llama-index-cockroachdb

LlamaIndex integration for CockroachDB's native VECTOR column type and the C-SPANN approximate nearest neighbor index, plus a full set of storage backends (KV, document, index, chat). Ships seven classes:

Vector / retrieval:

• CockroachDBVectorStore (llama_index.vector_stores.cockroachdb): drop-in BasePydanticVectorStore backed by CRDB.
• CockroachDBReader (llama_index.readers.cockroachdb): load existing CRDB rows as LlamaIndex Document objects.
• CockroachDBRetriever (llama_index.retrievers.cockroachdb): standalone retriever with C-SPANN beam-size tuning.

Storage backends (new in 0.2.0):

• CockroachDBKVStore (llama_index.storage.kvstore.cockroachdb): generic key-value store, sync + async.
• CockroachDBDocumentStore (llama_index.storage.docstore.cockroachdb): persist Document / Node objects.
• CockroachDBIndexStore (llama_index.storage.index_store.cockroachdb): persist IndexStruct objects.
• CockroachDBChatStore (llama_index.storage.chat_store.cockroachdb): persist ChatMessage history per session.

Requires CockroachDB v25.2+ and Python 3.10+.

Why not just use the pgvector store?

llama-index-vector-stores-postgres depends on the pgvector extension. CockroachDB ships its own native VECTOR(n) type and the C-SPANN distributed ANN index, so the pgvector store cannot be used as-is against a CRDB cluster: the DDL (CREATE EXTENSION vector, USING hnsw) and the per-session tuning vars (hnsw.ef_search, ivfflat.probes) don't exist there. This package targets CRDB's actual API surface.

Install

pip install llama-index-cockroachdb

You also need a CRDB v25.2+ cluster with the vector feature enabled once at the cluster level:

SET CLUSTER SETTING feature.vector_index.enabled = true;

The store will attempt to set this on first initialization if enable_feature_setting=True (the default) and the connected user has permission.

Quick start

from llama_index.vector_stores.cockroachdb import CockroachDBVectorStore
from llama_index.core import VectorStoreIndex, StorageContext, Document

store = CockroachDBVectorStore.from_params(
    host="localhost",
    port=26257,
    user="root",
    password="",
    database="defaultdb",
    table_name="my_index",
    embed_dim=1536,
    distance_metric="cosine",                # or "l2", "inner_product"
    cspann_kwargs={
        "min_partition_size": 16,
        "max_partition_size": 128,
    },
    sslmode="disable",                       # local dev only
)

index = VectorStoreIndex.from_documents(
    [Document(text="...")],
    storage_context=StorageContext.from_defaults(vector_store=store),
)
print(index.as_query_engine().query("..."))

See examples/ for seven runnable scripts covering async, MMR, metadata filters, the standalone retriever, and the reader.

Tuning C-SPANN

Two levers, both optional:

1. Build-time partitioning: pass cspann_kwargs={"min_partition_size": ..., "max_partition_size": ...} to from_params(). These map directly to the WITH clause on CREATE VECTOR INDEX.
2. Query-time beam size: pass vector_search_beam_size=N to the store constructor (applies to every query) or as a kwarg on individual query() calls. Higher = better recall, slightly more latency. Issued as SET LOCAL vector_search_beam_size per session.

store.query(query, vector_search_beam_size=128)

Reader

from llama_index.readers.cockroachdb import CockroachDBReader

reader = CockroachDBReader.from_params(
    host="localhost", port=26257, database="defaultdb", user="root", sslmode="disable",
)
docs = reader.load_data(
    table="articles",
    text_column="body",
    metadata_columns=["id", "author", "tag"],
    id_column="id",
)
# or pass a query=... with :named params

Retriever

from llama_index.retrievers.cockroachdb import CockroachDBRetriever
from llama_index.embeddings.openai import OpenAIEmbedding

retriever = CockroachDBRetriever(
    vector_store=store,
    embed_model=OpenAIEmbedding(model="text-embedding-3-small"),
    similarity_top_k=5,
    vector_search_beam_size=128,
)
nodes_with_scores = retriever.retrieve("What is C-SPANN?")

Storage backends

Each storage class uses the same cockroachdb+psycopg2 / cockroachdb+asyncpg dialects, so transactions retry automatically on SERIALIZATION_FAILURE.

from llama_index.storage.docstore.cockroachdb import CockroachDBDocumentStore
from llama_index.storage.index_store.cockroachdb import CockroachDBIndexStore
from llama_index.storage.chat_store.cockroachdb import CockroachDBChatStore
from llama_index.core import StorageContext

docstore = CockroachDBDocumentStore.from_params(
    host="localhost", port=26257, database="defaultdb",
    user="root", password="", sslmode="disable",
    table_name="docs",
)
index_store = CockroachDBIndexStore.from_params(
    host="localhost", port=26257, database="defaultdb",
    user="root", password="", sslmode="disable",
    table_name="idx",
)
chat_store = CockroachDBChatStore.from_params(
    host="localhost", port=26257, database="defaultdb",
    user="root", password="", sslmode="disable",
    table_name="chats",
)

storage_context = StorageContext.from_defaults(
    docstore=docstore,
    index_store=index_store,
    vector_store=store,
)

CockroachDBChatStore stores each session's messages as a single JSONB array (CRDB does not support ARRAY(JSON) as a column type). Atomic appends use JSONB concatenation (||).

Supported query modes

Supported metadata filter operators

EQ, NE, GT, GTE, LT, LTE, IN, NIN, CONTAINS, TEXT_MATCH, TEXT_MATCH_INSENSITIVE, IS_EMPTY. Filters are AND/OR/NOT nestable via MetadataFilters.

For frequently filtered keys, declare indexed_metadata_keys={("category", "text"), ("year", "int")} on the store to get JSONB-extracted BTREE indices.

Development

uv sync --all-extras
uv run pre-commit install
uv run pytest         # spins up a CRDB testcontainer per session
uv run ruff check .

Tests can target an existing cluster instead of a container by exporting CRDB_TEST_URL=postgresql://user:pass@host:port/db?sslmode=disable.

License

Apache 2.0