Quickstart

This guide walks through the main LynseDB workflow: connect, create a collection, insert vectors or documents, search, filter metadata, inspect results, and clean up.

1. Install and import

pip install lynsedb

For documents= and search(document=...), install the optional local embedding adapter:

pip install "lynsedb[embeddings]"

Native Linux, macOS, and Windows environments are supported. On Windows, the Rust core is built natively with MSVC.

import numpy as np
import lynse

print(lynse.__version__)

2. Connect

Use local mode for a single Python process:

client = lynse.VectorDBClient(uri="./lynsedb-data")

Use remote mode when several processes or services need the same database:

lynse serve --host 127.0.0.1 --port 7637 --data-dir ./server-data

client = lynse.VectorDBClient("http://127.0.0.1:7637")

With server authentication:

lynse serve --host 127.0.0.1 --port 7637 --data-dir ./server-data --api-key your_key

client = lynse.VectorDBClient("http://127.0.0.1:7637", api_key="your_key")

Health and operations endpoints:

curl http://127.0.0.1:7637/healthz
curl http://127.0.0.1:7637/readyz
curl http://127.0.0.1:7637/metrics
curl http://127.0.0.1:7637/openapi.json

3. Create a database and collection

drop_if_exists=True is destructive. Use it only for tests or when you really want to truncate existing data.

db = client.create_database("quickstart", drop_if_exists=True)

collection = db.require_collection(
    "documents",
    dim=4,
    drop_if_exists=True,
    description="quickstart collection",
)

New collections build a FLAT-IP index automatically after the first primary vector write. Use default_index=None to disable this, or pass an index mode such as default_index="FLAT-COS" when creating the collection.

Open an existing collection safely:

collection = db.get_collection("documents")

4. Insert vectors

Each row has a public string or non-negative integer ID. Metadata fields are optional JSON-like dicts and can be used later for filtering, BM25 search, or result display.

with collection:
    collection.add(
        ids=["intro", "guide", "note-fr", "note-archive"],
        vectors=[
            [0.10, 0.20, 0.30, 0.40],
            [0.11, 0.19, 0.29, 0.39],
            [0.80, 0.10, 0.20, 0.10],
            [0.75, 0.12, 0.18, 0.12],
        ],
        fields=[
            {"title": "LynseDB intro", "lang": "en", "rank": 1, "tags": ["vector", "rust"]},
            {"title": "Vector guide", "lang": "en", "rank": 2, "tags": ["vector"]},
            {"title": "French note", "lang": "fr", "rank": 3, "tags": ["note"]},
            {"title": "Another note", "lang": "fr", "rank": 4, "tags": ["note", "archive"]},
        ],
    )

The collection context manager calls commit() on successful exit.

For large dense arrays, keep the same add() API and choose a larger batch size:

vectors = np.random.rand(10_000, 4).astype(np.float32)
added = collection.add(ids=[f"vec-{i}" for i in range(10_000)], vectors=vectors, batch_size=5000)
print(added)

You can also pass text documents and let LynseDB use the default local embedding adapter:

docs = db.require_collection("text_docs", drop_if_exists=True)

with docs:
    docs.add(
        ids=["local", "server"],
        documents=[
            "Use embedded mode for notebooks, tests, and jobs.",
            "Run lynse serve when several workers need shared search.",
        ],
        fields=[
            {"mode": "local"},
            {"mode": "server"},
        ],
    )

result = docs.search(document="How do I share search across workers?", k=1)
print(result.ids)

5. Build or tune an index

Flat search is the simplest and most recall-friendly default. LynseDB builds a FLAT-IP index automatically for new collections after the first write so the default metric is explicit. Use HNSW, IVF, or SPANN as data grows and latency matters, or rebuild with a metric-specific flat index when that matches your embedding model.

collection.build_index("FLAT-L2")
print(collection.index_mode)

IVF and SPANN use n_clusters; other index families allow the argument and ignore it:

collection.build_index("IVF-L2", n_clusters=256)
collection.build_index("SPANN-L2", n_clusters=256)

6. Search

query = np.array([0.10, 0.20, 0.30, 0.40], dtype=np.float32)

result = collection.search(query, k=2, return_fields=True)

print(result.ids)
print(result.distances)
print(result.fields)
print(result.to_list())

Filter by metadata during vector search:

result = collection.search(
    query,
    k=3,
    where="lang = 'en' AND rank <= 2",
    return_fields=True,
)
print(result.to_list())

For IVF, SPANN, and HNSW, nprobe controls search breadth. Higher values generally improve recall and increase latency.

result = collection.search(query, k=3, nprobe=20)

Approximate flat search is available for IP, L2, cosine, L1, Chebyshev, Canberra, and Bray–Curtis metrics:

result = collection.search(query, k=3, approx=True, eps=1e-4)

Flat, PQ, RaBitQ, PolarVec, and named vector-field searches ignore nprobe. Other domain and binary metrics keep their exact search paths.

7. Query metadata and vectors

Use query() when you need IDs and fields. Use query_vectors() when you need stored vectors too.

rows = collection.query(where="tags CONTAINS 'vector'")
print(rows.ids)
print(rows.fields)

vectors = collection.query_vectors(filter_ids=[1, 2])
print(vectors.ids)
print(vectors.vectors.shape)

Calling query() or query_vectors() without where or filter_ids returns an empty ResultView; it does not perform a full scan.

8. Text and hybrid search

BM25 search uses stored metadata fields:

text_result = collection.bm25_search(
    "vector guide",
    k=3,
    text_fields=["title"],
    return_fields=True,
)
print(text_result.to_list())

Hybrid search combines vector and text candidates:

hybrid = collection.hybrid_search(
    vector=query,
    text="vector",
    text_fields=["title", "tags"],
    fusion="rrf",
    k=3,
    return_fields=True,
)
print(hybrid.to_list())

9. Named and sparse vectors

Named vector fields store additional embeddings for the same IDs. This is useful for multimodal records, for example text and image embeddings on one item.

collection.create_vector_field("image", dim=3, metric="l2")

image_vectors = np.array(
    [
        [0.10, 0.20, 0.30],
        [0.12, 0.19, 0.28],
        [0.90, 0.20, 0.10],
    ],
    dtype=np.float32,
)

collection.add_named_vectors("image", image_vectors, ids=[1, 2, 3])
collection.build_index("HNSW-L2", field_name="image")
collection.commit()

image_result = collection.search(
    [0.11, 0.20, 0.29],
    k=2,
    vector_field="image",
    return_fields=True,
)
print(image_result.to_list())

Sparse vectors store feature-ID weights and search with inner product:

collection.add_sparse_vectors(
    vectors=[
        {10: 1.0, 42: 0.5},
        {11: 1.0, 42: 0.8},
    ],
    ids=["intro", "guide"],
)
collection.commit()

sparse_result = collection.search_sparse({42: 1.0}, k=2, return_fields=True)
print(sparse_result.to_list())

10. Update, delete, and compact

Use upsert when the same external ID should be replaced or inserted:

collection.upsert(
    ids="intro",
    vectors=[0.12, 0.20, 0.31, 0.41],
    fields={"title": "updated intro", "lang": "en", "rank": 1},
)
collection.commit()

Deletes are soft deletes. Deleted IDs disappear from search and query results, but their raw storage is kept until compaction:

collection.delete(["note-archive"])
print(collection.list_deleted_ids())

collection.restore(["note-archive"])
print(collection.list_deleted_ids())

collection.delete(["note-archive"])
removed = collection.compact()
print(removed)

11. What you learned

This quickstart touched the whole everyday workflow:

• choose local or remote mode with VectorDBClient;
• create a database and collection;
• insert vectors with stable public IDs and metadata fields;
• commit writes through insert_session();
• build and tune an index;
• search by vector, filter by metadata, query fields, and retrieve vectors;
• use document, BM25, hybrid, named vector, and sparse vector retrieval;
• update, soft-delete, restore, and compact rows.

For a complete curriculum, continue with the Learning path.

10. ResultView

Search, query, head, tail, and range APIs return ResultView.

result = collection.search(query, k=2, return_fields=True)

ids, distances, fields = result

print(result.ids)          # numpy array
print(result.distances)    # numpy array
print(result.fields)       # list[dict]
print(result.to_dict())
print(result.to_list())
print(result.to_json())

Use to_list() for row iteration:

for row in result.to_list():
    print(row["id"], row.get("title"))

Optional dataframe conversions are available when the dependency is installed:

result.to_pandas()
result.to_polars()
result.to_arrow()

11. Delete, restore, and compact

Deletes are soft deletes. Deleted IDs are excluded from search and can be restored until compaction.

collection.delete(["note-fr"])
print(collection.list_deleted_ids())

collection.restore(["note-fr"])
print(collection.list_deleted_ids())

collection.delete(["note-archive"])
removed = collection.compact()
print(removed)

12. Back up and close

collection.checkpoint()

db.snapshot_collection("documents", "./documents.snapshot")
db.export_collection("documents", "./documents-export")

collection.close()
client.close()

commit() is used for normal fast write boundaries. checkpoint() is the call to make before backup or controlled shutdown because it forces a durable on-disk checkpoint and clears WAL state.