diff --git a/vsb/databases/__init__.py b/vsb/databases/__init__.py
new file mode 100644
index 0000000..e47ac98
--- /dev/null
+++ b/vsb/databases/__init__.py
@@ -0,0 +1,19 @@
+from enum import Enum
+from .base import DB
+
+
+class Database(Enum):
+    """Set of supported database backends, the value is the string used to
+    specify via --database="""
+    Pinecone = "pinecone"
+    PGVector = "pgvector"
+
+    def build(self) -> DB:
+        """Construct an instance of """
+        match self:
+            case Database.Pinecone:
+                from .pinecone.pinecone import PineconeDB
+                return PineconeDB()
+            case Database.PGVector:
+                from .pgvector.pgvector import PGVectorDB
+                return PGVectorDB()
diff --git a/vsb/databases/base.py b/vsb/databases/base.py
new file mode 100644
index 0000000..370d3e3
--- /dev/null
+++ b/vsb/databases/base.py
@@ -0,0 +1,45 @@
+from abc import ABC, abstractmethod
+from enum import Enum, auto
+
+
+class Request(Enum):
+    Upsert = auto()
+    Search = auto()
+
+
+class Index(ABC):
+    """Abstract class with represents an index or one or more vector records.
+    Specific implementations should subclass this and implement all abstract
+    methods.
+    Instance of this (derived) class are typically created via the corresponding
+    (concrete) DB create_index() method.
+    """
+
+    @abstractmethod
+    def upsert(self, ident, vector, metadata):
+        raise NotImplementedError
+
+    @abstractmethod
+    def search(self, query_vector):
+        raise NotImplementedError
+
+    def do_request(self, request):
+        print(f"Got request: {request}")
+        match request.operation:
+            case Request.Upsert:
+                self.upsert(request.id, request.vector, request.metadata)
+                return
+            case Request.Search:
+                response = self.search(request.q_vector)
+                # Record timing, calculate Recall etc.
+
+
+class DB(ABC):
+    """Abstract class which represents a database which can store vector
+    records in one or more Indexes. Specific Vector DB implementations should
+    subclass this and implement all abstract methods.
+    """
+
+    @abstractmethod
+    def create_index(self, tenant: str) -> Index:
+        raise NotImplementedError
diff --git a/vsb/databases/pgvector/pgvector.py b/vsb/databases/pgvector/pgvector.py
new file mode 100644
index 0000000..e64b49c
--- /dev/null
+++ b/vsb/databases/pgvector/pgvector.py
@@ -0,0 +1,12 @@
+from ..vectordb import VectorDB
+
+
+class PGVectorDB(VectorDB):
+    def __init__(self):
+        print("PGVectorDB::__init__")
+
+    def upsert(self, ident, vector, metadata):
+        pass
+
+    def search(self, query_vector):
+        pass
diff --git a/vsb/databases/pinecone/pinecone.py b/vsb/databases/pinecone/pinecone.py
new file mode 100644
index 0000000..095cc73
--- /dev/null
+++ b/vsb/databases/pinecone/pinecone.py
@@ -0,0 +1,21 @@
+from ..base import DB, Index
+
+
+class PineconeIndex(Index):
+    def __init__(self, tenent: str):
+        pass
+
+    def upsert(self, ident, vector, metadata):
+        pass
+
+    def search(self, query_vector):
+        pass
+
+
+class PineconeDB(DB):
+    def __init__(self):
+        print("PineconeDB::__init__")
+
+    def create_index(self, tenant: str) -> Index:
+        return PineconeIndex()
+
diff --git a/vsb/vsb.py b/vsb/vsb.py
new file mode 100755
index 0000000..c807586
--- /dev/null
+++ b/vsb/vsb.py
@@ -0,0 +1,112 @@
+#!/usr/bin/env python3
+
+from enum import Enum, auto
+from locust import User, events, task, TaskSet
+from locust.env import Environment
+from locust.exception import StopUser
+from locust.log import setup_logging
+from locust.runners import WorkerRunner
+from locust.stats import stats_history, stats_printer
+
+from databases import Database
+from workloads import Workload
+import argparse
+import gevent
+import logging
+import sys
+
+
+setup_logging("INFO")
+
+
+class VectorSearchUser(User):
+    class Phase(Enum):
+        Load = auto()
+        Run = auto()
+
+    """Represents a single user (aka client) performing requests against
+    a particular Backend."""
+    def __init__(self, environment):
+        super().__init__(environment)
+        self.count = 0
+        self.database = Database(environment.options.database).build()
+        self.workload = Workload(environment.options.workload).build()
+        self.phase = VectorSearchUser.Phase.Load
+
+    @task
+    def request(self):
+        match self.phase:
+            case VectorSearchUser.Phase.Load:
+                self.do_load()
+            case VectorSearchUser.Phase.Run:
+                self.do_run()
+
+    def do_load(self):
+        if batch := self.workload.next_record_batch():
+            print(f"Batch: {batch}")
+            print(f"Loading batch of size:", len(batch))
+            self.database.upsert(batch)
+        else:
+            # No more data to load, advance to Run phase.
+            self.phase = VectorSearchUser.Phase.Run
+
+    def do_run(self):
+        if self.workload.execute_next_request(self.workload):
+            print(f"Issue request {self.count}: request")
+        else:
+            runner = self.environment.runner
+            logging.info(f"User count: {runner.user_count}")
+            if runner.user_count == 1:
+                logging.info("Last user stopped, quitting runner")
+                if isinstance(runner, WorkerRunner):
+                    runner._send_stats()  # send a final report
+                # need to trigger this in a separate greenlet, in case test_stop handlers do something async
+                gevent.spawn_later(0.1, runner.quit)
+            raise StopUser()
+
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        prog='VCB',
+        description='Vector Search Bench')
+    parser.add_argument("--database", required=True,
+                        choices=tuple(e.value for e in Database))
+    parser.add_argument("--workload", required=True,
+                        choices=tuple(e.value for e in Workload))
+
+    options = parser.parse_args()
+
+    # setup Environment and Runner
+    env = Environment(user_classes=[VectorSearchUser], events=events)
+    env.options = options
+
+    runner = env.create_local_runner()
+
+    # start a WebUI instance
+    #web_ui = env.create_web_ui("127.0.0.1", 8089)
+
+    # execute init event handlers (only really needed if you have registered any)
+    env.events.init.fire(environment=env, runner=runner)# , web_ui=web_ui)
+
+    # start a greenlet that periodically outputs the current stats
+    gevent.spawn(stats_printer(env.stats))
+
+    # start a greenlet that save current stats to history
+    gevent.spawn(stats_history, env.runner)
+
+    # start the test
+    runner.start(1, spawn_rate=10)
+
+    # in 30 seconds stop the runner
+    gevent.spawn_later(30, runner.quit)
+
+    # wait for the greenlets
+    runner.greenlet.join()
+
+    # stop the web server for good measures
+#    web_ui.stop()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/vsb/workloads/__init__.py b/vsb/workloads/__init__.py
new file mode 100644
index 0000000..4ab9632
--- /dev/null
+++ b/vsb/workloads/__init__.py
@@ -0,0 +1,17 @@
+from enum import Enum
+from .base import VectorWorkload
+
+
+class Workload(Enum):
+    """Set of supported workloads, the value is the string used to
+    specify via --benchmark=
+    """
+
+    MNIST = "mnist"
+
+    def build(self) -> VectorWorkload:
+        """Construct an instance of Benchmark based on the value of the enum"""
+        match self:
+            case Workload.MNIST:
+                from .mnist.mnist import MNIST
+                return MNIST()
diff --git a/vsb/workloads/base.py b/vsb/workloads/base.py
new file mode 100644
index 0000000..9f42b13
--- /dev/null
+++ b/vsb/workloads/base.py
@@ -0,0 +1,19 @@
+from abc import ABC, abstractmethod
+
+
+class VectorWorkload(ABC):
+    @abstractmethod
+    def next_record_batch(self):
+        """
+        For initial dataset ingest, returns the next
+        :return:
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def execute_next_request(self, db: 'DB') -> bool:
+        """Obtain the next request for this workload and execute against the given
+        database. Returns true if execution should continue after this request,
+        or false if the workload is complete.
+        """
+        raise NotImplementedError
diff --git a/dataset.py b/vsb/workloads/dataset.py
similarity index 100%
rename from dataset.py
rename to vsb/workloads/dataset.py
diff --git a/vsb/workloads/mnist/mnist.py b/vsb/workloads/mnist/mnist.py
new file mode 100644
index 0000000..e9323b9
--- /dev/null
+++ b/vsb/workloads/mnist/mnist.py
@@ -0,0 +1,23 @@
+from ..base import VectorWorkload
+from ..dataset import Dataset
+
+
+class MNIST(VectorWorkload):
+    def __init__(self):
+        print("MNIST::__init__")
+        self.dataset = Dataset(name="mnist")
+        self.dataset.load_documents()
+        print(self.dataset.documents)
+        self.records = Dataset.split_dataframe(self.dataset.documents, 100)
+        print(self.records)
+        self.operation_count = 0
+        self.operation_limit = 10
+
+    def next_record_batch(self):
+        print("MNIST::next_record_batch")
+
+    def execute_next_request(self, db) ->bool:
+        print("MNIST::execute_next_request")
+        self.operation_count += 1
+        return self.operation_count < self.operation_limit
+