FlowQuery (Python)

A pure Python implementation of FlowQuery, a declarative OpenCypher-based query language for virtual graphs and data processing pipelines. This package has full functional fidelity with the TypeScript version.

Installation

pip install flowquery

Quick Start

Command Line Interface

Start the interactive REPL:

flowquery

Programmatic Usage

import asyncio
from flowquery import Runner

runner = Runner("WITH 1 as x RETURN x + 1 as result")
asyncio.run(runner.run())
print(runner.results)  # [{'result': 2}]

In Jupyter notebooks, you can use await directly:

from flowquery import Runner

runner = Runner("WITH 1 as x RETURN x + 1 as result")
await runner.run()
print(runner.results)  # [{'result': 2}]

Statement Info: Labels, Properties, and Source Lineage

The Runner exposes a metadata property that mirrors the TypeScript implementation. It reports counts of virtual nodes and relationships created/deleted plus an optional info: StatementInfo describing the structure the query touches - independent of execution.

StatementInfo captures:

• The node labels and relationship types referenced.
• The data sources backing the underlying virtual definitions.
• The node/relationship properties consumed by the query - alias.prop accesses anywhere in MATCH, WHERE, WITH, RETURN, ORDER BY, or function arguments, plus inline pattern properties like (u:User {id: 'rick.o'}).
• The properties declared by each virtual's RETURN clause via info.declared, so you can validate that a query references only declared properties.
• Literal values supplied for properties at the call site via info.nodes[label].literal_values - collected from inline pattern properties and from equality / IN predicates such as WHERE u.id = 'rick.o' or WHERE u.id IN ['a', 'b'].

The per-entity nodes and relationships maps give end-to-end lineage from a property to its data source:

from flowquery import Runner

runner = Runner("""
    CREATE VIRTUAL (:City) AS {
        LOAD JSON FROM "https://example.com/cities" AS c
        RETURN c.id AS id, c.name AS name, c.country AS country
    };
    CREATE VIRTUAL (:City)-[:FLIGHT]-(:City) AS {
        LOAD JSON FROM "https://example.com/flights" AS f
        RETURN f.left_id AS left_id, f.right_id AS right_id, f.airline AS airline
    };
    MATCH (a:City {name: 'NYC'})-[r:FLIGHT]->(b:City)
    WHERE b.country IN ['US', 'CA']
    RETURN a.name AS origin, b.name AS destination, r.airline AS airline
""")
info = runner.metadata.info

print(info.nodes)
# {'City': NodeInfo(
#     properties=['country', 'name'],
#     sources=['https://example.com/cities'],
#     literal_values={'country': ['US', 'CA'], 'name': ['NYC']},
# )}
print(info.relationships)
# {'FLIGHT': RelationshipInfo(
#     properties=['airline'],
#     sources=['https://example.com/flights'],
#     literal_values={},
# )}
print(info.declared.nodes['City'])
# DeclaredEntityInfo(
#     properties=['country', 'id', 'name'],
#     sources=['https://example.com/cities'],
# )
print(info.sources)
# ['https://example.com/cities', 'https://example.com/flights']

StatementInfo resolves sources and declared schemas for any virtual the query touches - both inline CREATE VIRTUAL clauses and previously-registered virtuals reached via MATCH or DELETE. The flat node_labels, relationship_types, sources, node_properties, and relationship_properties fields stay in sync with the per-entity nodes and relationships maps. Only purely literal AST subtrees end up in literal_values - values that depend on parameters, references, f-strings, or subqueries are skipped.

The same StatementInfoCrawler can be used directly on any parsed AST without going through a Runner:

from flowquery import StatementInfoCrawler
crawler = StatementInfoCrawler()
info = crawler.crawl(parsed_ast)

Lineage and Provenance

FlowQuery exposes two complementary lineage APIs that combine to trace every result cell back to the records, properties, and data sources that produced it.

Statement Info (always on)

Runner.metadata.info returns a StatementInfo snapshot derived from the parsed AST. It lists the labels and types touched by the statement, the properties read off each, the upstream data sources (URLs, file URIs, async function names, let:// references), any literal property filters present in WHERE, and per-output-column lineage in info.returns:

from flowquery import Runner

runner = Runner("""
    MATCH (c:City)
    WHERE c.country = 'US'
    RETURN c.name AS origin, c.country AS region
""")
info = runner.metadata.info
print(info.node_labels)              # ['City']
print(info.node_properties)          # {'City': ['country', 'name']}
print(info.nodes['City'].literal_values)  # {'country': ['US']}
print(info.returns['origin'].kind)   # 'property'
print(info.returns['origin'].references[0].alias)     # 'c'
print(info.returns['origin'].references[0].property)  # 'name'

info.returns is keyed by output column name and describes how each column was computed (literal, property, expression, or aggregate) plus every alias.property access that fed it.

Row-Level Provenance (opt in)

Pass RunnerOptions(provenance=True) to capture per-row lineage in runner.provenance, aligned by index with runner.results:

from flowquery import Runner, RunnerOptions

runner = Runner(
    "MATCH (a:City {name: 'New York'})-[r:FLIGHT]->(b:City) "
    "RETURN a.name AS origin, b.name AS destination",
    options=RunnerOptions(provenance=True),
)
await runner.run()

for row, prov in zip(runner.results, runner.provenance):
    print(row, "<-", [n.id for n in prov.nodes])
    for rel in prov.relationships:
        print(" ", rel.type, rel.path)  # full node-id path

Each RowProvenance carries the NodeBindings and RelationshipBindings active when the row was projected. Hops carry the matched relationship properties; nodes carry the matched node properties. For aggregate rows, prov.rows lists one RowSegment per contributing input row so collect()-style outputs align positionally with their lineage.

trace_row(i) and lineage()

Runner.trace_row(i) and Runner.lineage() bundle the structural column lineage with the runtime row provenance:

runner = Runner(
    "MATCH (c:City) WHERE c.country = 'US' RETURN c.name AS origin",
    options=RunnerOptions(provenance=True),
)
await runner.run()

trace = runner.trace_row(0)
print(trace['origin'].value)                          # 'New York'
print(trace['origin'].lineage.kind)                   # 'property'
print(trace['origin'].bindings[0].node.id)            # 'nyc'
print(trace['origin'].bindings[0].value)              # 'New York'

report = runner.lineage()
# report.columns -> column lineage map (deep copy of info.returns)
# report.rows    -> per-row {column: CellTrace} maps

For aggregate columns (e.g. collect(c.id)), bindings lists one entry per contributing input row, paired with the matched value at that row.

Threading Lineage Through Virtual Sub-Queries

When a MATCH reads from a CREATE VIRTUAL definition, the inner sub-query is itself executed with provenance, and the resulting RowProvenance is threaded onto the outer binding's source field:

await Runner("""
    CREATE VIRTUAL (:DerivedCity) AS {
        MATCH (s:SrcCity) WHERE s.country = 'US' RETURN s.id AS id
    }
""").run()

runner = Runner(
    "MATCH (d:DerivedCity) RETURN d.id AS id",
    options=RunnerOptions(provenance=True),
)
await runner.run()

d = runner.provenance[0].nodes[0]    # outer NodeBinding
inner = d.source                      # inner RowProvenance
print(inner.nodes[0].alias, inner.nodes[0].id)  # 's', 'nyc'

Recursion is unbounded: a virtual built atop another virtual produces nested source chains all the way down.

Data Sources and LET Chaining

LOAD JSON FROM <url> operations add a DataSourceBinding to the row's provenance segment. When the LOAD reads from a LET-bound sub-query, the binding's source is let://<name> and source_provenance carries the inner LET provenance, which in turn lists its own data_sources:

await Runner('LET cities = { LOAD JSON FROM "file://..." AS c '
             'RETURN c.id AS id, c.name AS name }').run()
await Runner("""
    CREATE VIRTUAL (:Mirror) AS {
        LOAD JSON FROM cities AS c RETURN c.id AS id, c.name AS name
    }
""").run()

runner = Runner(
    "MATCH (m:Mirror) RETURN m.name AS name",
    options=RunnerOptions(provenance=True),
)
await runner.run()

inner = runner.provenance[0].nodes[0].source
let_ds = next(d for d in inner.data_sources
              if d.source == 'let://cities')
file_ds = next(d for d in let_ds.source_provenance.data_sources
               if d.source.startswith('file://'))

The same chain is reflected statically in info.sources so callers can introspect the upstream dependency graph without running the query.

Backend-Aware Virtual Sources

By default a virtual definition is loaded eagerly: its sub-query runs once per query and all rows are held in memory. For large or remote backends, a source provider lets FlowQuery read only what a pattern needs. This is opt-in: existing CREATE VIRTUAL definitions behave exactly as before.

The simplest way to write a provider is to subclass NodeSource or RelationshipSource and implement scan(request, context). Implement lookup (nodes by id) or expand (relationships by left_id/right_id) when the backend can fetch selectively; the base class then declares the matching capability for you. version defaults to the current contract version, and capability keys may be written in snake_case (max_batch_size, string_ordering). A method may return a list of rows, an awaitable list, or an (async) iterable of pages.

The request is a SourceRequest: a dict with the same camelCase keys as the TypeScript implementation (request["pageSize"]) that also accepts snake_case keys (request["page_size"]) and attribute access (request.ids, request.side, request.predicate, request.page_size).

from flowquery import FlowQuery, NodeSource, RelationshipSource, cursor_pages

class People(NodeSource):
    capabilities = {
        "cacheable": True,       # identical reads within one query may be cached
        "max_batch_size": 50,    # ids per lookup/expand request
        "predicates": {"operators": ["eq", "in", "gt", "gte", "lt", "lte"]},
    }

    def scan(self, request, context):
        async def fetch(cursor):
            page = await api.people(filter=request.predicate, cursor=cursor)
            return {"rows": page.rows, "cursor": page.next_cursor}
        return cursor_pages(fetch)

    async def lookup(self, request, context):  # declares "lookup": True
        return await api.people_by_ids(request.ids)

class Knows(RelationshipSource):
    async def scan(self, request, context):
        return await api.all_knows()

    async def expand(self, request, context):  # declares both expand sides
        key = "from" if request.side == "left" else "to"
        return await api.knows(**{key: request.ids})

FlowQuery.register_node_source("Person", People())
FlowQuery.register_relationship_source("KNOWS", "Person", "Person", Knows())

fq = FlowQuery("MATCH (a:Person {id: 7})-[:KNOWS]->(b:Person) RETURN b.name LIMIT 10")
await fq.run()
print(fq.source_stats)  # calls, pages, rows, cacheHits, batches, plans...

Subclassing is optional: any object with a capabilities dict and a scan method works, and SourceProvider is a typing.Protocol describing that shape.

The engine then loads correlated rows in batches instead of reading every person and relationship:

• Correlated loading and batching. It issues id lookups, expands per page of anchor nodes, and makes batched lookups for target nodes.
• Selective start points. A chain may start from its more selective end. The decision is recorded in source_stats["plans"], and it can change result order when there is no ORDER BY.
• Predicate push-down. Pattern properties and AND-ed WHERE comparisons (=, <, <=, >, >=, IN) against literals or parameters are sent as request["predicate"].
• Streaming with early stop under LIMIT. Reading stops once the limit is reached and the provider's iterator is closed.
• Query-scoped caching.
• Cancellation. Call runner.cancel() or pass SourceOptions(signal=...) (an asyncio.Event or SourceSignal). The query raises QueryCancelledError.
• Read contexts. Optional open_read_context() and close_read_context(snapshot) methods are called once per query.

Providers may return a superset, because FlowQuery re-checks every row. matches_predicate(predicate, row) is a reference implementation of the predicate semantics.

Errors are reported through typed exceptions:

• SourceCapabilityError: an invalid declaration, raised at registration.
• SourceResponseError: malformed pages or a repeated cursor.
• SourceError: any other provider failure.

A provider may raise SourceUnsupportedError from lookup/expand to fall back to a scan.

Query-level tuning goes in RunnerOptions(sources=SourceOptions(...)): page_size (256), max_batch_size (100), cache_capacity (10000), max_prefetch_rows (1024) and reorder (True).

In-memory sources. in_memory_node_source(rows) and in_memory_relationship_source(rows) are ready-made providers over lists. They support lookup, expansion from both sides, every predicate operator and paging, which makes them handy for tests, examples and small reference data. Keyword options are cacheable (default True) and max_batch_size.

from flowquery import in_memory_node_source, in_memory_relationship_source

FlowQuery.register_node_source(
    "Person", in_memory_node_source([{"id": 1, "name": "Ada"}, {"id": 2, "name": "Bob"}])
)
FlowQuery.register_relationship_source(
    "KNOWS", "Person", "Person", in_memory_relationship_source([{"left_id": 1, "right_id": 2}])
)

Statement-backed sources. A CREATE VIRTUAL definition can opt into the same contract declaratively with a WITH CAPABILITIES { ... } clause between the pattern and AS. Each request runs its sub-query with $request bound to the request (operation, label/type, ids, side, predicate, pageSize):

await FlowQuery("""
    CREATE VIRTUAL (:Ticket) WITH CAPABILITIES {lookup: true, streaming: true} AS {
        LOAD JSON FROM f"https://api.example.com/tickets?ids={join(coalesce($request.ids, []), ',')}" AS t
        RETURN t.id AS id, t.title AS title
    }
""").run()

The map takes the same (camelCase) keys as the capabilities dict; version defaults to the current contract version. Invalid capabilities reject the whole definition. Redefining the label without WITH CAPABILITIES, or deleting it, restores eager loading. The same opt-in is available programmatically with FlowQuery.set_node_source_capabilities(label, capabilities) and FlowQuery.set_relationship_source_capabilities(...); pass None to opt out.

Documentation

• Language Reference (clauses, expressions, functions, graph operations, and more)
• Quick Cheat Sheet
• Full Documentation
• Contributing Guide
• Virtual Graph Demo Notebook - a demo of virtual graph capabilities and custom function extensibility

Extending FlowQuery with Custom Functions

The query language itself is identical between the TypeScript and Python versions. The only difference is that custom functions are written in Python here instead of TypeScript.

Creating a Custom Scalar Function

Scalar functions operate on individual values and return a result:

from flowquery.extensibility import Function, FunctionDef

@FunctionDef({
    "description": "Doubles a number",
    "category": "scalar",
    "parameters": [{"name": "value", "description": "Number to double", "type": "number"}],
    "output": {"description": "Doubled value", "type": "number"},
})
class Double(Function):
    def __init__(self):
        super().__init__("double")
        self._expected_parameter_count = 1

    def value(self):
        return self.get_children()[0].value() * 2

Once defined, use it in your queries:

WITH 5 AS num RETURN double(num) AS result
// Returns: [{"result": 10}]

Creating a Custom String Function

from flowquery.extensibility import Function, FunctionDef

@FunctionDef({
    "description": "Reverses a string",
    "category": "scalar",
    "parameters": [{"name": "text", "description": "String to reverse", "type": "string"}],
    "output": {"description": "Reversed string", "type": "string"},
})
class StrReverse(Function):
    def __init__(self):
        super().__init__("strreverse")
        self._expected_parameter_count = 1

    def value(self) -> str:
        return str(self.get_children()[0].value())[::-1]

Usage:

WITH 'hello' AS s RETURN strreverse(s) AS reversed
// Returns: [{"reversed": "olleh"}]

Creating a Custom Aggregate Function

Aggregate functions process multiple values and return a single result. They require a ReducerElement to track state:

from flowquery.extensibility import AggregateFunction, FunctionDef, ReducerElement

class MinReducerElement(ReducerElement):
    def __init__(self):
        self._value = None

    @property
    def value(self):
        return self._value

    @value.setter
    def value(self, val):
        self._value = val

@FunctionDef({
    "description": "Collects the minimum value",
    "category": "aggregate",
    "parameters": [{"name": "value", "description": "Value to compare", "type": "number"}],
    "output": {"description": "Minimum value", "type": "number"},
})
class MinValue(AggregateFunction):
    def __init__(self):
        super().__init__("minvalue")
        self._expected_parameter_count = 1

    def reduce(self, element):
        current = self.first_child().value()
        if element.value is None or current < element.value:
            element.value = current

    def element(self):
        return MinReducerElement()

Usage:

UNWIND [5, 2, 8, 1, 9] AS num RETURN minvalue(num) AS min
// Returns: [{"min": 1}]

Creating a Custom Async Data Provider

Async providers allow you to create custom data sources that can be used with LOAD JSON FROM:

from flowquery.extensibility import AsyncFunction, FunctionDef

@FunctionDef({
    "description": "Provides example data for testing",
    "category": "async",
    "parameters": [],
    "output": {"description": "Example data object", "type": "object"},
})
class GetExampleData(AsyncFunction):
    def __init__(self):
        super().__init__("getexampledata")
        self._expected_parameter_count = 0

    async def generate(self):
        yield {"id": 1, "name": "Alice"}
        yield {"id": 2, "name": "Bob"}

Usage:

LOAD JSON FROM getexampledata() AS data RETURN data.id AS id, data.name AS name
// Returns: [{"id": 1, "name": "Alice"}, {"id": 2, "name": "Bob"}]

Using Custom Functions with Expressions

Custom functions integrate seamlessly with FlowQuery expressions and can be combined with other functions:

// Using custom function with expressions
WITH 5 * 3 AS num RETURN addhundred(num) + 1 AS result

// Using multiple custom functions together
WITH 2 AS num RETURN triple(num) AS tripled, square(num) AS squared

Introspecting Registered Functions

You can use the built-in functions() function to discover registered functions including your custom ones:

WITH functions() AS funcs
UNWIND funcs AS f
WITH f WHERE f.name = 'double'
RETURN f.name AS name, f.description AS description, f.category AS category

License

MIT License - see LICENSE for details.

Links

• Homepage
• Repository
• Issues