY_TEXT, MY_METADATA, TO_NVARCHAR(MY_VECTOR) AS MY_VECTOR FROM {my_own_table_name} LIMIT 1")
rows = cur.fetchall()
print(rows[0][0])  # Value of column "SOME_OTHER_COLUMN". Should be NULL/None
print(rows[0][1])  # The text
print(rows[0][2])  # The metadata
embedding = parseFvecs(rows[0][3])
print(len(embedding), embedding[:3] + ['...'] + embedding[-3:])  # The vector

cur.close()

None
Some other text
{"start": 400, "end": 450, "doc_name": "other.txt"}
768 [0.016170687973499298, -0.01129427831619978, -0.0005921399570070207, '...', 0.017849743366241455, 0.0003932560794055462, -0.00045805066474713385]

向自定义表添加另一文档并执行相似度搜索。

docs = [
    Document(
        page_content="Some more text",
        metadata={"start": 800, "end": 950, "doc_name": "more.txt"},
    )
]
db.add_documents(docs)

query = "What's up?"
docs = db.similarity_search(query, k=2)
for doc in docs:
    print("-" * 80)
    print(doc.page_content)

--------------------------------------------------------------------------------
Some more text
--------------------------------------------------------------------------------
Some other text

自定义列性能优化

为支持灵活的元数据值，默认将所有元数据存储为元数据列的 JSON。若已知部分元数据键及其值类型，可通过在 HanaDB 构造函数中传递 specific_metadata_columns 列表指定目标表结构。插入文档时，匹配的元数据键会复制到特殊列。过滤器将优先使用特殊列而非元数据 JSON 列（针对 specific_metadata_columns 列表中的键）。

# Create a new table "PERFORMANT_CUSTOMTEXT_FILTER" with three "standard" columns and one additional column
my_own_table_name = "PERFORMANT_CUSTOMTEXT_FILTER"
cur = connection.cursor()
cur.execute(
    (
        f"CREATE TABLE {my_own_table_name} ("
        "CUSTOMTEXT NVARCHAR(500), "
        "MY_TEXT NVARCHAR(2048), "
        "MY_METADATA NVARCHAR(1024), "
        "MY_VECTOR REAL_VECTOR )"
    )
)

# Create a HanaDB instance with the own table
db = HanaDB(
    connection=connection,
    embedding=embeddings,
    table_name=my_own_table_name,
    content_column="MY_TEXT",
    metadata_column="MY_METADATA",
    vector_column="MY_VECTOR",
    specific_metadata_columns=["CUSTOMTEXT"],
)

# Add a simple document with some metadata
docs = [
    Document(
        page_content="Some other text",
        metadata={
            "start": 400,
            "end": 450,
            "doc_name": "other.txt",
            "CUSTOMTEXT": "Filters on this value are very performant",
        },
    )
]
db.add_documents(docs)

# Check if data has been inserted into our own table
cur.execute(f"SELECT * FROM {my_own_table_name} LIMIT 1")
rows = cur.fetchall()
print(
    rows[0][0]
)  # Value of column "CUSTOMTEXT". Should be "Filters on this value are very performant"
print(rows[0][1])  # The text
print(
    rows[0][2]
)  # The metadata without the "CUSTOMTEXT" data, as this is extracted into a separate column
embedding = parseFvecs(rows[0][3])
print(len(embedding), embedding[:3] + ['...'] + embedding[-3:])  # The vector

cur.close()

Filters on this value are very performant
Some other text
{"start": 400, "end": 450, "doc_name": "other.txt", "CUSTOMTEXT": "Filters on this value are very performant"}
768 [0.016170687973499298, -0.01129427831619978, -0.0005921399570070207, '...', 0.017849743366241455, 0.0003932560794055462, -0.00045805066474713385]

特殊列对 LangChain 接口完全透明。所有原有功能均保持不变，但性能有所提升。

docs = [
    Document(
        page_content="Some more text",
        metadata={
            "start": 800,
            "end": 950,
            "doc_name": "more.txt",
            "CUSTOMTEXT": "Another customtext value",
        },
    )
]
db.add_documents(docs)

advanced_filter = {"CUSTOMTEXT": {"$like": "%value%"}}
query = "What's up?"
docs = db.similarity_search(query, k=2, filter=advanced_filter)
for doc in docs:
    print("-" * 80)
    print(doc.page_content)

--------------------------------------------------------------------------------
Some more text
--------------------------------------------------------------------------------
Some other text

简单示例

加载样本文档 “state_of_the_union.txt” 并生成文本块。

langchain-community包已不再维护。从langchain_community导入的示例可能过时或无法使用。请谨慎使用。

from langchain_community.document_loaders import TextLoader
from langchain_core.documents import Document
from langchain_openai import OpenAIEmbeddings
from langchain_text_splitters import CharacterTextSplitter

text_documents = TextLoader(
    "../../how_to/state_of_the_union.txt", encoding="UTF-8"
).load()
text_splitter = CharacterTextSplitter(chunk_size=500, chunk_overlap=0)
text_chunks = text_splitter.split_documents(text_documents)
print(f"Number of document chunks: {len(text_chunks)}")

Number of document chunks: 88

将加载的文档块插入表。本示例会删除表内可能存在的先前内容（来自之前运行的残留数据）。

# Delete already existing documents from the table
db.delete(filter={})

# add the loaded document chunks
db.add_documents(text_chunks)

对前一步插入的文档块执行查询，获取匹配度最高的两个文档块。默认使用余弦相似度作为搜索指标。

query = "What did the president say about Ketanji Brown Jackson"
docs = db.similarity_search(query, k=2)

for doc in docs:
    print("-" * 80)
    print(doc.page_content)

--------------------------------------------------------------------------------
One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court.

And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.
--------------------------------------------------------------------------------
As I said last year, especially to our younger transgender Americans, I will always have your back as your President, so you can be yourself and reach your God-given potential.

While it often appears that we never agree, that isn’t true. I signed 80 bipartisan bills into law last year. From preventing government shutdowns to protecting Asian-Americans from still-too-common hate crimes to reforming military justice.

使用欧氏距离重新执行相同查询。结果应与余弦相似度查询一致。

from langchain_hana.utils import DistanceStrategy

db = HanaDB(
    embedding=embeddings,
    connection=connection,
    distance_strategy=DistanceStrategy.EUCLIDEAN_DISTANCE,
    table_name="STATE_OF_THE_UNION",
)

query = "What did the president say about Ketanji Brown Jackson"
docs = db.similarity_search(query, k=2)
for doc in docs:
    print("-" * 80)
    print(doc.page_content)

--------------------------------------------------------------------------------
One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court.

And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.
--------------------------------------------------------------------------------
As I said last year, especially to our younger transgender Americans, I will always have your back as your President, so you can be yourself and reach your God-given potential.

While it often appears that we never agree, that isn’t true. I signed 80 bipartisan bills into law last year. From preventing government shutdowns to protecting Asian-Americans from still-too-common hate crimes to reforming military justice.

最大边际相关性搜索（MMR）

Maximal marginal relevance 优化查询与文档的相似度 AND 文档间的多样性。首先从数据库获取前 20（fetch_k）条结果，MMR 算法将从中选出最佳 2（k）条匹配项。

docs = db.max_marginal_relevance_search(query, k=2, fetch_k=20)
for doc in docs:
    print("-" * 80)
    print(doc.page_content)

--------------------------------------------------------------------------------
One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court.

And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.
--------------------------------------------------------------------------------
Groups of citizens blocking tanks with their bodies. Everyone from students to retirees teachers turned soldiers defending their homeland.

In this struggle as President Zelenskyy said in his speech to the European Parliament “Light will win over darkness.” The Ukrainian Ambassador to the United States is here tonight.

Let each of us here tonight in this Chamber send an unmistakable signal to Ukraine and to the world.

创建 HNSW 向量索引

使用 create_hnsw_index 函数创建分层导航小世界（HNSW）向量索引，可显著加速向量查询的 top-k 最近邻搜索。

# HanaDB instance uses cosine similarity as default:
db_cosine = HanaDB(
    embedding=embeddings, connection=connection, table_name="STATE_OF_THE_UNION"
)

# Attempting to create the HNSW index with default parameters
db_cosine.create_hnsw_index()  # If no other parameters are specified, the default values will be used
# Default values: m=64, ef_construction=128, ef_search=200
# The default index name will be: STATE_OF_THE_UNION_COSINE_idx


# Creating a HanaDB instance with L2 distance as the similarity function and defined values
db_l2 = HanaDB(
    embedding=embeddings,
    connection=connection,
    table_name="STATE_OF_THE_UNION",
    distance_strategy=DistanceStrategy.EUCLIDEAN_DISTANCE,  # Specify L2 distance
)

# This will create an index based on L2 distance strategy.
db_l2.create_hnsw_index(
    index_name="STATE_OF_THE_UNION_L2_index",
    m=100,  # Max number of neighbors per graph node (valid range: 4 to 1000)
    ef_construction=200,  # Max number of candidates during graph construction (valid range: 1 to 100000)
    ef_search=500,  # Min number of candidates during the search (valid range: 1 to 100000)
)

# Use L2 index to perform MMR
docs = db_l2.max_marginal_relevance_search(query, k=2, fetch_k=20)
for doc in docs:
    print("-" * 80)
    print(doc.page_content)

--------------------------------------------------------------------------------
One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court.

And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.
--------------------------------------------------------------------------------
Groups of citizens blocking tanks with their bodies. Everyone from students to retirees teachers turned soldiers defending their homeland.

In this struggle as President Zelenskyy said in his speech to the European Parliament “Light will win over darkness.” The Ukrainian Ambassador to the United States is here tonight.

Let each of us here tonight in this Chamber send an unmistakable signal to Ukraine and to the world.

关键点：

相似度函数：索引默认使用余弦相似度。若需使用其他相似度函数（如 L2 距离），需在 HanaDB 实例初始化时指定。
默认参数：在 create_hnsw_index 函数中，若未提供 m、ef_construction 或 ef_search 等参数的定制值，将自动使用 m=64、ef_construction=128、ef_search=200 等默认值。这些值确保索引在合理性能下无需用户干预即可创建。

官方原文：src/oss/python/integrations/vectorstores/sap_hanavector.mdx · 非官方翻译，译文由 AI 生成，以英文原文为准。

​设置

​凭证

​初始化

​使用外部嵌入

​使用内部嵌入

​管理向量存储

​向向量存储添加项

​从向量存储删除项

​查询向量存储

​直接查询

​相似度搜索

​MMR 搜索

​将向量存储转换为检索器

​距离相似性算法

​创建 HNSW 索引

​高级过滤

​检索增强生成（RAG）应用

​结果重排序

​基础重排序

​带元数据的重排序

​带分数的重排序

​独立重排序使用 HanaReranker

​标准表与”自定义”向量数据表

​自定义列性能优化

​简单示例

​最大边际相关性搜索（MMR）

​创建 HNSW 向量索引

设置

凭证

初始化

使用外部嵌入

使用内部嵌入

管理向量存储

向向量存储添加项

从向量存储删除项

查询向量存储

直接查询

相似度搜索

MMR 搜索

将向量存储转换为检索器

距离相似性算法

创建 HNSW 索引

高级过滤

检索增强生成（RAG）应用

结果重排序

基础重排序

带元数据的重排序

带分数的重排序

独立重排序使用 HanaReranker

标准表与”自定义”向量数据表

自定义列性能优化

简单示例

最大边际相关性搜索（MMR）

创建 HNSW 向量索引