LanceDB 中文文档

外观

本页尚未翻译

本页内容仍为英文原文，源自官方文档 integrations/embedding/superlinked。中文翻译正在分批进行中。

Superlinked ​

Superlinked is a self-hosted inference engine (SIE) for embedding, reranking, and extraction. The sie-lancedb package registers SIE as a first-class embedding function in LanceDB's embeddings registry, so embeddings are computed automatically on insert and search. You need a running SIE instance - see the Superlinked quickstart for deployment options.

Installation ​

bash
pip install sie-lancedb
bash
npm install @superlinked/sie-lancedb @lancedb/lancedb

Registered functions ​

Importing sie_lancedb registers two embedding functions in LanceDB's registry:

NamePurpose
"sie"Dense text embeddings
"sie-multivector"ColBERT-style late interaction with MaxSim scoring

Supported parameters on .create():

ParameterTypeDescription
modelstrAny of 85+ SIE-supported models (e.g. BAAI/bge-m3, NovaSearch/stella_en_400M_v5, jinaai/jina-colbert-v2)
base_urlstrURL of the SIE endpoint (e.g. http://localhost:8080)

Usage ​

python
import tempfile
from pathlib import Path

from lancedb.embeddings import get_registry
from lancedb.pydantic import LanceModel, Vector
import sie_lancedb  # registers "sie" and "sie-multivector"

sie = get_registry().get("sie").create(
    model="BAAI/bge-m3",
    base_url="http://localhost:8080",
)

class Documents(LanceModel):
    id: int
    text: str = sie.SourceField()
    vector: Vector(sie.ndims()) = sie.VectorField()

db = lancedb.connect(str(Path(tempfile.mkdtemp()) / "superlinked-demo"))
table = db.create_table("docs", schema=Documents)

table.add([
    {"id": 1, "text": "Machine learning is a subset of AI."},
    {"id": 2, "text": "Neural networks use multiple layers."},
    {"id": 3, "text": "Python is popular for ML development."},
])

results = table.search("What is deep learning?").limit(3).to_list()

LanceDB handles embedding generation for both inserts and queries automatically, based on the SourceField / VectorField declarations on the schema. The id column gives each document a stable key, which entity extraction uses to merge its results back.

Hybrid search with reranker ​

SIEReranker plugs into LanceDB's hybrid search pipeline. It uses SIE's cross-encoder score() to rerank combined vector + full-text search results. You need a full-text search index on the column first:

python
from sie_lancedb import SIEReranker

# Create FTS index for hybrid search
table.create_fts_index("text", replace=True)

results = (
    table.search("What is deep learning?", query_type="hybrid")
    .rerank(SIEReranker(model="jinaai/jina-reranker-v2-base-multilingual"))
    .limit(5)
    .to_list()
)

for r in results:
    print(f"{r['_relevance_score']:.3f}  {r['text']}")

The reranker also works with pure vector or pure FTS search via .rerank().

ColBERT and multivector ​

SIEMultiVectorEmbeddingFunction (registered as "sie-multivector") works with LanceDB's native MultiVector type and MaxSim scoring for ColBERT and ColPali models. In LanceDB Python 0.40.0, a multivector that LanceDB generates on insert is stored as null unless the document's token count happens to equal the per-token dimension, so this example computes the document multivectors with compute_source_embeddings() and passes them in the vector column. LanceDB still generates the query multivector automatically:

python
from lancedb.pydantic import MultiVector

sie_colbert = get_registry().get("sie-multivector").create(
    model="jinaai/jina-colbert-v2",
    base_url="http://localhost:8080",
)

class ColBERTDocs(LanceModel):
    text: str = sie_colbert.SourceField()
    vector: MultiVector(sie_colbert.ndims()) = sie_colbert.VectorField()

table = db.create_table("colbert_docs", schema=ColBERTDocs)

# Compute the document multivectors and store them with their text
texts = ["Machine learning is a subset of AI."]
vectors = sie_colbert.compute_source_embeddings(texts)
table.add([{"text": text, "vector": vector} for text, vector in zip(texts, vectors)])

# MaxSim search - query and document multivectors compared token-by-token
results = table.search("What is ML?").limit(5).to_list()

Entity extraction ​

SIEExtractor adds entity extraction to LanceDB's data-enrichment workflows. Extract entities from a text column and merge the results back as a structured Arrow column - enabling filtered search on extracted entities. This example enriches the docs table from Usage, matching rows on its id column:

python
from sie_lancedb import SIEExtractor

extractor = SIEExtractor(
    base_url="http://localhost:8080",
    model="urchade/gliner_multi-v2.1",
)

docs = db.open_table("docs")
extractor.enrich_table(
    docs,
    source_column="text",
    target_column="entities",
    labels=["person", "technology", "organization"],
    id_column="id",
)

The entities column stores structured Arrow data (list<struct<text, label, score, start, end, bbox>>), so you can filter on extracted entities in queries.