One database for your records, their relationships, and search over them.
Agents need memory. Apps need connected data. The standard answer involves multiple databases, schema design, and an embedding pipeline before you write a single useful line of business logic.
RushDB skips all of that. Push any JSON — nested structure becomes a traversable graph, string properties become semantically searchable, type inference happens automatically.
Works with LangChain, CrewAI, AutoGen, or any Python AI framework.
pip install rushdb| Package | RushDB 3.x and RushDB Cloud | RushDB 2.x (legacy) |
|---|---|---|
@rushdb/javascript-sdk |
3.x | 2.11.0: npm i @rushdb/javascript-sdk@2 |
rushdb (PyPI) |
3.x | 2.10.0: pip install "rushdb<3" |
@rushdb/mcp-server |
3.x | 2.13.0: npx @rushdb/mcp-server@2 |
The SDK works with RushDB in one of two ways:
| RushDB Cloud | Self-hosted RushDB engine (direct mode) | |
|---|---|---|
| What you run | nothing: the managed service at app.rushdb.com | one rushdb binary, on your own infrastructure |
| Connect with | an API key | the engine URL and the engine token |
The engine is what you can self-host. RushDB Cloud runs that engine for you and adds the dashboard, managed embeddings and the AI features (natural-language search, relationship-pattern and join suggestions), which exist only in RushDB Cloud.
from rushdb import RushDB
# RushDB Cloud
db = RushDB("RUSHDB_API_KEY")
# A self-hosted engine
db = RushDB(
"ENGINE_TOKEN",
direct=True,
engine_url="http://127.0.0.1:3000",
scope="tenant-1", # only for an engine started with --multitenant
)What each offers:
| Feature | Cloud | Self-hosted engine |
|---|---|---|
Records: create, read, update, set, upsert, delete, find with where, orderBy, select, groupBy |
✓ | ✓ |
Relationships: attach/detach, relationships.create_many/delete_many/find/update |
✓ | ✓ |
| Labels, properties, property values | ✓ | ✓ |
| Import JSON and CSV, export CSV | ✓ | ✓ |
External vector indexes; vector search with a queryVector you supply |
✓ | ✓ |
Raw openCypher (query.raw; the engine needs --languages cypher) |
✓ | ✓ |
Schema (ai.get_schema, ai.get_schema_markdown) |
✓ | ✓ |
Catalog lookups: ai.search_values (a value across every property, also near spellings), ai.get_schema_summary, ai.resolve_names, and value search in properties.values (query, match, domain_on_miss) |
✓ | ✓ |
Keyword (BM25) search (records.text_search), text indexes (text_indexes.*), hybrid search |
✓ | ✓ |
| Transactions | ✓ | record creates, updates, sets, deletes by id, attaches and imports are staged; other writes refuse a transaction |
Managed embeddings: managed indexes, records.vector_search by text query |
✓ | — |
Natural-language search on SQL (ai.search, ai.run_sql) |
✓ | — |
Relationship-pattern suggestions (relationships.patterns.*) and join suggestions (ai.get_join_candidates) |
✓ | — |
Background file imports (imports.*) and SaaS connections (connections.*) |
✓ | — |
Project settings (settings.get), dashboard, team access |
✓ | — |
In direct mode a Cloud-only method raises CloudOnlyError (with method and reason) before
sending anything, and no request ever goes to RushDB Cloud. A write the engine cannot stage into a
transaction raises DirectModeError, also before sending anything. The list is available in code:
from rushdb import CLOUD_ONLY_METHODS, CloudOnlyError, is_cloud_only_methoddb.engine.execute() runs any Protocol v1 operation directly.
Direct mode keeps its connection to the engine alive: each thread reuses one connection, so a call
does not pay for a new TCP (or TLS) handshake. A connection the engine closed is replaced; a read
that lost its connection is sent once more, a write is not (it raises RushDBError).
db.engine.close() closes the open connections. When HTTP_PROXY/HTTPS_PROXY applies to the
engine URL, calls go through the proxy without keep-alive.
Get an API key at app.rushdb.com.
from rushdb import RushDB
db = RushDB('RUSHDB_API_KEY')
# Store an agent action — graph links sessions and context automatically
db.records.create(
label='MEMORY',
data={
'agent_id': 'agent-42',
'session_id': 'sess-001',
'action': 'summarized',
'topic': 'Q4 results',
'output': summary_text,
},
)
# Recall — traverse relationships, filter by properties
results = db.records.find({
'labels': ['MEMORY'],
'where': {
'agent_id': 'agent-42',
'topic': {'$contains': 'Q4'},
},
'limit': 10,
})
for memory in results:
print(memory.get('output'))# Push nested JSON — relationships created automatically
db.records.create_many('COMPANY', {
'name': 'Acme Corp',
'DEPARTMENT': [{
'name': 'Engineering',
'EMPLOYEE': [{
'name': 'Alice',
'role': 'Staff Engineer',
}]
}]
})
# Traverse the auto-created graph
engineers = db.records.find({
'labels': ['EMPLOYEE'],
'where': {
'role': {'$contains': 'Engineer'},
'DEPARTMENT': {'COMPANY': {'name': 'Acme Corp'}},
},
})
# Constrain by relationship type and direction
authored_posts = db.records.find({
'labels': ['USER'],
'where': {
'POST': {
'$relation': {'type': 'AUTHORED', 'direction': 'out'},
'title': {'$contains': 'graph'},
}
},
'limit': 10,
})
# Multi-hop: add hops to $relation — everyone in Alice's reporting chain, up to 4 levels
chain = db.records.find({
'labels': ['EMPLOYEE'],
'where': {
'EMPLOYEE': {
'$relation': {'type': 'REPORTS_TO', 'direction': 'out', 'hops': {'min': 1, 'max': 4}},
'name': {'$contains': 'Alice'},
}
},
})
# Cycle detection: accounts on a circular transfer ring (fraud rings, circular ownership)
ring_members = db.records.find({
'labels': ['ACCOUNT'],
'where': {
'RING': { # key is a display name — the $cycle block holds only $relation
'$cycle': True,
'$relation': {'type': 'TRANSFERRED_TO', 'direction': 'out', 'hops': {'min': 2, 'max': 6}},
}
},
})
# Manage relationships explicitly
user = db.records.find_uniq({'labels': ['USER'], 'where': {'name': 'Alice'}})
company = db.records.find_uniq({'labels': ['COMPANY'], 'where': {'name': 'Acme Corp'}})
user.attach(
target=company,
options={'type': 'WORKS_AT', 'direction': 'out', 'properties': {'source': 'profile'}},
)
# Relationship search: where filters edge type/properties, source/target filter endpoint records
relationships = db.relationships.find({
'source': {'labels': ['USER'], 'where': {'name': 'Alice'}},
'target': {'labels': ['COMPANY']},
'where': {'type': 'WORKS_AT', 'source': 'profile'},
})
# Change properties on existing relationships (RushDB 3.0.4+): set writes, unset removes
db.relationships.update(user, company, type='WORKS_AT', set={'source': 'hr'}, unset=['legacy'])
# types marks values JSON can't type: 'datetime' stores an ISO string as a datetime
db.relationships.update(
user, company, type='WORKS_AT', set={'since': '2024-03-01'}, types={'since': 'datetime'}
)csv_data = "name,email,age\nJohn,john@example.com,30\nJane,jane@example.com,25"
db.records.import_csv(
label='USER',
data=csv_data,
# skipEmptyValues: treat empty cells ("" / []) as unset instead of storing them (0/False are kept)
options={'returnResult': True, 'suggestTypes': True, 'skipEmptyValues': True},
parse_config={'header': True, 'skipEmptyLines': True, 'dynamicTyping': True},
)Large files and SaaS data load in the background: the client hands the data over and the run
continues on the server. Every file import and every connection sync is an import run; follow it
with db.imports.get(id) or db.imports.wait(id) until its status is completed, partial,
failed or cancelled.
# One call: create → upload every file → start (→ wait)
run = db.imports.import_files(
[
{'data': 'customers.csv', 'label': 'Customer'}, # a path: name and size come from it
open('orders.csv.gz', 'rb'), # a binary file object works too
],
mappings={0: {'keyColumn': 'id'}}, # re-importing upserts by `id`
links=[{'source': 1, 'sourceColumn': 'customer_id', 'target': 0, 'targetColumn': 'id', 'type': 'PLACED_BY'}],
parallelism=4,
wait=True,
wait_interval=2,
on_progress=lambda p: print(f"{p['loaded']}/{p['total']}" if p['phase'] == 'upload' else p['run']['status']),
)
print(run.data['recordsWritten'], run.data['summary'])Or step by step, to show a preview before starting:
import os
created = db.imports.create([{'name': 'orders.csv', 'size': os.path.getsize('orders.csv')}]).data
db.imports.upload(created['id'], created['files'][0], 'orders.csv', parallelism=4, retries=3)
preview = db.imports.preview(created['id']).data # column types, suggested key and links, personal data
columns = preview['files'][0]['columns'] # each has `suggestion` ('include' | 'skip' | 'hash') and `pii` when it looks personal
mapping = {
'skip': [c['name'] for c in columns if c['suggestion'] == 'skip'],
'hash': [c['name'] for c in columns if c['suggestion'] == 'hash'], # keyed hash: joins, not readable
}
if preview['files'][0].get('suggestedKey'):
mapping['keyColumn'] = preview['files'][0]['suggestedKey']
db.imports.start(created['id'], files=[{'id': created['files'][0]['id'], 'mapping': mapping}])
finished = db.imports.wait(created['id'], timeout=30 * 60).data # TimeoutError after 30 minutesupload follows the upload mode the server returns: s3-multipart sends parts to their presigned
URLs (parallelism at a time on a thread pool, default 4; each failed part retried, retries
default 3) and completes with their ETags; direct sends the bytes through the API. Upload data can
be a file path (str or Path), bytes/bytearray/memoryview, or a binary file object (read
from its current position). A non-seekable stream (a pipe, a socket) needs its size up front in
import_files, and a direct upload of one is not retried.
Also: imports.list(kind=, connection_id=, limit=, skip=), cancel, retry, delete,
add_files, remove_file.
import os
catalog = db.connections.catalog().data # connectors with their streams and config fields
created = db.connections.create({
'connectorId': 'hubspot',
'auth': {'type': 'bearer', 'token': os.environ['HUBSPOT_TOKEN']},
'mode': 'scheduled',
'scheduleMinutes': 60,
'streams': [{'stream': 'contacts', 'policy': {'mode': 'auto', 'exclude': ['mobilephone']}}],
}).data
connection_id = created['connection']['id']
db.connections.test(connection_id)
db.connections.preview(connection_id, 'contacts') # first records, nothing written
# Field review: profiles with personal-data findings (masked examples), the policy, unreviewed fields
review = db.connections.fields(connection_id, 'contacts', limit=100).data
db.connections.review_fields(
connection_id,
'contacts',
[{'name': f['name'], 'action': f['suggestion']} for f in review['fields']], # or your own choices
)
sync_run = db.connections.sync(connection_id).data # or sync(connection_id, streams=['contacts'], full=True)
db.imports.wait(sync_run['id'])A stream's policy decides which fields reach the graph: auto keeps everything except exclude
and hash; review_fields switches it to an allowlist of the fields marked include or hash,
and fields that appear later wait in the stream's unreviewed list until the next review.
On-prem databases sync through rushdb-agent: create an agent connection and hand the token
(shown once) to the agent.
agent = db.connections.create({'connectorId': 'agent', 'name': 'Warehouse', 'agent': {'intervalMinutes': 15}}).data
print(agent['agentToken']) # connection['agent'] shows what the agent last reportedAlso: list, get, update, delete(id, delete_data=True), runs(id, limit=, skip=), pause,
resume, rotate_webhook_token (a new webhook URL, or a new agent token), push_templates()
(Zapier, Make and n8n recipes for push connections). Credentials are sealed on the server and never
returned.
db.records.find() returns a SearchResult — a list-like container with pagination metadata.
result = db.records.find({
'where': {'status': 'active'},
'limit': 10,
'skip': 0,
})
# List-like usage
print(f"Loaded {len(result)} of {result.total} total")
print(f"Has more: {result.has_more}")
for record in result:
print(record.get('name'))
# Indexing and slicing
first = result[0]
top_five = result[:5]
# Boolean check
if result:
process(result[0])| Property | Type | Description |
|---|---|---|
data |
List[Record] |
The result items |
total |
int |
Total matching records in the database |
has_more |
bool |
Whether more records exist beyond this page |
search_query |
dict |
The query that produced this result |
sql |
str |
For Smart search, the SQL statement it ran |
Use db.records.vector_search() for direct semantic/vector retrieval over an
embedding index:
results = db.records.vector_search({
'labels': ['MEMORY'],
'propertyName': 'content',
'query': 'how agents remember things',
'where': {'agent_id': 'agent-42'},
'limit': 5,
})
for record in results:
print(record.score, record.get('content'))Use db.ai.search() when you want RushDB to turn a natural-language request
into one read-only SQL statement over your schema and run it (RushDB Cloud):
results = db.ai.search('Find active memories about Q4 results for agent-42')
print(results.name) # a short title, e.g. to save the query under
print(results.sql) # the statement it ran
more = results.next() # the next page of the statement's rows
# Run a statement directly, e.g. a saved one:
rows = db.ai.run_sql('SELECT agent_id, count(*) AS memories FROM "MEMORY" GROUP BY agent_id')Rows that select a record's __id come back as full records; aggregates come
back as plain rows.
db.ai.search({...}) still works as a deprecated vector-search alias, but new
code should use db.records.vector_search({...}).
Declare a BM25 text index, then search it by keywords:
db.text_indexes.create({'label': 'MEMORY', 'property': 'content'})
results = db.records.text_search({
'query': 'quarterly results',
'labels': ['MEMORY'],
'where': {'agent_id': 'agent-42'},
'limit': 5,
})
print(results.total, results.meta['textSearch'])Add a text leg to a vector search to fuse keyword and vector ranking. fusion
is rrf (default) or dbsf; weight and prefetchLimit tune each leg:
results = db.records.vector_search({
'labels': ['MEMORY'],
'propertyName': 'content',
'queryVector': embedding,
'text': {'query': 'quarterly results', 'weight': 0.5},
'fusion': 'dbsf',
'limit': 5,
})Map the words of a question to what the data actually stores before building a query. Each is one read of the live catalog:
# Where is a value stored? Every property, with label, property and record count
found = db.ai.search_values('spiderman', match='similar') # contains | prefix | exact | similar
for m in found.data['matches']:
print(m['label'], m['property'], m['value'], m['count'])
# Every label with its properties, distinct counts, ranges and the values of small fields
summary = db.ai.get_schema_summary(labels=['Patient'], few_values=10)
# The real label or property behind a name that is not in the schema
db.ai.resolve_names('patients.birthdate') # -> Patient.Birthday
# Value search within one property: ranked by record count, with counts;
# when nothing matches, list the whole domain of a small property
db.properties.values(property_id, query='positive', domain_on_miss=20)
# How labels reference each other, with evidence (RushDB Cloud only)
db.ai.get_join_candidates(labels=['Order'], min_confidence=0.5)user = db.records.create('USER', {
'name': 'Alice',
'email': 'alice@example.com',
})
# Safe field access
name = user.get('name') # 'Alice'
phone = user.get('phone', 'N/A') # 'N/A'
# Clean data (excludes internal __id, __label fields)
data = user.get_data() # {'name': 'Alice', 'email': '...'}
full = user.get_data(exclude_internal=False) # includes __id, __label, etc.
# Existence check (no exception if record was deleted)
if user.exists:
user.update({'status': 'active'})
# String representations
repr(user) # Record(id='abc-123', label='USER')
str(user) # USER: Alicewith db.transactions.begin() as tx:
record_a = db.records.create('NODE', {'value': 1}, transaction=tx)
record_b = db.records.create('NODE', {'value': 2}, transaction=tx)
record_a.attach(target=record_b, options={'type': 'LINKED'}, transaction=tx)
# auto-committed on exit, rolled back on exceptionfrom rushdb import RushDB
db = RushDB(
'RUSHDB_API_KEY',
url='https://api.rushdb.com/api/v1', # a different RushDB Cloud API endpoint; this is the default
timeout=30,
)docs.rushdb.com/3.x/connect/sdks and the Python reference — full API reference, vector search, aggregations, and more.
- GitHub Issues — bug reports and feature requests
- Email — direct support
See CONTRIBUTING.md. Issues and PRs welcome.