Feat/metadata search with postgres fts

[aaronjae22]

This PR adds Metadata search with Postgres FTS.

Closes #295

The new search method rank schemas by full-text relevance against query_text, with name weighted above description. When query_text is blank the queryset is returned unchanged, so the caller's existing ordering (e.g. alphabetical on the index page) is preserved for plain browsing.

class Schema(BaseModel):
    ...
    ...
    ...
    search_vector = models.GeneratedField(
        expression=(
            SearchVector("name", weight="A", config="english")
            + SearchVector("description", weight="B", config="english")
        ),
        output_field=SearchVectorField(),
        db_persist=True,
    )

    class Meta:
        indexes = [
            models.Index(fields=["published_at"]),
            GinIndex(fields=["search_vector"], name="schema_search_vec_gin"),
        ]

GeneratedField is a database-computed column. Postgres itself calculates the value from the expression. It makes no use of triggers, no manual .update(), no overriding save(). The database keeps it in sync automatically whenever name or description changes.

expression is the full-text search document. In this case just concatenates two SearchVectors, name with weight='A' and description with weight=B. So a match in the name ranks higher than a match in the description.

config="english" appliies English stemming/stop-words rules.

output_field=SearchVectorField() tells Django the generated column's type is a tsvector.

db_persist=True means the value is stored on disk (Postgres STORED generated column), not recomputed on every read. This is required for Postgres, and it’s what lets the GIN index work.

---

Row	name / description	Stored search_vector
A	Invoice Schema / A directory of Schemas	'directori':4B 'invoic':1A 'schema':2A,6B
B	ACME Format / None	'acm':1A 'format':2A

• ’invoic':1A — "Invoice" stemmed, position 1, weight A (from name).
• 'schema':2A,6B — this one entry is the whole feature. The stem schema appears at position 2 weight A (from name "Invoice Schema") and position 6 weight B (from description "...Schemas"). Singular and plural collapsed into one lexeme, with both origins retained.
• 'directori':4B — "directory" stemmed, weight B.
• Stop-words (A, of) are dropped, which is why positions skip.
• Row B proves NULL is safe: a None description contributed nothing and the vector is still valid.

[aaronjae22]

@alexbainter can you check this out when you have some time

[alexbainter]

This is great! Had one comment about a test that looks like it might be missing a couple assertions, but not a big deal. Looking forward to using this for #293!

[alexbainter]

Oh nice! I introduced a SchemaQuerySet on my branch for #293 too.

[aaronjae22]

Great!

[alexbainter]

The test title and comment seem to imply this should also verify the order of the results. Did you mean to do that too?

[aaronjae22]

Yes, I completely agree with you. I modified the test in order to check the plain browsing order in which they should get returned.

Thanks for checking that out.

[aaronjae22]

I updated the test as you suggested and I agreed with it.

This update should mirror what happens in our stg/prod in which the index view builds its base queryset with .order_by("name") before calling .search(), so an empty search box on the homepage browses schemas alphabetically, and only when we type something does .search() override that with .order_by("-rank", "name").

Thanks for noticing it. I'll move forward and merge this.