diff --git a/docs/tech-specs/kafka-backend.md b/docs/tech-specs/kafka-backend.md
new file mode 100644
index 00000000..178ddfe4
--- /dev/null
+++ b/docs/tech-specs/kafka-backend.md
@@ -0,0 +1,200 @@
+---
+layout: default
+title: "Kafka Pub/Sub Backend Technical Specification"
+parent: "Tech Specs"
+---
+
+# Kafka Pub/Sub Backend Technical Specification
+
+## Overview
+
+Add Apache Kafka as a third pub/sub backend alongside Pulsar and RabbitMQ.
+Kafka's topic model maps naturally to TrustGraph's pub/sub abstraction:
+topics are first-class, consumer groups provide competing-consumer
+semantics, and the AdminClient handles topic lifecycle.
+
+## Problem
+
+TrustGraph currently supports Pulsar and RabbitMQ. Kafka is widely
+deployed and operationally familiar to many teams. Its log-based
+architecture provides durable, replayable message streams with
+well-understood scaling properties.
+
+## Design
+
+### Concept Mapping
+
+| TrustGraph concept | Kafka equivalent |
+|---|---|
+| Topic (`class:topicspace:topic`) | Kafka topic (named `topicspace.class.topic`) |
+| Subscription (competing consumers) | Consumer group |
+| `create_topic` / `delete_topic` | `AdminClient.create_topics()` / `delete_topics()` |
+| `ensure_topic` | `AdminClient.create_topics()` (idempotent) |
+| Producer | `KafkaProducer` |
+| Consumer | `KafkaConsumer` in a consumer group |
+| Message acknowledge | Commit offset |
+| Message negative acknowledge | Seek back to message offset |
+
+### Topic Naming
+
+The topic name follows the same convention as the RabbitMQ exchange
+name:
+
+```
+class:topicspace:topic  ->  topicspace.class.topic
+```
+
+Examples:
+- `flow:tg:text-completion-request` -> `tg.flow.text-completion-request`
+- `request:tg:librarian` -> `tg.request.librarian`
+- `response:tg:config` -> `tg.response.config`
+
+### Topic Classes and Retention
+
+Kafka topics are always durable (log-based). The class prefix determines
+retention policy rather than durability:
+
+| Class | Retention | Partitions | Notes |
+|---|---|---|---|
+| `flow` | Long or infinite | 1 | Data pipeline, order preserved |
+| `request` | Short (e.g. 300s) | 1 | RPC requests, ephemeral |
+| `response` | Short (e.g. 300s) | 1 | RPC responses, shared (see below) |
+| `notify` | Short (e.g. 300s) | 1 | Broadcast signals |
+
+Single partition per topic preserves message ordering and makes
+offset-based acknowledgment equivalent to per-message ack. This matches
+the current `prefetch_count=1` model used across all backends.
+
+### Producers
+
+Straightforward `KafkaProducer` wrapping. Messages are serialised as
+JSON (consistent with the RabbitMQ backend). Message properties/headers
+map to Kafka record headers.
+
+### Consumers
+
+#### Flow and Request Class (Competing Consumers)
+
+Consumer group ID = subscription name. Multiple consumers in the same
+group share the workload (Kafka's native consumer group rebalancing).
+
+```
+group_id = subscription  # e.g. "triples-store--default--input"
+```
+
+#### Response and Notify Class (Per-Subscriber)
+
+This is where Kafka differs from RabbitMQ. Kafka has no anonymous
+exclusive auto-delete queues.
+
+Design: use a **shared response topic with unique consumer groups**.
+Each subscriber gets its own consumer group (using the existing
+UUID-based subscription name from `RequestResponseSpec`). Every
+subscriber reads all messages from the topic and filters by correlation
+ID, discarding non-matching messages.
+
+This is slightly wasteful — N subscribers each read every response — but
+request/response traffic is low-volume compared to the data pipeline.
+The alternative (per-instance temporary topics) would require dynamic
+topic creation/deletion for every API gateway request, which is
+expensive in Kafka (AdminClient operations involve controller
+coordination).
+
+### Acknowledgment
+
+#### Acknowledge (Success)
+
+Commit the message's offset. With a single partition and sequential
+processing, this is equivalent to per-message ack:
+
+```
+consumer.commit(offsets={partition: offset + 1})
+```
+
+#### Negative Acknowledge (Failure / Retry)
+
+Kafka has no native nack-with-redelivery. On processing failure, seek
+the consumer back to the failed message's offset:
+
+```
+consumer.seek(partition, offset)
+```
+
+The message is redelivered on the next poll. This matches the current
+RabbitMQ `basic_nack(requeue=True)` behaviour: the message is retried
+by the same consumer.
+
+### Topic Lifecycle
+
+The flow service creates and deletes topics via the Kafka AdminClient:
+
+- **Flow start**: `AdminClient.create_topics()` for each unique topic
+  in the blueprint. Topic config includes `retention.ms` based on class.
+- **Flow stop**: `AdminClient.delete_topics()` for the flow's topics.
+- **Service startup**: `ensure_topic` creates the topic if it doesn't
+  exist (idempotent via `create_topics` with `validate_only=False`).
+
+Unlike RabbitMQ where consumers declare their own queues, Kafka topics
+must exist before consumers connect. The flow service and service
+startup `ensure_topic` calls handle this.
+
+### Message Encoding
+
+JSON body, consistent with the RabbitMQ backend. Serialisation uses the
+existing `dataclass_to_dict` / `dict_to_dataclass` helpers. Message
+properties map to Kafka record headers (byte-encoded string values).
+
+### Configuration
+
+New CLI arguments following the existing pattern:
+
+```
+--pubsub-backend kafka
+--kafka-bootstrap-servers localhost:9092
+--kafka-security-protocol PLAINTEXT
+--kafka-sasl-mechanism (optional)
+--kafka-sasl-username (optional)
+--kafka-sasl-password (optional)
+```
+
+The factory in `pubsub.py` creates a `KafkaBackend` instance when
+`pubsub_backend='kafka'`.
+
+### Dependencies
+
+`kafka-python-ng` or `confluent-kafka`. The `confluent-kafka` package
+provides both producer/consumer and AdminClient in one library with
+better performance (C-backed librdkafka), but requires a C extension
+build. `kafka-python-ng` is pure Python, simpler to install.
+
+## Key Design Decisions
+
+1. **Shared response topic with filtering** over per-instance temporary
+   topics. Avoids expensive dynamic topic creation for every RPC
+   exchange. Acceptable because response traffic is low-volume.
+
+2. **Seek-back for negative acknowledge** over not-committing or retry
+   topics. Provides immediate redelivery consistent with the RabbitMQ
+   nack behaviour.
+
+3. **Single partition per topic** to preserve ordering and simplify
+   offset management. Parallelism comes from multiple topics and
+   multiple services, not from partitioning within a topic.
+
+4. **Retention-based class semantics** instead of durability flags.
+   Kafka topics are always durable; short retention achieves the
+   ephemeral behaviour needed for request/response/notify classes.
+
+## Open Questions
+
+- **Retention values**: exact `retention.ms` for short-lived topic
+  classes. 300s (5 minutes) is a starting point; may need tuning based
+  on worst-case restart/reconnect times.
+
+- **Library choice**: `confluent-kafka` vs `kafka-python-ng`. Performance
+  vs install simplicity trade-off. Could support both behind a thin
+  wrapper.
+
+- **Consumer poll timeout**: needs to align with the existing
+  `receive(timeout_millis)` API. Kafka's `poll()` takes a timeout
+  directly, so this maps cleanly.
diff --git a/tests/unit/test_pubsub/test_kafka_backend.py b/tests/unit/test_pubsub/test_kafka_backend.py
new file mode 100644
index 00000000..456386f0
--- /dev/null
+++ b/tests/unit/test_pubsub/test_kafka_backend.py
@@ -0,0 +1,131 @@
+"""
+Unit tests for Kafka backend — topic parsing and factory dispatch.
+Does not require a running Kafka instance.
+"""
+
+import pytest
+import argparse
+
+from trustgraph.base.kafka_backend import KafkaBackend
+from trustgraph.base.pubsub import get_pubsub, add_pubsub_args
+
+
+class TestKafkaParseTopic:
+
+    @pytest.fixture
+    def backend(self):
+        b = object.__new__(KafkaBackend)
+        return b
+
+    def test_flow_is_durable(self, backend):
+        name, cls, durable = backend._parse_topic('flow:tg:text-completion-request')
+        assert durable is True
+        assert cls == 'flow'
+        assert name == 'tg.flow.text-completion-request'
+
+    def test_notify_is_not_durable(self, backend):
+        name, cls, durable = backend._parse_topic('notify:tg:config')
+        assert durable is False
+        assert cls == 'notify'
+        assert name == 'tg.notify.config'
+
+    def test_request_is_not_durable(self, backend):
+        name, cls, durable = backend._parse_topic('request:tg:config')
+        assert durable is False
+        assert cls == 'request'
+        assert name == 'tg.request.config'
+
+    def test_response_is_not_durable(self, backend):
+        name, cls, durable = backend._parse_topic('response:tg:librarian')
+        assert durable is False
+        assert cls == 'response'
+        assert name == 'tg.response.librarian'
+
+    def test_custom_topicspace(self, backend):
+        name, cls, durable = backend._parse_topic('flow:prod:my-queue')
+        assert name == 'prod.flow.my-queue'
+        assert durable is True
+
+    def test_no_colon_defaults_to_flow(self, backend):
+        name, cls, durable = backend._parse_topic('simple-queue')
+        assert name == 'tg.flow.simple-queue'
+        assert cls == 'flow'
+        assert durable is True
+
+    def test_invalid_class_raises(self, backend):
+        with pytest.raises(ValueError, match="Invalid topic class"):
+            backend._parse_topic('unknown:tg:topic')
+
+    def test_topic_with_flow_suffix(self, backend):
+        """Topic names with flow suffix (e.g. :default) are preserved."""
+        name, cls, durable = backend._parse_topic('request:tg:prompt:default')
+        assert name == 'tg.request.prompt:default'
+
+
+class TestKafkaRetention:
+
+    @pytest.fixture
+    def backend(self):
+        b = object.__new__(KafkaBackend)
+        return b
+
+    def test_flow_gets_long_retention(self, backend):
+        assert backend._retention_ms('flow') == 7 * 24 * 60 * 60 * 1000
+
+    def test_request_gets_short_retention(self, backend):
+        assert backend._retention_ms('request') == 300 * 1000
+
+    def test_response_gets_short_retention(self, backend):
+        assert backend._retention_ms('response') == 300 * 1000
+
+    def test_notify_gets_short_retention(self, backend):
+        assert backend._retention_ms('notify') == 300 * 1000
+
+
+class TestGetPubsubKafka:
+
+    def test_factory_creates_kafka_backend(self):
+        backend = get_pubsub(pubsub_backend='kafka')
+        assert isinstance(backend, KafkaBackend)
+
+    def test_factory_passes_config(self):
+        backend = get_pubsub(
+            pubsub_backend='kafka',
+            kafka_bootstrap_servers='myhost:9093',
+            kafka_security_protocol='SASL_SSL',
+            kafka_sasl_mechanism='PLAIN',
+            kafka_sasl_username='user',
+            kafka_sasl_password='pass',
+        )
+        assert isinstance(backend, KafkaBackend)
+        assert backend._bootstrap_servers == 'myhost:9093'
+        assert backend._admin_config['security.protocol'] == 'SASL_SSL'
+        assert backend._admin_config['sasl.mechanism'] == 'PLAIN'
+        assert backend._admin_config['sasl.username'] == 'user'
+        assert backend._admin_config['sasl.password'] == 'pass'
+
+
+class TestAddPubsubArgsKafka:
+
+    def test_kafka_args_present(self):
+        parser = argparse.ArgumentParser()
+        add_pubsub_args(parser)
+        args = parser.parse_args([
+            '--pubsub-backend', 'kafka',
+            '--kafka-bootstrap-servers', 'myhost:9093',
+        ])
+        assert args.pubsub_backend == 'kafka'
+        assert args.kafka_bootstrap_servers == 'myhost:9093'
+
+    def test_kafka_defaults_container(self):
+        parser = argparse.ArgumentParser()
+        add_pubsub_args(parser)
+        args = parser.parse_args([])
+        assert args.kafka_bootstrap_servers == 'kafka:9092'
+        assert args.kafka_security_protocol == 'PLAINTEXT'
+
+    def test_kafka_standalone_defaults_to_localhost(self):
+        parser = argparse.ArgumentParser()
+        add_pubsub_args(parser, standalone=True)
+        args = parser.parse_args([])
+        assert args.kafka_bootstrap_servers == 'localhost:9092'
diff --git a/tests/unit/test_pubsub/test_rabbitmq_backend.py b/tests/unit/test_pubsub/test_rabbitmq_backend.py
index ffe18fd7..5bafb78f 100644
--- a/tests/unit/test_pubsub/test_rabbitmq_backend.py
+++ b/tests/unit/test_pubsub/test_rabbitmq_backend.py
@@ -1,18 +1,16 @@
 """
-Unit tests for RabbitMQ backend — queue name mapping and factory dispatch.
+Unit tests for RabbitMQ backend — topic parsing and factory dispatch.
 Does not require a running RabbitMQ instance.
 """
 
 import pytest
 import argparse
 
-pika = pytest.importorskip("pika", reason="pika not installed")
-
 from trustgraph.base.rabbitmq_backend import RabbitMQBackend
 from trustgraph.base.pubsub import get_pubsub, add_pubsub_args
 
 
-class TestRabbitMQMapQueueName:
+class TestRabbitMQParseTopic:
 
     @pytest.fixture
     def backend(self):
@@ -20,43 +18,48 @@ class TestRabbitMQMapQueueName:
         return b
 
     def test_flow_is_durable(self, backend):
-        name, durable = backend.map_queue_name('flow:tg:text-completion-request')
+        exchange, cls, durable = backend._parse_topic('flow:tg:text-completion-request')
         assert durable is True
-        assert name == 'tg.flow.text-completion-request'
+        assert cls == 'flow'
+        assert exchange == 'tg.flow.text-completion-request'
 
     def test_notify_is_not_durable(self, backend):
-        name, durable = backend.map_queue_name('notify:tg:config')
+        exchange, cls, durable = backend._parse_topic('notify:tg:config')
         assert durable is False
-        assert name == 'tg.notify.config'
+        assert cls == 'notify'
+        assert exchange == 'tg.notify.config'
 
     def test_request_is_not_durable(self, backend):
-        name, durable = backend.map_queue_name('request:tg:config')
+        exchange, cls, durable = backend._parse_topic('request:tg:config')
         assert durable is False
-        assert name == 'tg.request.config'
+        assert cls == 'request'
+        assert exchange == 'tg.request.config'
 
     def test_response_is_not_durable(self, backend):
-        name, durable = backend.map_queue_name('response:tg:librarian')
+        exchange, cls, durable = backend._parse_topic('response:tg:librarian')
         assert durable is False
-        assert name == 'tg.response.librarian'
+        assert cls == 'response'
+        assert exchange == 'tg.response.librarian'
 
     def test_custom_topicspace(self, backend):
-        name, durable = backend.map_queue_name('flow:prod:my-queue')
-        assert name == 'prod.flow.my-queue'
+        exchange, cls, durable = backend._parse_topic('flow:prod:my-queue')
+        assert exchange == 'prod.flow.my-queue'
         assert durable is True
 
     def test_no_colon_defaults_to_flow(self, backend):
-        name, durable = backend.map_queue_name('simple-queue')
-        assert name == 'tg.simple-queue'
-        assert durable is False
+        exchange, cls, durable = backend._parse_topic('simple-queue')
+        assert exchange == 'tg.flow.simple-queue'
+        assert cls == 'flow'
+        assert durable is True
 
     def test_invalid_class_raises(self, backend):
-        with pytest.raises(ValueError, match="Invalid queue class"):
-            backend.map_queue_name('unknown:tg:topic')
+        with pytest.raises(ValueError, match="Invalid topic class"):
+            backend._parse_topic('unknown:tg:topic')
 
-    def test_flow_with_flow_suffix(self, backend):
-        """Queue names with flow suffix (e.g. :default) are preserved."""
-        name, durable = backend.map_queue_name('request:tg:prompt:default')
-        assert name == 'tg.request.prompt:default'
+    def test_topic_with_flow_suffix(self, backend):
+        """Topic names with flow suffix (e.g. :default) are preserved."""
+        exchange, cls, durable = backend._parse_topic('request:tg:prompt:default')
+        assert exchange == 'tg.request.prompt:default'
 
 
 class TestGetPubsubRabbitMQ:
diff --git a/trustgraph-base/pyproject.toml b/trustgraph-base/pyproject.toml
index 4f1bce76..c4ef3c27 100644
--- a/trustgraph-base/pyproject.toml
+++ b/trustgraph-base/pyproject.toml
@@ -15,6 +15,7 @@ dependencies = [
     "requests",
     "python-logging-loki",
     "pika",
+    "confluent-kafka",
     "pyyaml",
 ]
 classifiers = [
diff --git a/trustgraph-base/trustgraph/base/backend.py b/trustgraph-base/trustgraph/base/backend.py
index 0f95ca1b..a105ca17 100644
--- a/trustgraph-base/trustgraph/base/backend.py
+++ b/trustgraph-base/trustgraph/base/backend.py
@@ -121,7 +121,7 @@ class PubSubBackend(Protocol):
         Create a producer for a topic.
 
         Args:
-            topic: Generic topic format (qos/tenant/namespace/queue)
+            topic: Queue identifier in class:topicspace:topic format
             schema: Dataclass type for messages
             **options: Backend-specific options (e.g., chunking_enabled)
 
@@ -159,59 +159,55 @@ class PubSubBackend(Protocol):
         """
         ...
 
-    async def create_queue(self, topic: str, subscription: str) -> None:
+    async def create_topic(self, topic: str) -> None:
         """
-        Pre-create a queue so it exists before any consumer connects.
+        Create the broker-side resources for a logical topic.
 
-        The topic and subscription together identify the queue, mirroring
-        create_consumer where the queue name is derived from both.
+        For RabbitMQ this creates a fanout exchange.  For Pulsar this is
+        a no-op (topics auto-create on first use).
 
-        Idempotent — creating an already-existing queue succeeds silently.
+        Idempotent — creating an already-existing topic succeeds silently.
 
         Args:
-            topic: Queue identifier in class:topicspace:topic format
-            subscription: Subscription/consumer group name
+            topic: Topic identifier in class:topicspace:topic format
         """
         ...
 
-    async def delete_queue(self, topic: str, subscription: str) -> None:
+    async def delete_topic(self, topic: str) -> None:
         """
-        Delete a queue and any messages it contains.
+        Delete a topic and discard any in-flight messages.
 
-        The topic and subscription together identify the queue, mirroring
-        create_consumer where the queue name is derived from both.
+        For RabbitMQ this deletes the fanout exchange; consumer queues
+        lose their binding and drain naturally.
 
-        Idempotent — deleting a non-existent queue succeeds silently.
+        Idempotent — deleting a non-existent topic succeeds silently.
 
         Args:
-            topic: Queue identifier in class:topicspace:topic format
-            subscription: Subscription/consumer group name
+            topic: Topic identifier in class:topicspace:topic format
         """
         ...
 
-    async def queue_exists(self, topic: str, subscription: str) -> bool:
+    async def topic_exists(self, topic: str) -> bool:
         """
-        Check whether a queue exists.
+        Check whether a topic exists.
 
         Args:
-            topic: Queue identifier in class:topicspace:topic format
-            subscription: Subscription/consumer group name
+            topic: Topic identifier in class:topicspace:topic format
 
         Returns:
-            True if the queue exists, False otherwise.
+            True if the topic exists, False otherwise.
         """
         ...
 
-    async def ensure_queue(self, topic: str, subscription: str) -> None:
+    async def ensure_topic(self, topic: str) -> None:
         """
-        Ensure a queue exists, creating it if necessary.
+        Ensure a topic exists, creating it if necessary.
 
         Convenience wrapper — checks existence, creates if missing.
-        Used by system services on startup.
+        Used by the flow service and system services on startup.
 
         Args:
-            topic: Queue identifier in class:topicspace:topic format
-            subscription: Subscription/consumer group name
+            topic: Topic identifier in class:topicspace:topic format
         """
         ...
 
diff --git a/trustgraph-base/trustgraph/base/kafka_backend.py b/trustgraph-base/trustgraph/base/kafka_backend.py
new file mode 100644
index 00000000..8dfe8bfa
--- /dev/null
+++ b/trustgraph-base/trustgraph/base/kafka_backend.py
@@ -0,0 +1,400 @@
+"""
+Kafka backend implementation for pub/sub abstraction.
+
+Each logical topic maps to a Kafka topic.  The topic name encodes
+the full identity:
+
+    class:topicspace:topic  ->  topicspace.class.topic
+
+Producers publish to the topic directly.
+Consumers use consumer groups for competing-consumer semantics:
+
+  - flow / request: named consumer group (competing consumers)
+  - response / notify: unique consumer group per instance, filtering
+    messages by correlation ID (all subscribers see all messages)
+
+The flow service manages topic lifecycle via AdminClient.
+
+Architecture:
+  Producer  -->  [Kafka topic]  -->  Consumer Group A  --> Consumer
+                                -->  Consumer Group A  --> Consumer
+                                -->  Consumer Group B  --> Consumer (response)
+"""
+
+import asyncio
+import json
+import logging
+import uuid
+from typing import Any
+
+from confluent_kafka import (
+    Producer as KafkaProducer,
+    Consumer as KafkaConsumer,
+    TopicPartition,
+    KafkaError,
+    KafkaException,
+)
+from confluent_kafka.admin import AdminClient, NewTopic
+
+from .backend import PubSubBackend, BackendProducer, BackendConsumer, Message
+from .serialization import dataclass_to_dict, dict_to_dataclass
+
+logger = logging.getLogger(__name__)
+
+# Retention defaults (milliseconds)
+LONG_RETENTION_MS = 7 * 24 * 60 * 60 * 1000   # 7 days
+SHORT_RETENTION_MS = 300 * 1000                 # 5 minutes
+
+
+class KafkaMessage:
+    """Wrapper for Kafka messages to match Message protocol."""
+
+    def __init__(self, msg, schema_cls):
+        self._msg = msg
+        self._schema_cls = schema_cls
+        self._value = None
+
+    def value(self) -> Any:
+        """Deserialize and return the message value as a dataclass."""
+        if self._value is None:
+            data_dict = json.loads(self._msg.value().decode('utf-8'))
+            self._value = dict_to_dataclass(data_dict, self._schema_cls)
+        return self._value
+
+    def properties(self) -> dict:
+        """Return message properties from Kafka headers."""
+        headers = self._msg.headers() or []
+        return {
+            k: v.decode('utf-8') if isinstance(v, bytes) else v
+            for k, v in headers
+        }
+
+
+class KafkaBackendProducer:
+    """Publishes messages to a Kafka topic.
+
+    confluent-kafka Producer is thread-safe, so a single instance
+    can be shared across threads.
+    """
+
+    def __init__(self, bootstrap_servers, topic_name, durable):
+        self._topic_name = topic_name
+        self._durable = durable
+        self._producer = KafkaProducer({
+            'bootstrap.servers': bootstrap_servers,
+            'acks': 'all' if durable else '1',
+        })
+
+    def send(self, message: Any, properties: dict = {}) -> None:
+        data_dict = dataclass_to_dict(message)
+        json_data = json.dumps(data_dict).encode('utf-8')
+
+        headers = [
+            (k, str(v).encode('utf-8'))
+            for k, v in properties.items()
+        ] if properties else None
+
+        self._producer.produce(
+            topic=self._topic_name,
+            value=json_data,
+            headers=headers,
+        )
+        self._producer.flush()
+
+    def flush(self) -> None:
+        self._producer.flush()
+
+    def close(self) -> None:
+        self._producer.flush()
+
+
+class KafkaBackendConsumer:
+    """Consumes from a Kafka topic using a consumer group.
+
+    Uses confluent-kafka Consumer.poll() for message delivery.
+    Not thread-safe — each instance must be used from a single thread,
+    which matches the ThreadPoolExecutor pattern in consumer.py.
+    """
+
+    def __init__(self, bootstrap_servers, topic_name, group_id,
+                 schema_cls, auto_offset_reset='latest'):
+        self._bootstrap_servers = bootstrap_servers
+        self._topic_name = topic_name
+        self._group_id = group_id
+        self._schema_cls = schema_cls
+        self._auto_offset_reset = auto_offset_reset
+        self._consumer = None
+
+    def _connect(self):
+        self._consumer = KafkaConsumer({
+            'bootstrap.servers': self._bootstrap_servers,
+            'group.id': self._group_id,
+            'auto.offset.reset': self._auto_offset_reset,
+            'enable.auto.commit': False,
+        })
+        self._consumer.subscribe([self._topic_name])
+        logger.info(
+            f"Kafka consumer connected: topic={self._topic_name}, "
+            f"group={self._group_id}"
+        )
+
+    def _is_alive(self):
+        return self._consumer is not None
+
+    def ensure_connected(self) -> None:
+        """Eagerly connect and subscribe.
+
+        For response/notify consumers this must be called before the
+        corresponding request is published, so that the consumer is
+        assigned a partition and will see the response message.
+        """
+        if not self._is_alive():
+            self._connect()
+
+            # Force a partition assignment by polling briefly.
+            # Without this, the consumer may not be assigned partitions
+            # until the first real poll(), creating a race where the
+            # request is sent before assignment completes.
+            self._consumer.poll(timeout=1.0)
+
+    def receive(self, timeout_millis: int = 2000) -> Message:
+        """Receive a message. Raises TimeoutError if none available."""
+        if not self._is_alive():
+            self._connect()
+
+        timeout_seconds = timeout_millis / 1000.0
+        msg = self._consumer.poll(timeout=timeout_seconds)
+
+        if msg is None:
+            raise TimeoutError("No message received within timeout")
+
+        if msg.error():
+            error = msg.error()
+            if error.code() == KafkaError._PARTITION_EOF:
+                raise TimeoutError("End of partition reached")
+            raise KafkaException(error)
+
+        return KafkaMessage(msg, self._schema_cls)
+
+    def acknowledge(self, message: Message) -> None:
+        """Commit the message's offset (next offset to read)."""
+        if isinstance(message, KafkaMessage) and message._msg:
+            tp = TopicPartition(
+                message._msg.topic(),
+                message._msg.partition(),
+                message._msg.offset() + 1,
+            )
+            self._consumer.commit(offsets=[tp], asynchronous=False)
+
+    def negative_acknowledge(self, message: Message) -> None:
+        """Seek back to the message's offset for redelivery."""
+        if isinstance(message, KafkaMessage) and message._msg:
+            tp = TopicPartition(
+                message._msg.topic(),
+                message._msg.partition(),
+                message._msg.offset(),
+            )
+            self._consumer.seek(tp)
+
+    def unsubscribe(self) -> None:
+        if self._consumer:
+            try:
+                self._consumer.unsubscribe()
+            except Exception:
+                pass
+
+    def close(self) -> None:
+        if self._consumer:
+            try:
+                self._consumer.close()
+            except Exception:
+                pass
+            self._consumer = None
+
+
+class KafkaBackend:
+    """Kafka pub/sub backend using one topic per logical topic."""
+
+    def __init__(self, bootstrap_servers='localhost:9092',
+                 security_protocol='PLAINTEXT',
+                 sasl_mechanism=None,
+                 sasl_username=None,
+                 sasl_password=None):
+        self._bootstrap_servers = bootstrap_servers
+
+        # AdminClient config
+        self._admin_config = {
+            'bootstrap.servers': bootstrap_servers,
+        }
+
+        if security_protocol != 'PLAINTEXT':
+            self._admin_config['security.protocol'] = security_protocol
+            if sasl_mechanism:
+                self._admin_config['sasl.mechanism'] = sasl_mechanism
+            if sasl_username:
+                self._admin_config['sasl.username'] = sasl_username
+            if sasl_password:
+                self._admin_config['sasl.password'] = sasl_password
+
+        logger.info(
+            f"Kafka backend: {bootstrap_servers} "
+            f"protocol={security_protocol}"
+        )
+
+    def _parse_topic(self, topic_id: str) -> tuple[str, str, bool]:
+        """
+        Parse topic identifier into Kafka topic name, class, and durability.
+
+        Format: class:topicspace:topic
+        Returns: (topic_name, class, durable)
+        """
+        if ':' not in topic_id:
+            return f'tg.flow.{topic_id}', 'flow', True
+
+        parts = topic_id.split(':', 2)
+        if len(parts) != 3:
+            raise ValueError(
+                f"Invalid topic format: {topic_id}, "
+                f"expected class:topicspace:topic"
+            )
+
+        cls, topicspace, topic = parts
+
+        if cls == 'flow':
+            durable = True
+        elif cls in ('request', 'response', 'notify'):
+            durable = False
+        else:
+            raise ValueError(
+                f"Invalid topic class: {cls}, "
+                f"expected flow, request, response, or notify"
+            )
+
+        topic_name = f"{topicspace}.{cls}.{topic}"
+
+        return topic_name, cls, durable
+
+    def _retention_ms(self, cls):
+        """Return retention.ms for a topic class."""
+        if cls == 'flow':
+            return LONG_RETENTION_MS
+        return SHORT_RETENTION_MS
+
+    def create_producer(self, topic: str, schema: type,
+                        **options) -> BackendProducer:
+        topic_name, cls, durable = self._parse_topic(topic)
+        logger.debug(f"Creating producer: topic={topic_name}")
+        return KafkaBackendProducer(
+            self._bootstrap_servers, topic_name, durable,
+        )
+
+    def create_consumer(self, topic: str, subscription: str, schema: type,
+                        initial_position: str = 'latest',
+                        **options) -> BackendConsumer:
+        """Create a consumer subscribed to a Kafka topic.
+
+        Behaviour is determined by the topic's class prefix:
+          - flow: named consumer group, competing consumers
+          - request: named consumer group, competing consumers
+          - response: unique consumer group per instance
+          - notify: unique consumer group per instance
+        """
+        topic_name, cls, durable = self._parse_topic(topic)
+
+        if cls in ('response', 'notify'):
+            # Per-subscriber: unique group so every instance sees
+            # every message.  Filter by correlation ID happens at
+            # the Subscriber layer above.
+            group_id = f"{subscription}-{uuid.uuid4()}"
+            auto_offset_reset = 'latest'
+        else:
+            # Shared: named group, competing consumers
+            group_id = subscription
+            auto_offset_reset = (
+                'earliest' if initial_position == 'earliest'
+                else 'latest'
+            )
+
+        logger.debug(
+            f"Creating consumer: topic={topic_name}, "
+            f"group={group_id}, cls={cls}"
+        )
+
+        return KafkaBackendConsumer(
+            self._bootstrap_servers, topic_name, group_id,
+            schema, auto_offset_reset,
+        )
+
+    def _create_topic_sync(self, topic_name, retention_ms):
+        """Blocking topic creation via AdminClient."""
+        admin = AdminClient(self._admin_config)
+        new_topic = NewTopic(
+            topic_name,
+            num_partitions=1,
+            replication_factor=1,
+            config={
+                'retention.ms': str(retention_ms),
+            },
+        )
+        fs = admin.create_topics([new_topic])
+        for name, f in fs.items():
+            try:
+                f.result()
+                logger.info(f"Created topic: {name}")
+            except KafkaException as e:
+                # Topic already exists — idempotent
+                if e.args[0].code() == KafkaError.TOPIC_ALREADY_EXISTS:
+                    logger.debug(f"Topic already exists: {name}")
+                else:
+                    raise
+
+    async def create_topic(self, topic: str) -> None:
+        """Create a Kafka topic with appropriate retention."""
+        topic_name, cls, durable = self._parse_topic(topic)
+        retention_ms = self._retention_ms(cls)
+        await asyncio.to_thread(
+            self._create_topic_sync, topic_name, retention_ms,
+        )
+
+    def _delete_topic_sync(self, topic_name):
+        """Blocking topic deletion via AdminClient."""
+        admin = AdminClient(self._admin_config)
+        fs = admin.delete_topics([topic_name])
+        for name, f in fs.items():
+            try:
+                f.result()
+                logger.info(f"Deleted topic: {name}")
+            except KafkaException as e:
+                # Topic doesn't exist — idempotent
+                if e.args[0].code() == KafkaError.UNKNOWN_TOPIC_OR_PART:
+                    logger.debug(f"Topic not found: {name}")
+                else:
+                    raise
+            except Exception as e:
+                logger.debug(f"Topic delete for {name}: {e}")
+
+    async def delete_topic(self, topic: str) -> None:
+        """Delete a Kafka topic."""
+        topic_name, cls, durable = self._parse_topic(topic)
+        await asyncio.to_thread(self._delete_topic_sync, topic_name)
+
+    def _topic_exists_sync(self, topic_name):
+        """Blocking topic existence check via AdminClient."""
+        admin = AdminClient(self._admin_config)
+        metadata = admin.list_topics(timeout=10)
+        return topic_name in metadata.topics
+
+    async def topic_exists(self, topic: str) -> bool:
+        """Check whether a Kafka topic exists."""
+        topic_name, cls, durable = self._parse_topic(topic)
+        return await asyncio.to_thread(
+            self._topic_exists_sync, topic_name,
+        )
+
+    async def ensure_topic(self, topic: str) -> None:
+        """Ensure a topic exists, creating it if necessary."""
+        if not await self.topic_exists(topic):
+            await self.create_topic(topic)
+
+    def close(self) -> None:
+        pass
diff --git a/trustgraph-base/trustgraph/base/pubsub.py b/trustgraph-base/trustgraph/base/pubsub.py
index a7ae3719..fb4765c1 100644
--- a/trustgraph-base/trustgraph/base/pubsub.py
+++ b/trustgraph-base/trustgraph/base/pubsub.py
@@ -17,6 +17,12 @@ DEFAULT_RABBITMQ_USERNAME = os.getenv("RABBITMQ_USERNAME", 'guest')
 DEFAULT_RABBITMQ_PASSWORD = os.getenv("RABBITMQ_PASSWORD", 'guest')
 DEFAULT_RABBITMQ_VHOST = os.getenv("RABBITMQ_VHOST", '/')
 
+DEFAULT_KAFKA_BOOTSTRAP = os.getenv("KAFKA_BOOTSTRAP_SERVERS", 'kafka:9092')
+DEFAULT_KAFKA_PROTOCOL = os.getenv("KAFKA_SECURITY_PROTOCOL", 'PLAINTEXT')
+DEFAULT_KAFKA_SASL_MECHANISM = os.getenv("KAFKA_SASL_MECHANISM", None)
+DEFAULT_KAFKA_SASL_USERNAME = os.getenv("KAFKA_SASL_USERNAME", None)
+DEFAULT_KAFKA_SASL_PASSWORD = os.getenv("KAFKA_SASL_PASSWORD", None)
+
 
 def get_pubsub(**config: Any) -> Any:
     """
@@ -47,6 +53,25 @@ def get_pubsub(**config: Any) -> Any:
             password=config.get('rabbitmq_password', DEFAULT_RABBITMQ_PASSWORD),
             vhost=config.get('rabbitmq_vhost', DEFAULT_RABBITMQ_VHOST),
         )
+    elif backend_type == 'kafka':
+        from .kafka_backend import KafkaBackend
+        return KafkaBackend(
+            bootstrap_servers=config.get(
+                'kafka_bootstrap_servers', DEFAULT_KAFKA_BOOTSTRAP,
+            ),
+            security_protocol=config.get(
+                'kafka_security_protocol', DEFAULT_KAFKA_PROTOCOL,
+            ),
+            sasl_mechanism=config.get(
+                'kafka_sasl_mechanism', DEFAULT_KAFKA_SASL_MECHANISM,
+            ),
+            sasl_username=config.get(
+                'kafka_sasl_username', DEFAULT_KAFKA_SASL_USERNAME,
+            ),
+            sasl_password=config.get(
+                'kafka_sasl_password', DEFAULT_KAFKA_SASL_PASSWORD,
+            ),
+        )
     else:
         raise ValueError(f"Unknown pub/sub backend: {backend_type}")
 
@@ -65,6 +90,7 @@ def add_pubsub_args(parser: ArgumentParser, standalone: bool = False) -> None:
     pulsar_host = STANDALONE_PULSAR_HOST if standalone else DEFAULT_PULSAR_HOST
     pulsar_listener = 'localhost' if standalone else None
     rabbitmq_host = 'localhost' if standalone else DEFAULT_RABBITMQ_HOST
+    kafka_bootstrap = 'localhost:9092' if standalone else DEFAULT_KAFKA_BOOTSTRAP
 
     parser.add_argument(
         '--pubsub-backend',
@@ -122,3 +148,34 @@ def add_pubsub_args(parser: ArgumentParser, standalone: bool = False) -> None:
         default=DEFAULT_RABBITMQ_VHOST,
         help=f'RabbitMQ vhost (default: {DEFAULT_RABBITMQ_VHOST})',
     )
+
+    # Kafka options
+    parser.add_argument(
+        '--kafka-bootstrap-servers',
+        default=kafka_bootstrap,
+        help=f'Kafka bootstrap servers (default: {kafka_bootstrap})',
+    )
+
+    parser.add_argument(
+        '--kafka-security-protocol',
+        default=DEFAULT_KAFKA_PROTOCOL,
+        help=f'Kafka security protocol (default: {DEFAULT_KAFKA_PROTOCOL})',
+    )
+
+    parser.add_argument(
+        '--kafka-sasl-mechanism',
+        default=DEFAULT_KAFKA_SASL_MECHANISM,
+        help='Kafka SASL mechanism',
+    )
+
+    parser.add_argument(
+        '--kafka-sasl-username',
+        default=DEFAULT_KAFKA_SASL_USERNAME,
+        help='Kafka SASL username',
+    )
+
+    parser.add_argument(
+        '--kafka-sasl-password',
+        default=DEFAULT_KAFKA_SASL_PASSWORD,
+        help='Kafka SASL password',
+    )
diff --git a/trustgraph-base/trustgraph/base/pulsar_backend.py b/trustgraph-base/trustgraph/base/pulsar_backend.py
index 2100483d..e27d16af 100644
--- a/trustgraph-base/trustgraph/base/pulsar_backend.py
+++ b/trustgraph-base/trustgraph/base/pulsar_backend.py
@@ -266,22 +266,22 @@ class PulsarBackend:
 
         return PulsarBackendConsumer(pulsar_consumer, schema)
 
-    async def create_queue(self, topic: str, subscription: str) -> None:
+    async def create_topic(self, topic: str) -> None:
         """No-op — Pulsar auto-creates topics on first use.
         TODO: Use admin REST API for explicit persistent topic creation."""
         pass
 
-    async def delete_queue(self, topic: str, subscription: str) -> None:
+    async def delete_topic(self, topic: str) -> None:
         """No-op — to be replaced with admin REST API calls.
-        TODO: Delete subscription and persistent topic via admin API."""
+        TODO: Delete persistent topic via admin API."""
         pass
 
-    async def queue_exists(self, topic: str, subscription: str) -> bool:
+    async def topic_exists(self, topic: str) -> bool:
         """Returns True — Pulsar auto-creates on subscribe.
         TODO: Use admin REST API for actual existence check."""
         return True
 
-    async def ensure_queue(self, topic: str, subscription: str) -> None:
+    async def ensure_topic(self, topic: str) -> None:
         """No-op — Pulsar auto-creates topics on first use.
         TODO: Use admin REST API for explicit creation."""
         pass
diff --git a/trustgraph-base/trustgraph/base/rabbitmq_backend.py b/trustgraph-base/trustgraph/base/rabbitmq_backend.py
index 43c717c3..73b80cb9 100644
--- a/trustgraph-base/trustgraph/base/rabbitmq_backend.py
+++ b/trustgraph-base/trustgraph/base/rabbitmq_backend.py
@@ -1,22 +1,24 @@
 """
 RabbitMQ backend implementation for pub/sub abstraction.
 
-Uses a single topic exchange per topicspace. The logical queue name
-becomes the routing key. Consumer behavior is determined by the
-subscription name:
+Each logical topic maps to its own fanout exchange.  The exchange name
+encodes the full topic identity:
 
-- Same subscription + same topic = shared queue (competing consumers)
-- Different subscriptions = separate queues (broadcast / fan-out)
+    class:topicspace:topic  →  exchange  topicspace.class.topic
 
-This mirrors Pulsar's subscription model using idiomatic RabbitMQ.
+Producers publish to the exchange with an empty routing key.
+Consumers declare and bind their own queues:
+
+  - flow / request: named durable/non-durable queue (competing consumers)
+  - response / notify: anonymous exclusive auto-delete queue (per-subscriber)
+
+The flow service manages topic lifecycle (create/delete exchanges).
+Consumers manage their own queue lifecycle (declare + bind on connect).
 
 Architecture:
-  Producer  -->  [tg exchange]  --routing key-->  [named queue]  --> Consumer
-                                --routing key-->  [named queue]  --> Consumer
-                                --routing key-->  [exclusive q]  --> Subscriber
-
-Uses basic_consume (push) instead of basic_get (polling) for
-efficient message delivery.
+  Producer  -->  [fanout exchange]  -->  [named queue]      --> Consumer
+                                    -->  [named queue]      --> Consumer
+                                    -->  [exclusive queue]  --> Subscriber
 """
 
 import asyncio
@@ -58,18 +60,16 @@ class RabbitMQMessage:
 
 
 class RabbitMQBackendProducer:
-    """Publishes messages to a topic exchange with a routing key.
+    """Publishes messages to a fanout exchange.
 
     Uses thread-local connections so each thread gets its own
     connection/channel. This avoids wire corruption from concurrent
     threads writing to the same socket (pika is not thread-safe).
     """
 
-    def __init__(self, connection_params, exchange_name, routing_key,
-                 durable):
+    def __init__(self, connection_params, exchange_name, durable):
         self._connection_params = connection_params
         self._exchange_name = exchange_name
-        self._routing_key = routing_key
         self._durable = durable
         self._local = threading.local()
 
@@ -90,7 +90,7 @@ class RabbitMQBackendProducer:
             chan = conn.channel()
             chan.exchange_declare(
                 exchange=self._exchange_name,
-                exchange_type='topic',
+                exchange_type='fanout',
                 durable=True,
             )
             self._local.connection = conn
@@ -113,7 +113,7 @@ class RabbitMQBackendProducer:
                 channel = self._get_channel()
                 channel.basic_publish(
                     exchange=self._exchange_name,
-                    routing_key=self._routing_key,
+                    routing_key='',
                     body=json_data.encode('utf-8'),
                     properties=amqp_properties,
                 )
@@ -144,19 +144,17 @@ class RabbitMQBackendProducer:
 
 
 class RabbitMQBackendConsumer:
-    """Consumes from a queue bound to a topic exchange.
+    """Consumes from a queue bound to a fanout exchange.
 
     Uses basic_consume (push model) with messages delivered to an
     internal thread-safe queue. process_data_events() drives both
     message delivery and heartbeat processing.
     """
 
-    def __init__(self, connection_params, exchange_name, routing_key,
-                 queue_name, schema_cls, durable, exclusive=False,
-                 auto_delete=False):
+    def __init__(self, connection_params, exchange_name, queue_name,
+                 schema_cls, durable, exclusive=False, auto_delete=False):
         self._connection_params = connection_params
         self._exchange_name = exchange_name
-        self._routing_key = routing_key
         self._queue_name = queue_name
         self._schema_cls = schema_cls
         self._durable = durable
@@ -171,17 +169,16 @@ class RabbitMQBackendConsumer:
         self._connection = pika.BlockingConnection(self._connection_params)
         self._channel = self._connection.channel()
 
-        # Declare the topic exchange (idempotent, also done by producers)
+        # Declare the fanout exchange (idempotent)
         self._channel.exchange_declare(
             exchange=self._exchange_name,
-            exchange_type='topic',
+            exchange_type='fanout',
             durable=True,
         )
 
         if self._exclusive:
             # Anonymous ephemeral queue (response/notify class).
-            # These are per-consumer and must be created here — the
-            # broker assigns the name.
+            # Per-consumer, broker assigns the name.
             result = self._channel.queue_declare(
                 queue='',
                 durable=False,
@@ -189,20 +186,22 @@ class RabbitMQBackendConsumer:
                 auto_delete=True,
             )
             self._queue_name = result.method.queue
-
-            self._channel.queue_bind(
-                queue=self._queue_name,
-                exchange=self._exchange_name,
-                routing_key=self._routing_key,
-            )
         else:
-            # Named queue (flow/request class). Queue must already
-            # exist — created by the flow service or ensure_queue.
-            # We just verify it exists and bind to consume.
+            # Named queue (flow/request class).
+            # Consumer owns its queue — declare and bind here.
             self._channel.queue_declare(
-                queue=self._queue_name, passive=True,
+                queue=self._queue_name,
+                durable=self._durable,
+                exclusive=False,
+                auto_delete=False,
             )
 
+        # Bind queue to the fanout exchange
+        self._channel.queue_bind(
+            queue=self._queue_name,
+            exchange=self._exchange_name,
+        )
+
         self._channel.basic_qos(prefetch_count=1)
 
         # Register push-based consumer
@@ -318,7 +317,7 @@ class RabbitMQBackendConsumer:
 
 
 class RabbitMQBackend:
-    """RabbitMQ pub/sub backend using a topic exchange per topicspace."""
+    """RabbitMQ pub/sub backend using one fanout exchange per topic."""
 
     def __init__(self, host='localhost', port=5672, username='guest',
                  password='guest', vhost='/'):
@@ -331,20 +330,23 @@ class RabbitMQBackend:
         )
         logger.info(f"RabbitMQ backend: {host}:{port} vhost={vhost}")
 
-    def _parse_queue_id(self, queue_id: str) -> tuple[str, str, str, bool]:
+    def _parse_topic(self, topic_id: str) -> tuple[str, str, bool]:
         """
-        Parse queue identifier into exchange, routing key, and durability.
+        Parse topic identifier into exchange name and durability.
 
         Format: class:topicspace:topic
-        Returns: (exchange_name, routing_key, class, durable)
-        """
-        if ':' not in queue_id:
-            return 'tg', queue_id, 'flow', False
+        Returns: (exchange_name, class, durable)
 
-        parts = queue_id.split(':', 2)
+        The exchange name encodes the full topic identity:
+            class:topicspace:topic  →  topicspace.class.topic
+        """
+        if ':' not in topic_id:
+            return f'tg.flow.{topic_id}', 'flow', True
+
+        parts = topic_id.split(':', 2)
         if len(parts) != 3:
             raise ValueError(
-                f"Invalid queue format: {queue_id}, "
+                f"Invalid topic format: {topic_id}, "
                 f"expected class:topicspace:topic"
             )
 
@@ -356,36 +358,28 @@ class RabbitMQBackend:
             durable = False
         else:
             raise ValueError(
-                f"Invalid queue class: {cls}, "
+                f"Invalid topic class: {cls}, "
                 f"expected flow, request, response, or notify"
             )
 
-        # Exchange per topicspace, routing key includes class
-        exchange_name = topicspace
-        routing_key = f"{cls}.{topic}"
+        exchange_name = f"{topicspace}.{cls}.{topic}"
 
-        return exchange_name, routing_key, cls, durable
-
-    # Keep map_queue_name for backward compatibility with tests
-    def map_queue_name(self, queue_id: str) -> tuple[str, bool]:
-        exchange, routing_key, cls, durable = self._parse_queue_id(queue_id)
-        return f"{exchange}.{routing_key}", durable
+        return exchange_name, cls, durable
 
     def create_producer(self, topic: str, schema: type,
                         **options) -> BackendProducer:
-        exchange, routing_key, cls, durable = self._parse_queue_id(topic)
+        exchange, cls, durable = self._parse_topic(topic)
         logger.debug(
-            f"Creating producer: exchange={exchange}, "
-            f"routing_key={routing_key}"
+            f"Creating producer: exchange={exchange}"
         )
         return RabbitMQBackendProducer(
-            self._connection_params, exchange, routing_key, durable,
+            self._connection_params, exchange, durable,
         )
 
     def create_consumer(self, topic: str, subscription: str, schema: type,
                         initial_position: str = 'latest',
                         **options) -> BackendConsumer:
-        """Create a consumer with a queue bound to the topic exchange.
+        """Create a consumer with a queue bound to the topic's exchange.
 
         Behaviour is determined by the topic's class prefix:
           - flow: named durable queue, competing consumers (round-robin)
@@ -393,7 +387,7 @@ class RabbitMQBackend:
           - response: anonymous ephemeral queue, per-subscriber (auto-delete)
           - notify: anonymous ephemeral queue, per-subscriber (auto-delete)
         """
-        exchange, routing_key, cls, durable = self._parse_queue_id(topic)
+        exchange, cls, durable = self._parse_topic(topic)
 
         if cls in ('response', 'notify'):
             # Per-subscriber: anonymous queue, auto-deleted on disconnect
@@ -403,45 +397,33 @@ class RabbitMQBackend:
             auto_delete = True
         else:
             # Shared: named queue, competing consumers
-            queue_name = f"{exchange}.{routing_key}.{subscription}"
+            queue_name = f"{exchange}.{subscription}"
             queue_durable = durable
             exclusive = False
             auto_delete = False
 
         logger.debug(
             f"Creating consumer: exchange={exchange}, "
-            f"routing_key={routing_key}, queue={queue_name or '(anonymous)'}, "
-            f"cls={cls}"
+            f"queue={queue_name or '(anonymous)'}, cls={cls}"
         )
 
         return RabbitMQBackendConsumer(
-            self._connection_params, exchange, routing_key,
+            self._connection_params, exchange,
             queue_name, schema, queue_durable, exclusive, auto_delete,
         )
 
-    def _create_queue_sync(self, exchange, routing_key, queue_name, durable):
-        """Blocking queue creation — run via asyncio.to_thread."""
+    def _create_topic_sync(self, exchange_name):
+        """Blocking exchange creation — run via asyncio.to_thread."""
         connection = None
         try:
             connection = pika.BlockingConnection(self._connection_params)
             channel = connection.channel()
             channel.exchange_declare(
-                exchange=exchange,
-                exchange_type='topic',
+                exchange=exchange_name,
+                exchange_type='fanout',
                 durable=True,
             )
-            channel.queue_declare(
-                queue=queue_name,
-                durable=durable,
-                exclusive=False,
-                auto_delete=False,
-            )
-            channel.queue_bind(
-                queue=queue_name,
-                exchange=exchange,
-                routing_key=routing_key,
-            )
-            logger.info(f"Created queue: {queue_name}")
+            logger.info(f"Created topic (exchange): {exchange_name}")
         finally:
             if connection and connection.is_open:
                 try:
@@ -449,34 +431,30 @@ class RabbitMQBackend:
                 except Exception:
                     pass
 
-    async def create_queue(self, topic: str, subscription: str) -> None:
-        """Pre-create a named queue bound to the topic exchange.
+    async def create_topic(self, topic: str) -> None:
+        """Create the fanout exchange for a logical topic.
 
-        Only applies to shared queues (flow/request class). Response and
-        notify queues are anonymous/auto-delete and created by consumers.
+        Only applies to flow and request class topics. Response and
+        notify exchanges are created on demand by consumers.
         """
-        exchange, routing_key, cls, durable = self._parse_queue_id(topic)
+        exchange, cls, durable = self._parse_topic(topic)
 
         if cls in ('response', 'notify'):
             return
 
-        queue_name = f"{exchange}.{routing_key}.{subscription}"
-        await asyncio.to_thread(
-            self._create_queue_sync, exchange, routing_key,
-            queue_name, durable,
-        )
+        await asyncio.to_thread(self._create_topic_sync, exchange)
 
-    def _delete_queue_sync(self, queue_name):
-        """Blocking queue deletion — run via asyncio.to_thread."""
+    def _delete_topic_sync(self, exchange_name):
+        """Blocking exchange deletion — run via asyncio.to_thread."""
         connection = None
         try:
             connection = pika.BlockingConnection(self._connection_params)
             channel = connection.channel()
-            channel.queue_delete(queue=queue_name)
-            logger.info(f"Deleted queue: {queue_name}")
+            channel.exchange_delete(exchange=exchange_name)
+            logger.info(f"Deleted topic (exchange): {exchange_name}")
         except Exception as e:
-            # Idempotent — queue may already be gone
-            logger.debug(f"Queue delete for {queue_name}: {e}")
+            # Idempotent — exchange may already be gone
+            logger.debug(f"Exchange delete for {exchange_name}: {e}")
         finally:
             if connection and connection.is_open:
                 try:
@@ -484,31 +462,27 @@ class RabbitMQBackend:
                 except Exception:
                     pass
 
-    async def delete_queue(self, topic: str, subscription: str) -> None:
-        """Delete a named queue and any messages it contains.
+    async def delete_topic(self, topic: str) -> None:
+        """Delete a topic's fanout exchange.
 
-        Only applies to shared queues (flow/request class). Response and
-        notify queues are anonymous/auto-delete and managed by the broker.
+        Consumer queues lose their binding and drain naturally.
         """
-        exchange, routing_key, cls, durable = self._parse_queue_id(topic)
+        exchange, cls, durable = self._parse_topic(topic)
+        await asyncio.to_thread(self._delete_topic_sync, exchange)
 
-        if cls in ('response', 'notify'):
-            return
-
-        queue_name = f"{exchange}.{routing_key}.{subscription}"
-        await asyncio.to_thread(self._delete_queue_sync, queue_name)
-
-    def _queue_exists_sync(self, queue_name):
-        """Blocking queue existence check — run via asyncio.to_thread.
+    def _topic_exists_sync(self, exchange_name):
+        """Blocking exchange existence check — run via asyncio.to_thread.
         Uses passive=True which checks without creating."""
         connection = None
         try:
             connection = pika.BlockingConnection(self._connection_params)
             channel = connection.channel()
-            channel.queue_declare(queue=queue_name, passive=True)
+            channel.exchange_declare(
+                exchange=exchange_name, passive=True,
+            )
             return True
         except pika.exceptions.ChannelClosedByBroker:
-            # 404 NOT_FOUND — queue does not exist
+            # 404 NOT_FOUND — exchange does not exist
             return False
         finally:
             if connection and connection.is_open:
@@ -517,26 +491,25 @@ class RabbitMQBackend:
                 except Exception:
                     pass
 
-    async def queue_exists(self, topic: str, subscription: str) -> bool:
-        """Check whether a named queue exists.
+    async def topic_exists(self, topic: str) -> bool:
+        """Check whether a topic's exchange exists.
 
-        Only applies to shared queues (flow/request class). Response and
-        notify queues are anonymous/ephemeral — always returns False.
+        Only applies to flow and request class topics. Response and
+        notify topics are ephemeral — always returns False.
         """
-        exchange, routing_key, cls, durable = self._parse_queue_id(topic)
+        exchange, cls, durable = self._parse_topic(topic)
 
         if cls in ('response', 'notify'):
             return False
 
-        queue_name = f"{exchange}.{routing_key}.{subscription}"
         return await asyncio.to_thread(
-            self._queue_exists_sync, queue_name
+            self._topic_exists_sync, exchange
         )
 
-    async def ensure_queue(self, topic: str, subscription: str) -> None:
-        """Ensure a queue exists, creating it if necessary."""
-        if not await self.queue_exists(topic, subscription):
-            await self.create_queue(topic, subscription)
+    async def ensure_topic(self, topic: str) -> None:
+        """Ensure a topic exists, creating it if necessary."""
+        if not await self.topic_exists(topic):
+            await self.create_topic(topic)
 
     def close(self) -> None:
         pass
diff --git a/trustgraph-flow/trustgraph/chunking/recursive/chunker.py b/trustgraph-flow/trustgraph/chunking/recursive/chunker.py
index 257edc8c..81fb7303 100755
--- a/trustgraph-flow/trustgraph/chunking/recursive/chunker.py
+++ b/trustgraph-flow/trustgraph/chunking/recursive/chunker.py
@@ -10,6 +10,8 @@ from prometheus_client import Histogram
 from ... schema import TextDocument, Chunk, Metadata, Triples
 from ... base import ChunkingService, ConsumerSpec, ProducerSpec
 
+RecursiveCharacterTextSplitter = None
+
 from ... provenance import (
     chunk_uri as make_chunk_uri, derived_entity_triples,
     set_graph, GRAPH_SOURCE,
@@ -41,8 +43,12 @@ class Processor(ChunkingService):
         self.default_chunk_size = chunk_size
         self.default_chunk_overlap = chunk_overlap
 
-        from langchain_text_splitters import RecursiveCharacterTextSplitter
-        self.RecursiveCharacterTextSplitter = RecursiveCharacterTextSplitter
+        global RecursiveCharacterTextSplitter
+        if RecursiveCharacterTextSplitter is None:
+            from langchain_text_splitters import (
+                RecursiveCharacterTextSplitter as _cls,
+            )
+            RecursiveCharacterTextSplitter = _cls
 
         if not hasattr(__class__, "chunk_metric"):
             __class__.chunk_metric = Histogram(
diff --git a/trustgraph-flow/trustgraph/chunking/token/chunker.py b/trustgraph-flow/trustgraph/chunking/token/chunker.py
index d315c428..e0d58b41 100755
--- a/trustgraph-flow/trustgraph/chunking/token/chunker.py
+++ b/trustgraph-flow/trustgraph/chunking/token/chunker.py
@@ -10,6 +10,8 @@ from prometheus_client import Histogram
 from ... schema import TextDocument, Chunk, Metadata, Triples
 from ... base import ChunkingService, ConsumerSpec, ProducerSpec
 
+TokenTextSplitter = None
+
 from ... provenance import (
     chunk_uri as make_chunk_uri, derived_entity_triples,
     set_graph, GRAPH_SOURCE,
@@ -41,8 +43,10 @@ class Processor(ChunkingService):
         self.default_chunk_size = chunk_size
         self.default_chunk_overlap = chunk_overlap
 
-        from langchain_text_splitters import TokenTextSplitter
-        self.TokenTextSplitter = TokenTextSplitter
+        global TokenTextSplitter
+        if TokenTextSplitter is None:
+            from langchain_text_splitters import TokenTextSplitter as _cls
+            TokenTextSplitter = _cls
 
         if not hasattr(__class__, "chunk_metric"):
             __class__.chunk_metric = Histogram(
diff --git a/trustgraph-flow/trustgraph/config/service/service.py b/trustgraph-flow/trustgraph/config/service/service.py
index 75232315..fe44b852 100644
--- a/trustgraph-flow/trustgraph/config/service/service.py
+++ b/trustgraph-flow/trustgraph/config/service/service.py
@@ -124,9 +124,7 @@ class Processor(AsyncProcessor):
 
     async def start(self):
 
-        await self.pubsub.ensure_queue(
-            self.config_request_topic, self.config_request_subscriber
-        )
+        await self.pubsub.ensure_topic(self.config_request_topic)
         await self.push()  # Startup poke: empty types = everything
         await self.config_request_consumer.start()
 
diff --git a/trustgraph-flow/trustgraph/cores/service.py b/trustgraph-flow/trustgraph/cores/service.py
index 400f96d1..93017c30 100755
--- a/trustgraph-flow/trustgraph/cores/service.py
+++ b/trustgraph-flow/trustgraph/cores/service.py
@@ -119,9 +119,7 @@ class Processor(AsyncProcessor):
 
     async def start(self):
 
-        await self.pubsub.ensure_queue(
-            self.knowledge_request_topic, self.knowledge_request_subscriber
-        )
+        await self.pubsub.ensure_topic(self.knowledge_request_topic)
         await super(Processor, self).start()
         await self.knowledge_request_consumer.start()
         await self.knowledge_response_producer.start()
diff --git a/trustgraph-flow/trustgraph/decoding/pdf/pdf_decoder.py b/trustgraph-flow/trustgraph/decoding/pdf/pdf_decoder.py
index ab31b717..7f9ca71d 100755
--- a/trustgraph-flow/trustgraph/decoding/pdf/pdf_decoder.py
+++ b/trustgraph-flow/trustgraph/decoding/pdf/pdf_decoder.py
@@ -15,6 +15,9 @@ from ... schema import Document, TextDocument, Metadata
 from ... schema import librarian_request_queue, librarian_response_queue
 from ... schema import Triples
 from ... base import FlowProcessor, ConsumerSpec, ProducerSpec, LibrarianClient
+
+PyPDFLoader = None
+
 from ... provenance import (
     document_uri, page_uri as make_page_uri, derived_entity_triples,
     set_graph, GRAPH_SOURCE,
@@ -128,7 +131,12 @@ class Processor(FlowProcessor):
                 fp.write(base64.b64decode(v.data))
                 fp.close()
 
-            from langchain_community.document_loaders import PyPDFLoader
+            global PyPDFLoader
+            if PyPDFLoader is None:
+                from langchain_community.document_loaders import (
+                    PyPDFLoader as _cls,
+                )
+                PyPDFLoader = _cls
             loader = PyPDFLoader(temp_path)
             pages = loader.load()
 
diff --git a/trustgraph-flow/trustgraph/flow/service/flow.py b/trustgraph-flow/trustgraph/flow/service/flow.py
index 477c6a2c..b864faf9 100644
--- a/trustgraph-flow/trustgraph/flow/service/flow.py
+++ b/trustgraph-flow/trustgraph/flow/service/flow.py
@@ -7,7 +7,7 @@ import logging
 # Module logger
 logger = logging.getLogger(__name__)
 
-# Queue deletion retry settings
+# Topic deletion retry settings
 DELETE_RETRIES = 5
 DELETE_RETRY_DELAY = 2  # seconds
 
@@ -215,11 +215,11 @@ class FlowConfig:
 
             return result
 
-        # Pre-create flow-level queues so the data path is wired
+        # Pre-create topic exchanges so the data path is wired
         # before processors receive their config and start connecting.
-        queues = self._collect_flow_queues(cls, repl_template_with_params)
-        for topic, subscription in queues:
-            await self.pubsub.create_queue(topic, subscription)
+        topics = self._collect_flow_topics(cls, repl_template_with_params)
+        for topic in topics:
+            await self.pubsub.create_topic(topic)
 
         # Build all processor config updates, then write in a single batch.
         updates = []
@@ -283,8 +283,8 @@ class FlowConfig:
             error = None,
         )
 
-    async def ensure_existing_flow_queues(self):
-        """Ensure queues exist for all already-running flows.
+    async def ensure_existing_flow_topics(self):
+        """Ensure topics exist for all already-running flows.
 
         Called on startup to handle flows that were started before this
         version of the flow service was deployed, or before a restart.
@@ -315,7 +315,7 @@ class FlowConfig:
                 if blueprint_data is None:
                     logger.warning(
                         f"Blueprint '{blueprint_name}' not found for "
-                        f"flow '{flow_id}', skipping queue creation"
+                        f"flow '{flow_id}', skipping topic creation"
                     )
                     continue
 
@@ -333,65 +333,63 @@ class FlowConfig:
                         )
                     return result
 
-                queues = self._collect_flow_queues(cls, repl_template)
-                for topic, subscription in queues:
-                    await self.pubsub.ensure_queue(topic, subscription)
+                topics = self._collect_flow_topics(cls, repl_template)
+                for topic in topics:
+                    await self.pubsub.ensure_topic(topic)
 
                 logger.info(
-                    f"Ensured queues for existing flow '{flow_id}'"
+                    f"Ensured topics for existing flow '{flow_id}'"
                 )
 
             except Exception as e:
                 logger.error(
-                    f"Failed to ensure queues for flow '{flow_id}': {e}"
+                    f"Failed to ensure topics for flow '{flow_id}': {e}"
                 )
 
-    def _collect_flow_queues(self, cls, repl_template):
-        """Collect (topic, subscription) pairs for all flow-level queues.
+    def _collect_flow_topics(self, cls, repl_template):
+        """Collect unique topic identifiers from the blueprint.
 
-        Iterates the blueprint's "flow" section and reads only the
-        "topics" dict from each processor entry.
+        Iterates the blueprint's "flow" section and returns a
+        deduplicated set of resolved topic strings.  The flow service
+        manages topic lifecycle (create/delete exchanges), not
+        individual consumer queues.
         """
-        queues = []
+        topics = set()
 
         for k, v in cls["flow"].items():
-            processor, variant = k.split(":", 1)
-            variant = repl_template(variant)
-
             for spec_name, topic_template in v.get("topics", {}).items():
                 topic = repl_template(topic_template)
-                subscription = f"{processor}--{variant}--{spec_name}"
-                queues.append((topic, subscription))
+                topics.add(topic)
 
-        return queues
+        return topics
 
-    async def _delete_queues(self, queues):
-        """Delete queues with retries. Best-effort — logs failures but
+    async def _delete_topics(self, topics):
+        """Delete topics with retries. Best-effort — logs failures but
         does not raise."""
         for attempt in range(DELETE_RETRIES):
             remaining = []
 
-            for topic, subscription in queues:
+            for topic in topics:
                 try:
-                    await self.pubsub.delete_queue(topic, subscription)
+                    await self.pubsub.delete_topic(topic)
                 except Exception as e:
                     logger.warning(
-                        f"Queue delete failed (attempt {attempt + 1}/"
+                        f"Topic delete failed (attempt {attempt + 1}/"
                         f"{DELETE_RETRIES}): {topic}: {e}"
                     )
-                    remaining.append((topic, subscription))
+                    remaining.append(topic)
 
             if not remaining:
                 return
 
-            queues = remaining
+            topics = remaining
 
             if attempt < DELETE_RETRIES - 1:
                 await asyncio.sleep(DELETE_RETRY_DELAY)
 
-        for topic, subscription in queues:
+        for topic in topics:
             logger.error(
-                f"Failed to delete queue after {DELETE_RETRIES} "
+                f"Failed to delete topic after {DELETE_RETRIES} "
                 f"attempts: {topic}"
             )
 
@@ -426,8 +424,8 @@ class FlowConfig:
                 result = result.replace(f"{{{param_name}}}", str(param_value))
             return result
 
-        # Collect queue identifiers before removing config
-        queues = self._collect_flow_queues(cls, repl_template)
+        # Collect topic identifiers before removing config
+        topics = self._collect_flow_topics(cls, repl_template)
 
         # Phase 1: Set status to "stopping" and remove processor config.
         # The config push tells processors to shut down their consumers.
@@ -448,8 +446,8 @@ class FlowConfig:
 
         await self.config.delete_many(deletes)
 
-        # Phase 2: Delete queues with retries, then remove the flow record.
-        await self._delete_queues(queues)
+        # Phase 2: Delete topics with retries, then remove the flow record.
+        await self._delete_topics(topics)
 
         if msg.flow_id in await self.config.keys("flow"):
             await self.config.delete("flow", msg.flow_id)
diff --git a/trustgraph-flow/trustgraph/flow/service/service.py b/trustgraph-flow/trustgraph/flow/service/service.py
index a3f2fb6b..e1997452 100644
--- a/trustgraph-flow/trustgraph/flow/service/service.py
+++ b/trustgraph-flow/trustgraph/flow/service/service.py
@@ -101,11 +101,9 @@ class Processor(AsyncProcessor):
 
     async def start(self):
 
-        await self.pubsub.ensure_queue(
-            self.flow_request_topic, self.flow_request_subscriber
-        )
+        await self.pubsub.ensure_topic(self.flow_request_topic)
         await self.config_client.start()
-        await self.flow.ensure_existing_flow_queues()
+        await self.flow.ensure_existing_flow_topics()
         await self.flow_request_consumer.start()
 
     async def on_flow_request(self, msg, consumer, flow):
diff --git a/trustgraph-flow/trustgraph/librarian/service.py b/trustgraph-flow/trustgraph/librarian/service.py
index 83f97bf3..ed005298 100755
--- a/trustgraph-flow/trustgraph/librarian/service.py
+++ b/trustgraph-flow/trustgraph/librarian/service.py
@@ -263,12 +263,8 @@ class Processor(AsyncProcessor):
 
     async def start(self):
 
-        await self.pubsub.ensure_queue(
-            self.librarian_request_topic, self.librarian_request_subscriber
-        )
-        await self.pubsub.ensure_queue(
-            self.collection_request_topic, self.collection_request_subscriber
-        )
+        await self.pubsub.ensure_topic(self.librarian_request_topic)
+        await self.pubsub.ensure_topic(self.collection_request_topic)
         await super(Processor, self).start()
         await self.librarian_request_consumer.start()
         await self.librarian_response_producer.start()
diff --git a/trustgraph-vertexai/trustgraph/model/text_completion/googleaistudio/llm.py b/trustgraph-vertexai/trustgraph/model/text_completion/googleaistudio/llm.py
index 9f431d20..142fc45c 100644
--- a/trustgraph-vertexai/trustgraph/model/text_completion/googleaistudio/llm.py
+++ b/trustgraph-vertexai/trustgraph/model/text_completion/googleaistudio/llm.py
@@ -18,6 +18,12 @@ import logging
 # Module logger
 logger = logging.getLogger(__name__)
 
+from google import genai
+from google.genai import types
+from google.genai.types import HarmCategory, HarmBlockThreshold
+from google.genai.errors import ClientError
+from google.api_core.exceptions import ResourceExhausted
+
 from .... exceptions import TooManyRequests
 from .... base import LlmService, LlmResult, LlmChunk
 
@@ -71,20 +77,20 @@ class Processor(LlmService):
         block_level = self.HarmBlockThreshold.BLOCK_ONLY_HIGH
 
         self.safety_settings = [
-            self.types.SafetySetting(                           
-                category = self.HarmCategory.HARM_CATEGORY_HATE_SPEECH,
+            types.SafetySetting(                           
+                category = HarmCategory.HARM_CATEGORY_HATE_SPEECH,
                 threshold = block_level,
             ),
-            self.types.SafetySetting(                           
-                category = self.HarmCategory.HARM_CATEGORY_HARASSMENT,
+            types.SafetySetting(                           
+                category = HarmCategory.HARM_CATEGORY_HARASSMENT,
                 threshold = block_level,
             ),
-            self.types.SafetySetting(                           
-                category = self.HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
+            types.SafetySetting(                           
+                category = HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT,
                 threshold = block_level,
             ),
-            self.types.SafetySetting(                           
-                category = self.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
+            types.SafetySetting(                           
+                category = HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
                 threshold = block_level,
             ),
             # There is a documentation conflict on whether or not
@@ -104,7 +110,7 @@ class Processor(LlmService):
 
         if cache_key not in self.generation_configs:
             logger.info(f"Creating generation config for '{model_name}' with temperature {effective_temperature}")
-            self.generation_configs[cache_key] = self.types.GenerateContentConfig( 
+            self.generation_configs[cache_key] = types.GenerateContentConfig( 
                 temperature = effective_temperature,
                 top_p = 1,
                 top_k = 40,
@@ -160,7 +166,7 @@ class Processor(LlmService):
             # Leave rate limit retries to the default handler
             raise TooManyRequests()
 
-        except self.ClientError as e:                 
+        except ClientError as e:                 
             # google-genai SDK throws ClientError for 4xx errors
             if e.code == 429:
                 logger.warning(f"Rate limit exceeded (ClientError 429): {e}")
@@ -229,11 +235,11 @@ class Processor(LlmService):
 
             logger.debug("Streaming complete")
 
-        except self.ResourceExhausted:              
+        except ResourceExhausted:              
             logger.warning("Rate limit exceeded during streaming")
             raise TooManyRequests()
 
-        except self.ClientError as e:                 
+        except ClientError as e:                 
             # google-genai SDK throws ClientError for 4xx errors
             if e.code == 429:
                 logger.warning(f"Rate limit exceeded during streaming (ClientError 429): {e}")