how can I get a query set from multiple levels of sets / foreign keys?

After you've done the prefetch, it seems like the only cheap way to access the child records is through all(). Any filters seem to trigger another database query.

The short answer to your question about all the paragraphs in a book is to use a list comprehension with two levels:

    paragraphs = [paragraph
                  for page in book.page_set.all()
                  for paragraph in page.paragraph_set.all()]

Here's a runnable example:

# Tested with Django 1.11.13
from __future__ import print_function
import os
import sys

import django
from django.apps import apps
from django.apps.config import AppConfig
from django.conf import settings
from django.core.files.base import ContentFile, File
from django.db import connections, models, DEFAULT_DB_ALIAS
from django.db.models.base import ModelBase

from django_mock_queries.mocks import MockSet, mocked_relations

NAME = 'udjango'


def main():
    setup()

    class Book(models.Model):
        name = models.CharField(max_length=64)

    class Page(models.Model):
        number = models.IntegerField()
        book = models.ForeignKey(Book)

    class Paragraph(models.Model):
        number = models.IntegerField()
        page = models.ForeignKey(Page)

    syncdb(Book)
    syncdb(Page)
    syncdb(Paragraph)

    b = Book.objects.create(name='Gone With The Wind')
    p = b.page_set.create(number=1)
    p.paragraph_set.create(number=1)
    b = Book.objects.create(name='The Three Body Problem')
    p = b.page_set.create(number=1)
    p.paragraph_set.create(number=1)
    p.paragraph_set.create(number=2)
    p = b.page_set.create(number=2)
    p.paragraph_set.create(number=1)
    p.paragraph_set.create(number=2)

    books = Book.objects.all().prefetch_related('page_set',
                                                'page_set__paragraph_set')

    for book in books:
        print(book.name)
        paragraphs = [paragraph
                      for page in book.page_set.all()
                      for paragraph in page.paragraph_set.all()]
        for paragraph in paragraphs:
            print(paragraph.page.number, paragraph.number)


def setup():
    DB_FILE = NAME + '.db'
    with open(DB_FILE, 'w'):
        pass  # wipe the database
    settings.configure(
        DEBUG=True,
        DATABASES={
            DEFAULT_DB_ALIAS: {
                'ENGINE': 'django.db.backends.sqlite3',
                'NAME': DB_FILE}},
        LOGGING={'version': 1,
                 'disable_existing_loggers': False,
                 'formatters': {
                    'debug': {
                        'format': '%(asctime)s[%(levelname)s]'
                                  '%(name)s.%(funcName)s(): %(message)s',
                        'datefmt': '%Y-%m-%d %H:%M:%S'}},
                 'handlers': {
                    'console': {
                        'level': 'DEBUG',
                        'class': 'logging.StreamHandler',
                        'formatter': 'debug'}},
                 'root': {
                    'handlers': ['console'],
                    'level': 'WARN'},
                 'loggers': {
                    "django.db": {"level": "DEBUG"}}})
    app_config = AppConfig(NAME, sys.modules['__main__'])
    apps.populate([app_config])
    django.setup()
    original_new_func = ModelBase.__new__

    @staticmethod
    def patched_new(cls, name, bases, attrs):
        if 'Meta' not in attrs:
            class Meta:
                app_label = NAME
            attrs['Meta'] = Meta
        return original_new_func(cls, name, bases, attrs)
    ModelBase.__new__ = patched_new


def syncdb(model):
    """ Standard syncdb expects models to be in reliable locations.

    Based on https://github.com/django/django/blob/1.9.3
    /django/core/management/commands/migrate.py#L285
    """
    connection = connections[DEFAULT_DB_ALIAS]
    with connection.schema_editor() as editor:
        editor.create_model(model)

main()

And here's the end of the output. You can see that it only runs one query for each table.

2018-10-30 15:58:25[DEBUG]django.db.backends.execute(): (0.000) SELECT "udjango_book"."id", "udjango_book"."name" FROM "udjango_book"; args=()
2018-10-30 15:58:25[DEBUG]django.db.backends.execute(): (0.000) SELECT "udjango_page"."id", "udjango_page"."number", "udjango_page"."book_id" FROM "udjango_page" WHERE "udjango_page"."book_id" IN (1, 2); args=(1, 2)
2018-10-30 15:58:25[DEBUG]django.db.backends.execute(): (0.000) SELECT "udjango_paragraph"."id", "udjango_paragraph"."number", "udjango_paragraph"."page_id" FROM "udjango_paragraph" WHERE "udjango_paragraph"."page_id" IN (1, 2, 3); args=(1, 2, 3)
Gone With The Wind
1 1
The Three Body Problem
1 1
1 2
2 1
2 2

further to Don's answer, you can use Prefetch objects to apply any filters you want, e.g:

from django.db import models, connection

def query():
    paragraph_filter = models.Prefetch(
        'page_set__paragraph_set',
        Paragraph.objects.filter(number__gt=1))

    books = Book.objects.all().prefetch_related(
        'page_set', paragraph_filter)

    for book in books:
        for page in book.page_set.all():
            for paragraph in page.paragraph_set.all():
                print(paragraph)

    print(connection.queries)

Django takes care of ensuring that all appropriate objects gets loaded in a small number of queries (one per table, so you'll get three queries)