A pythonic full-text search - DjangoCon US 2022

-- A talk I gave at DjangoCon US 2022

Keeping in mind the pythonic principle that “simple is better than complex” we’ll see how to implement full-text search in a web service using only latest versions of Django and PostgreSQL and we’ll analyze the advantages compared to more complex solutions based on external services.

https://www.paulox.net/2022/10/19/djangocon-us-2022/

Transcript

1. A PYTHONIC
   FULL-TEXT SEARCH
   PAOLO MELCHIORRE ~ @pauloxnet
   

   View Slide

2. View Slide

3. Paolo Melchiorre ~ @pauloxnet
   @pauloxnet
   • CTO @ 20tab
   • Software engineer
   • Python developer
   • Django contributor
   Paolo Melchiorre
   3
   DjangoCon Europe 2019 - Bartek Pawlik (CC BY-NC-SA)
   

   View Slide

4. Paolo Melchiorre ~ @pauloxnet
   4
   Pythonic
   >>> import this
   The Zen of Python, by Tim Peters
   Beautiful is better than ugly.
   Explicit is better than implicit.
   Simple is better than complex.
   Complex is better than complicated.
   ...
   

   View Slide

5. Paolo Melchiorre ~ @pauloxnet
   5
   Full-text search
   “… techniques for searching
   … computer-stored document …
   in a full-text database.”
   — “Full-text search”, Wikipedia
   

   View Slide

6. Paolo Melchiorre ~ @pauloxnet
   6
   Popular search engines
   

   View Slide

7. Paolo Melchiorre ~ @pauloxnet
   7
   External search engines
   PROS
   Popular
   Full featured
   Resources
   CONS
   Driver
   Query language
   Synchronization
   

   View Slide

8. Paolo Melchiorre ~ @pauloxnet
   8
   External search engines synchronization
   

   View Slide

9. Paolo Melchiorre ~ @pauloxnet
   PostgreSQL
   9
   Photo by Nam Anh on Unsplash
   Elephant walking during daytime 2019
   

   View Slide

10. Paolo Melchiorre ~ @pauloxnet
    10
    PostgreSQL
    Full text search (v8.3 ~2008)
    Data type (tsquery, tsvector)
    Special indexes (GIN, GiST)
    Phrase search (v9.6 ~2016)
    JSON support (v10 ~2017)
    Web search (v11 ~2018)
    New languages (v12-14 ~2019-2021)
    

    View Slide

11. Paolo Melchiorre ~ @pauloxnet
    11
    Document
    “… the unit of searching
    in a full-text search system;
    e.g., a magazine article …”
    — “Full Text Search”, PostgreSQL Documentation
    

    View Slide

12. View Slide

13. Paolo Melchiorre ~ @pauloxnet
    Django
    13
    William Gottlieb (Public domain)
    Django Reinhardt at the Aquarium jazz club in New York, NY 1946
    

    View Slide

14. Paolo Melchiorre ~ @pauloxnet
    14
    Django
    Full text search (v1.10 ~2016)
    django.contrib.postgres
    Fields, expressions, functions
    GIN index (v1.11 ~2017)
    GiST index (v2.0 ~2018)
    Phrase search (v2.2 ~2019)
    Web search (v3.1 ~2020)
    

    View Slide

15. Paolo Melchiorre ~ @pauloxnet
    15
    Document-based search
    • Weighting
    • Categorization
    • Highlighting
    • Multiple languages
    

    View Slide

16. View Slide

17. Paolo Melchiorre ~ @pauloxnet
    """Blogs models."""
    from django.contrib.postgres import search
    from django.db import models
    class Author(models.Model):
    name = models.CharField(max_length=200)
    class Entry(models.Model):
    headline = models.CharField(max_length=255)
    body_text = models.TextField()
    authors = models.ManyToManyField(Author)
    17
    

    View Slide

18. Paolo Melchiorre ~ @pauloxnet
    """Field lookups."""
    from blog.models import Author
    Author.objects.filter(name__contains="Jerry")
    []
    Author.objects.filter(name__icontains="TERRY")
    [, ]
    18
    

    View Slide

19. Paolo Melchiorre ~ @pauloxnet
    """Application definition."""
    INSTALLED_APPS = [
    "django.contrib.admin",
    "django.contrib.auth",
    "django.contrib.contenttypes",
    "django.contrib.sessions",
    "django.contrib.messages",
    "django.contrib.staticfiles",
    "django.contrib.postgres",
    "blog",
    ]
    19
    

    View Slide

20. Paolo Melchiorre ~ @pauloxnet
    """Trigram extension."""
    from django.contrib.postgres import operations
    from django.db import migrations
    class Migration(migrations.Migration):
    operations = [operations.TrigramExtension()]
    20
    

    View Slide

21. Paolo Melchiorre ~ @pauloxnet
    """Trigram similar lookup."""
    from blog.models import Author
    Author.objects.filter(name__trigram_similar="jerry ones")
    [, ]
    21
    

    View Slide

22. Paolo Melchiorre ~ @pauloxnet
    """Search lookup."""
    from blog.models import Entry
    Entry.objects.filter(headline__search="any cheeses")
    []
    22
    

    View Slide

23. Paolo Melchiorre ~ @pauloxnet
    """SearchVector function."""
    from django.contrib.postgres import search
    from blog.models import Entry
    V = search.SearchVector("headline", "body_text")
    entries = Entry.objects.annotate(search=V)
    entries.filter(search="cheeses")
    [, ]
    23
    

    View Slide

24. Paolo Melchiorre ~ @pauloxnet
    """SearchQuery expression."""
    from django.contrib.postgres import search
    from blog.models import Entry
    V = search.SearchVector("headline", "body_text")
    Q = search.SearchQuery("cheese -top", search_type="websearch")
    entries = Entry.objects.annotate(search=V)
    entries.filter(search=Q)
    []
    24
    

    View Slide

25. Paolo Melchiorre ~ @pauloxnet
    """SearchConfig expression."""
    from django.contrib.postgres import search
    from blog.models import Entry
    V = search.SearchVector("body_text", config="french")
    Q = search.SearchQuery("œuf", config="french")
    entries = Entry.objects.annotate(search=V)
    entries.filter(search=Q)
    []
    25
    

    View Slide

26. Paolo Melchiorre ~ @pauloxnet
    """SearchRank function."""
    from django.contrib.postgres import search
    from blog.models import Entry
    V = search.SearchVector("headline","body_text")
    Q = search.SearchQuery("cheese")
    R = search.SearchRank(V, Q)
    entries = Entry.objects.annotate(rank=R).order_by("-rank")
    entries.filter(rank__gt=0.05).values_list("headline", "rank")
    [('Cheese on Toast recipes', 0.09), ('Pizza Recipes', 0.06)]
    26
    

    View Slide

27. Paolo Melchiorre ~ @pauloxnet
    """SearchVector weight attribute."""
    from django.contrib.postgres import search
    from blog.models import Entry
    V = search.SearchVector("headline", weight="A")
    W = search.SearchVector("body_text", weight="B")
    Q = search.SearchQuery("cheese")
    R = search.SearchRank(V + W, Q)
    entries = Entry.objects.annotate(rank=R).order_by("-rank")
    entries.filter(rank__gt=0.05).values_list("headline", "rank")
    [('Cheese on Toast recipes', 0.72), ('Pizza Recipes', 0.24)]
    27
    

    View Slide

28. Paolo Melchiorre ~ @pauloxnet
    """SearchHeadline function."""
    from django.contrib.postgres import search
    from blog.models import Entry
    V = search.SearchVector("headline","body_text")
    Q = search.SearchQuery("cheeses")
    H = search.SearchHeadline("headline", Q)
    entries = Entry.objects.annotate(search=V, highlight=H)
    entries.filter(search=Q).values_list("highlight", flat=True)
    ["Cheese on Toast recipes", "Pizza Recipes"]
    28
    

    View Slide

29. Paolo Melchiorre ~ @pauloxnet
    """SearchVector functional index."""
    from django.contrib.postgres import indexes, search
    from django.db import models
    V = search.SearchVector("headline", config="english")
    class Entry(models.Model):
    headline = models.CharField(max_length=255)
    body_text = models.TextField()
    authors = models.ManyToManyField("Author")
    class Meta:
    indexes = [indexes.GinIndex(V, name="search_idx")]
    29
    

    View Slide

30. Paolo Melchiorre ~ @pauloxnet
    """SearchVector functional index query."""
    from django.contrib.postgres import search
    from blog.models import Entry
    V = search.SearchVector("headline", config="english")
    entries = Entry.objects.annotate(search=V)
    entries.filter(search="cheeses")
    []
    30
    

    View Slide

31. Paolo Melchiorre ~ @pauloxnet
    """SearchVector field."""
    from django.contrib.postgres import search
    from django.db import models
    class Entry(models.Model):
    headline = models.CharField(max_length=255)
    body_text = models.TextField()
    authors = models.ManyToManyField("Author")
    search_vector = search.SearchVectorField()
    class Meta:
    indexes = [indexes.GinIndex(fields=["search_vector"])]
    31
    

    View Slide

32. Paolo Melchiorre ~ @pauloxnet
    """SearchVector field update."""
    from django.contrib.postgres import aggregates as agg, search
    from django.db import models
    from blog.models import Entry
    V = search.SearchVector("headline", "body_text")
    W = search.SearchVector(agg.StringAgg("authors__name", " "))
    entries = Entry.objects.filter(id=models.OuterRef("id"))
    results = entries.annotate(search=V + W).values("search")[:1]
    Entry.objects.update(search_vector=models.Subquery(results))
    3
    32
    

    View Slide

33. Paolo Melchiorre ~ @pauloxnet
    """SearchVector field query."""
    from blog.models import Entry
    Entry.objects.filter(search_vector="cheeses")
    [, ]
    33
    

    View Slide

34. View Slide

35. Paolo Melchiorre ~ @pauloxnet
    35
    An old search
    • English-only search
    • HTML tag in results
    • Sphinx generation
    • PostgreSQL database
    • External search engine
    

    View Slide

36. Paolo Melchiorre ~ @pauloxnet
    36
    Django developers feedback
    PROS
    Maintenance
    Light setup
    Dogfooding
    CONS
    Work to do
    Features
    DB workload
    

    View Slide

37. Paolo Melchiorre ~ @pauloxnet
    EuroPython Sprints 2017
    37
    Paolo Melchiorre (CC BY-SA)
    

    View Slide

38. View Slide

39. C
    A
    B
    D
    A
    

    View Slide

40. Paolo Melchiorre ~ @pauloxnet
    """Django documentation search document definition."""
    from django.contrib.postgres.search import SearchVector as V
    from django.db.models import F
    from django.db.models.fields.json import KeyTextTransform as K
    DOCUMENT_SEARCH_VECTOR = (
    V('title', weight='A', config=F('config')) +
    V(K('slug', 'metadata'), weight='A', config=F('config')) +
    V(K('toc', 'metadata'), weight='B', config=F('config')) +
    V(K('body', 'metadata'), weight='C', config=F('config')) +
    V(K('parents', 'metadata'), weight='D', config=F('config'))
    )
    40
    

    View Slide

41. Paolo Melchiorre ~ @pauloxnet
    41
    djangoproject.com
    Full-text search features
    • Multilingual
    • PostgreSQL-only
    • Clean results
    • Low maintenance
    • Easier to setup
    • Web search syntax
    

    View Slide

42. View Slide

43. Paolo Melchiorre ~ @pauloxnet
    • Irish
    • Italian
    • Lithuanian
    • Nepali
    • Norwegian
    • Portuguese
    • Romanian
    • Finnish
    • French
    • German
    • Greek
    • Hindi
    • Hungarian
    • Indonesian
    • Arabic
    • Armenian
    • Basque
    • Catalan
    • Danish
    • Dutch
    • English
    43
    DjangoProject.com supported languages*
    • Russian
    • Serbian
    • Spanish
    • Swedish
    • Tamil
    • Turkish
    • Yiddish
    

    View Slide

44. Paolo Melchiorre ~ @pauloxnet
    44
    What’s next
    • Misspelling support
    • Search suggestions
    • Search statistics
    • Autocomplete
    • ...
    

    View Slide

45. Paolo Melchiorre ~ @pauloxnet
    45
    Tips
    • docs in djangoproject.com
    • details in postgresql.org
    • source code in github.com
    • questions in stackoverflow.com
    

    View Slide

46. Paolo Melchiorre ~ @pauloxnet
    46
    License
    CC BY-SA 4.0
    This work is licensed under
    a Creative Commons
    Attribution-ShareAlike 4.0
    International License.
    

    View Slide

47. View Slide

48. @20tab
    20tab
    20tab
    info@20tab.com
    20tab.com
    

    View Slide

49. @pauloxnet
    paolomelchiorre
    pauloxnet
    paolo@melchiorre.org
    paulox.net
    

    View Slide

50. Paolo Melchiorre ~ @pauloxnet
    • Participants
    • Speakers
    • Organizers
    • Volunteers from all conferences
    50
    Thanks
    Grazie /ˈɡrat.t
    ͡ sje/
    DjangoCon US 2022 - Paolo Melchiorre (CC BY-SA)
    

    View Slide