Validate untrusted payloads, collect every error at once, coerce loose types — no dependencies.
Every backend receives data it does not control: request bodies, webhook callbacks,
uploaded config, rows scraped from someone else's page. Hand-rolled
if "email" in body and isinstance(body["email"], str) checks rot fast, stop at the
first problem, and tell the caller nothing useful.
Slate describes the shape once, as a class, and reports all problems at once.
from slate import Model, Str, Int, Email
class User(Model):
name = Str(min_len=2) # Field instances, not type annotations
age = Int()
email = Email()
user = User.parse({"name": "ada", "age": "36", "email": "ada@example.com"})
print(user.name, user.age, type(user.age).__name__)
# ada 36 int <- "36" was coerced
print(user.dump())
# {'name': 'ada', 'age': 36, 'email': 'ada@example.com'}parse() raises on failure. validate() never raises — it returns a result you can
inspect, and it collects every error rather than stopping at the first:
r = User.validate({"name": "x", "age": "nope", "email": "not-an-email"})
print(r.valid) # False
for e in r.errors:
print(f"{e.path}: {e.message}")name: length 1 is less than minimum 2
age: not a valid integer
email: string does not match pattern ^[^@\s]+@[^@\s]+\.[^@\s]+$
Three errors, one pass. A form can highlight every bad field instead of making the user fix them one reload at a time.
| Scalars | Containers |
|---|---|
Str Int Float Bool Const Any Nil |
List Dict Tuple |
Enum DateTime Email |
Optional Union Nested |
Nested(OtherModel) validates an embedded mapping and rewrites the error path, so
address.zip reports as address.zip, not zip.
from slate import TypeAdapter, Int, List
TypeAdapter(Int()).parse("42") # 42
TypeAdapter(List(int)).parse(["1", "2"]) # [1, 2]
TypeAdapter(Int()).validate("x").valid # Falsegraph LR
D[untrusted dict] --> M[Model.validate]
M --> F1[field: Str]
M --> F2[field: Int]
M --> F3[field: Email]
F1 --> C{coerce}
F2 --> C
F3 --> C
C -->|ok| V[ValidationResult.valid=True<br/>value = Model instance]
C -->|fail| E[ValidationResult.valid=False<br/>errors = 每字段一条]
V --> DUMP[.dump 可 JSON 序列化]
Every field is checked independently, so one bad value does not hide the others.
Drop the slate/ folder into your project. No pip install, no supply chain.
python -m pytest -q- Coercion is lenient by design, and that is a policy decision.
"36"becomes36;"yes"becomesTrue. If you need strict typing, tighten individual fields — the default favours accepting messy real-world input over rejecting it. - No JSON Schema export. The grammar is Python-side only;
schema()describes field shapes for debugging, not a portable contract. - No async or streaming validation. It validates a whole object in memory.
- Performance is not benchmarked. No throughput numbers are published here, because none were measured.
MIT