dyff-schema 0.23.0__py3-none-any.whl → 0.24.1__py3-none-any.whl

dyff/schema/v0/r1/adapters.py

@@ -6,14 +6,21 @@ from __future__ import annotations
 
 import functools
 import json
-from typing import Any, Iterable, NamedTuple, Protocol, Type
+import operator
+import re
+from typing import Any, Callable, Iterable, Literal, NamedTuple, Protocol, Type
 
 import jsonpath_ng as jsonpath
 from jsonpath_ng.exceptions import JSONPathError
+from jsonpath_ng.ext.parser import parse as jsonpath_parse_ext
 
 from dyff.schema.platform import SchemaAdapter
 
 
+def _json_deep_copy(data):
+    return json.loads(json.dumps(data))
+
+
 def map_structure(fn, data):
     """Given a JSON data structure ``data``, create a new data structure instance with
     the same shape as ``data`` by applying ``fn`` to each "leaf" value in the nested
@@ -70,90 +77,346 @@ class Adapter(Protocol):
         raise NotImplementedError()
 
 
+class _Literal:
+    def __init__(self, value):
+        self.value = value
+
+    def __call__(self, x):
+        return self.value
+
+
+class _Func_findall:
+    def __init__(self, *, pattern: str, flags: int = 0):
+        self.pattern = pattern
+        self.flags = flags
+
+    def __call__(self, x) -> list[str]:
+        return re.findall(self.pattern, x, self.flags)
+
+
+class _Func_join:
+    def __init__(self, *, separator: str = ""):
+        self._separator = separator
+
+    def __call__(self, x: list[str]) -> str:
+        return self._separator.join(x)
+
+
+class _Func_list:
+    def __call__(self, x) -> list:
+        return list(x)
+
+
+class _Func_reduce:
+    def __call__(self, x):
+        return functools.reduce(operator.add, x)
+
+
+class _Func_search:
+    def __init__(
+        self,
+        *,
+        pattern: str,
+        flags: int = 0,
+        group: int = 0,
+        default: str | None = None,
+    ):
+        self.pattern = pattern
+        self.flags = flags
+        self.group = group
+        self.default = default
+
+    def __call__(self, x) -> str | None:
+        m = re.search(self.pattern, x, self.flags)
+        return self.default if m is None else m.group(self.group)
+
+
+class _Func_split:
+    def __init__(self, *, pattern: str, maxsplit: int = 0, flags: int = 0):
+        self.pattern = pattern
+        self.maxsplit = maxsplit
+        self.flags = flags
+
+    def __call__(self, x) -> list[str]:
+        return re.split(self.pattern, x, self.maxsplit, self.flags)
+
+
+class _Func_sub:
+    def __init__(self, *, pattern: str, repl: str, count: int = 0, flags: int = 0):
+        self.pattern = pattern
+        self.repl = repl
+        self.count = count
+        self.flags = flags
+
+    def __call__(self, x) -> str:
+        return re.sub(self.pattern, self.repl, x, self.count, self.flags)
+
+
+class _Value_jsonpath:
+    def __init__(self, expr, *, kind: Literal["scalar", "list"] = "scalar"):
+        self._expr: jsonpath.JSONPath = jsonpath.parse(expr)
+        self._kind = kind
+
+    def __call__(self, x):
+        results = self._expr.find(x)
+        if self._kind == "list":
+            return [result.value for result in results]
+        elif self._kind == "scalar":
+            if len(results) == 0:
+                raise ValueError(f"no match for '{self._expr}' in '{x}'")
+            elif len(results) > 1:
+                raise ValueError(f"multiple results for '{self._expr}' in '{x}'")
+            return results[0].value
+        else:
+            raise AssertionError(f"kind {self._kind}")
+
+
+class _Value_list:
+    def __init__(self, exprs: list[Callable]):
+        self._exprs = exprs
+
+    def __call__(self, x) -> list:
+        return [e(x) for e in self._exprs]
+
+
+def _maybe_value_expr(expr: dict) -> Callable | None:
+    kinds = ["$literal", "$scalar", "$list"]
+    maybe_exprs = {k: expr.get(k) for k in kinds}
+    just_exprs = [k for k in kinds if maybe_exprs[k] is not None]
+    if len(just_exprs) == 0:
+        return None
+    if len(just_exprs) > 1:
+        raise ValueError(f"must specify exactly one of {kinds}: got {just_exprs}")
+
+    # remove sigil
+    kind: Literal["literal", "scalar", "list"] = just_exprs[0][1:]  # type: ignore
+    value = maybe_exprs[just_exprs[0]]
+    if kind == "literal":
+        return _Literal(value)
+
+    op: Callable = _Literal(value)
+    if isinstance(value, str):
+        if value.startswith("$"):
+            if value.startswith("$$"):
+                # Literal string -- remove "escape" character
+                op = _Literal(value[1:])
+            else:
+                op = _Value_jsonpath(value, kind=kind)
+    elif kind == "list" and isinstance(value, list):
+        exprs = [_maybe_value_expr(e) for e in value]
+        if any(e is None for e in exprs):
+            raise ValueError("$list elements must be value expressions")
+        op = _Value_list(exprs)  # type: ignore
+    if isinstance(op, _Literal) and kind != "literal":
+        raise ValueError("must use $literal when providing a literal value")
+    return op
+
+
+class _LeafExpression:
+    FUNCTIONS = {
+        "findall": _Func_findall,
+        "join": _Func_join,
+        "list": _Func_list,
+        "reduce": _Func_reduce,
+        "search": _Func_search,
+        "split": _Func_split,
+        "sub": _Func_sub,
+    }
+
+    def __init__(self, pipeline: dict | list[dict]):
+        if isinstance(pipeline, dict):
+            pipeline = [pipeline]
+
+        self._compiled_pipeline: list[Callable] = []
+        for step in pipeline:
+            if (value_op := _maybe_value_expr(step)) is not None:
+                self._compiled_pipeline.append(value_op)
+            elif (func := step.pop("$func", None)) is not None:
+                self._compiled_pipeline.append(_LeafExpression.FUNCTIONS[func](**step))
+            else:
+                raise ValueError(f"invalid $compute step: {step}")
+
+    def __call__(self, x):
+        output = None
+        for i, step in enumerate(self._compiled_pipeline):
+            if i == 0:
+                output = step(x)
+            else:
+                output = step(output)
+        return output
+
+
 class TransformJSON:
-    """Transform an input JSON structure by creating a new output JSON structure where
-    all of the "leaf" values are populated by either:
+    """Create a new JSON structure where the "leaf" values are populated by the results
+    of transformation functions applied to the input.
+
+    The "value" for each leaf can be::
+
+        1. A JSON literal value, or
+        2. The result of a jsonpath query on the input structure, or
+        3. The result of a computation pipeline starting from (1) or (2).
 
-        1. A provided JSON literal value, or
-        2. The result of a jsonpath query on the input structure.
+    To distinguish the specifications of leaf values from the specification of
+    the output structure, we apply the following rules::
 
-    For example, if the ``output_structure`` parameter is::
+        1. Composite values (``list`` and ``dict``) specify the structure of
+        the output.
+        2. Scalar values are output as-is, unless they are strings containing
+        JSONPath queries.
+        3. JSONPath queries are strings beginning with '$'. They are replaced
+        by the result of the query.
+        4. A ``dict`` containing the special key ``"$compute"`` introduces a
+        "compute context", which computes a leaf value from the input data.
+        Descendents of this key have "compute context semantics", which are
+        different from the "normal" semantics.
+
+    For example, if the ``configuration`` is::
 
         {
             "id": "$.object.id",
             "name": "literal",
             "children": {"left": "$.list[0]", "right": "$.list[1]"}
+            "characters": {
+                "letters": {
+                    "$compute": [
+                        {"$scalar": "$.object.id"},
+                        {
+                            "$func": "sub",
+                            "pattern": "[A-Za-z]",
+                            "repl": "",
+                        },
+                        {"$func": "list"}
+                    ]
+                }
+            }
         }
 
     and the data is::
 
         {
-            "object": {"id": 42, "name": "spam"},
+            "object": {"id": "abc123", "name": "spam"},
             "list": [1, 2]
         }
 
-    Then applying the transformer to the data will result in the new structure::
+    Then applying the transformation to the data will result in the new structure::
 
         {
-            "id": 42,
+            "id": "abc123",
             "name": "literal",
-            "children: {"left": 1, "right": 2}
+            "children: {"left": 1, "right": 2},
+            "characters": {
+                "letters": ["a", "b", "c"]
+            }
         }
 
-    A value is interpreted as a jsonpath query if it is a string that starts
-    with the '$' character. If you need a literal string that starts with
-    the '$' character, escape it with a second '$', e.g., "$$PATH" will appear
-    as the literal string "$PATH" in the output.
-
-    All of the jsonpath queries must return *exactly one value* when executed
-    against each input item. If not, a ``ValueError`` will be raised.
+    The ``.characters.letters`` field was derived by::
+
+        1. Extracting the value of the ``.object.id`` field in the input
+        2. Applying ``re.sub(r"[A-Za-z]", "", _)`` to the result of (1)
+        3. Applying ``list(_)`` to the result of (2)
+
+    Notice that descendents of the ``$compute`` key no longer describe the
+    structure of the output, but instead describe steps of the computation.
+    The value of ``"$compute"`` can be either an object or a list of objects.
+    A list is interpreted as a "pipeline" where each step is applied to the
+    output of the previous step.
+
+    Implicit queries
+    ================
+
+    Outside of the ``$compute`` context, string values that start with a ``$``
+    character are interpreted as jsonpath queries. Queries in this context must
+    return **exactly one value**, otherwise a ``ValueError`` will be raised.
+    This is because when multiple values are returned, there's no way to
+    distinguish a scalar-valued query that found 1 scalar from a list-valued
+    query that found a list with 1 element. In the ``$compute`` context, you
+    can specify which semantics you want.
+
+    If you need a literal string that starts with the '$' character, escape it
+    with a second '$', e.g., "$$PATH" will appear as the literal string "$PATH"
+    in the output. This works for both keys and values, e.g.,
+    ``{"$$key": "$$value"}`` outputs ``{"$key": "$value"}``. All keys that
+    begin with ``$`` are reserved, and you must always escape them.
+
+    The $compute context
+    ====================
+
+    A ``$compute`` context is introduced by a ``dict`` that contains the key
+    ``{"$compute": ...}``. Semantics in the ``$compute`` context are different
+    from semantics in the "normal" context.
+
+    $literal vs. $scalar vs. $list
+    ------------------------------
+
+    Inside a ``$compute`` context, we distinguish explicitly between literal
+    values, jsonpath queries that return scalars, and jsonpath queries that
+    return lists. You specify which semantics you intend by using
+    ``{"$literal": [1, 2]}``, ``{"$scalar": "$.foo"}``, or ``{"$list": $.foo[*]}``.
+    Items with ``$literal`` semantics are **never** interpreted as jsonpath
+    queries, even if they start with ``$``. In the ``$literal`` context, you
+    **should not** escape the leading ``$`` character.
+
+    A ``$scalar`` query has the same semantiics as a jsonpath query outside
+    of the ``$compute`` context, i.e., it must return exactly 1 item.
+    A ``$list`` query will return a list, which can be empty. Scalar-valued
+    queries in a ``$list`` context will return a list with 1 element.
+
+    $func
+    -----
+
+    You use blocks with a ``$func`` key to specify computation steps. The
+    available functions are: ``findall``, ``join``, ``list``, ``reduce``,
+    ``search``, ``split``, ``sub``. These behave the same way as the
+    corresponding functions from the Python standard library::
+
+        * ``findall``, ``search``, ``split``, and ``sub`` are from the
+        ``re`` module.
+        * ``reduce`` uses the ``+`` operator with no starting value; it will
+        raise an error if called on an empty list.
+
+    All of these functions take named parameters with the same names and
+    semantics as their parameters in Python.
     """
 
     def __init__(self, configuration: dict):
-        """
-        Parameters:
-            ``output_structure``: A JSON object where all the "leaf" values
-                are strings containing jsonpath queries.
-        """
         if configuration != json.loads(json.dumps(configuration)):
             raise ValueError("configuration is not valid JSON")
-        self.output_structure = configuration
-        try:
-            self._expressions = map_structure(
-                self._jsonpath_expr_or_literal, self.output_structure
-            )
-        except JSONPathError as ex:
-            raise ValueError(
-                "output_structure leaf values must be JSON literals or jsonpath query strings"
-            ) from ex
-
-    def _jsonpath_expr_or_literal(self, x):
-        if isinstance(x, str):
+        self.configuration = configuration
+        self._transformation = self._compile(self.configuration)
+
+    def _compile(self, x) -> Callable | list | dict:
+        if isinstance(x, dict):
+            if (compute := x.get("$compute")) is not None:
+                if len(x) != 1:
+                    raise ValueError("$compute must be the only key in the dict")
+                return _LeafExpression(compute)
+            else:
+                # Escape '$' in dict keys
+                d: dict[str, Any] = {}
+                for k, v in x.items():
+                    if k.startswith("$"):
+                        if k.startswith("$$"):
+                            k = k[1:]
+                        else:
+                            raise ValueError(
+                                f"dict key '{k}': keys beginning with '$' are reserved; use '$$' to escape"
+                            )
+                    d[k] = self._compile(v)
+                return d
+        elif isinstance(x, list):
+            return [self._compile(y) for y in x]
+        elif isinstance(x, str):
             if x.startswith("$"):
                 if x.startswith("$$"):
                     # Literal string -- remove "escape" character
-                    return x[1:]
+                    return _Literal(x[1:])
                 else:
-                    return jsonpath.parse(x)
-        return x
+                    return _Value_jsonpath(x, kind="scalar")
+        return _Literal(x)
 
     def __call__(self, stream: Iterable[dict]) -> Iterable[dict]:
-        def query(data, expr):
-            if not isinstance(expr, jsonpath.JSONPath):
-                # Literal
-                return expr
-            results = expr.find(data)
-            if len(results) == 0:
-                raise ValueError(f"no match for {expr}")
-            elif len(results) > 1:
-                raise ValueError(f"multiple results for {expr}")
-            return results[0].value
-
         for item in stream:
-            transformed = map_structure(
-                lambda expr: query(item, expr), self._expressions
-            )
-            yield transformed
+            yield map_structure(lambda compute: compute(item), self._transformation)
 
 
 class EmbedIndex:
@@ -573,13 +836,16 @@ def _test():
     print(list(transformer([data])))
 
     transformer = TransformJSON(
-        {"id": "$.object.id", "children": {"left": "$.list[0]", "right": "$.list[1]"}}
+        {
+            "id": "$.object.id",
+            "children": {"left": "$.list[0]", "right": "$.list[1]"},
+        }
     )
     print(
         list(
             transformer(
                 [
-                    {"object": {"id": 42, "name": "spam"}, "list": [1, 2]},
+                    {"object": {"id": "abc123", "name": "spam"}, "list": [1, 2]},
                 ]
             )
         )

dyff/schema/v0/r1/platform.py

@@ -1370,6 +1370,52 @@ class TaskSchema(DyffSchemaBaseModel):
     objective: str
 
 
+class EvaluationClientConfiguration(DyffSchemaBaseModel):
+    badRequestPolicy: Literal["Abort", "Skip"] = pydantic.Field(
+        default="Abort",
+        description="What to do if an inference call raises a 400 Bad Request"
+        " or a similar error that indicates a problem with the input instance."
+        " Abort (default): the evaluation fails immediately."
+        " Skip: output None for the bad instance and continue.",
+    )
+
+    transientErrorRetryLimit: int = pydantic.Field(
+        default=120,
+        ge=0,
+        description="How many times to retry transient errors before the"
+        " evaluation fails. The count is reset after a successful inference."
+        " Note that transient errors often occur during inference service"
+        " startup. The maximum time that the evaluation will wait for a"
+        " service (re)start is (retry limit) * (retry delay).",
+    )
+
+    transientErrorRetryDelaySeconds: int = pydantic.Field(
+        default=30,
+        ge=1,
+        description="How long to wait before retrying a transient error."
+        " Note that transient errors often occur during inference service"
+        " startup. The maximum time that the evaluation will wait for a"
+        " service (re)start is (retry limit) * (retry delay).",
+    )
+
+    duplicateOutputPolicy: Literal["Deduplicate", "Error", "Ignore"] = pydantic.Field(
+        default="Deduplicate",
+        description="What to do if there are missing outputs."
+        " Deduplicate (default): output only one of the duplicates, chosen"
+        " arbitrarily. Error: the evaluation fails. Ignore: duplicates are"
+        " retained in the output."
+        " Setting this to Error is discouraged because duplicates can"
+        " arise in normal operation if the client restarts due to"
+        " a transient failure.",
+    )
+
+    missingOutputPolicy: Literal["Error", "Ignore"] = pydantic.Field(
+        default="Error",
+        description="What to do if there are missing outputs."
+        " Error (default): the evaluation fails. Ignore: no error.",
+    )
+
+
 class EvaluationBase(DyffSchemaBaseModel):
     dataset: str = pydantic.Field(description="The Dataset to evaluate on.")
 
@@ -1377,11 +1423,17 @@ class EvaluationBase(DyffSchemaBaseModel):
         default=1, description="Number of replications to run."
     )
 
+    # TODO: This should be in the client config object
     workersPerReplica: Optional[int] = pydantic.Field(
         default=None,
         description="Number of data workers per inference service replica.",
     )
 
+    client: EvaluationClientConfiguration = pydantic.Field(
+        default_factory=EvaluationClientConfiguration,
+        description="Configuration for the evaluation client.",
+    )
+
 
 class Evaluation(DyffEntity, EvaluationBase):
     """A description of how to run an InferenceService on a Dataset to obtain a set of

{dyff_schema-0.23.0.dist-info → dyff_schema-0.24.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: dyff-schema
-Version: 0.23.0
+Version: 0.24.1
 Summary: Data models for the Dyff AI auditing platform.
 Author-email: Digital Safety Research Institute <contact@dsri.org>
 License: Apache-2.0

{dyff_schema-0.23.0.dist-info → dyff_schema-0.24.1.dist-info}/RECORD

@@ -22,9 +22,9 @@ dyff/schema/io/__init__.py,sha256=L5y8UhRnojerPYHumsxQJRcHCNz8Hj9NM8b47mewMNs,92
 dyff/schema/io/vllm.py,sha256=2q05M_-lTzq9oywKXHPPpCFCSDVCSsRQqtmERzWTtio,123
 dyff/schema/v0/__init__.py,sha256=L5y8UhRnojerPYHumsxQJRcHCNz8Hj9NM8b47mewMNs,92
 dyff/schema/v0/r1/__init__.py,sha256=L5y8UhRnojerPYHumsxQJRcHCNz8Hj9NM8b47mewMNs,92
-dyff/schema/v0/r1/adapters.py,sha256=2t2oxsnGfSEDKKDIEYw4qqLXMH7qlFIwPVuLyUmbsHs,23552
+dyff/schema/v0/r1/adapters.py,sha256=dmQS2ecgDX4ZvTMOW-6NzV_Oq_UpaiyFd7QnSNoOnK8,33057
 dyff/schema/v0/r1/base.py,sha256=IpvlYDr6JjYo6tn8XW4C1Fpgd_uqzZGZsG_cuEn_gQs,19441
-dyff/schema/v0/r1/platform.py,sha256=xaJohO5SB_Nh6kIRmbAe1gc-xV_vGoM3R7UgMJcbByM,75980
+dyff/schema/v0/r1/platform.py,sha256=KpTGCS8x1FAkfrMEgbjfDpTGgwNTF28WSuohYnROvj8,78287
 dyff/schema/v0/r1/requests.py,sha256=4TM1IKG9IP4MyprIy9E9XA_JqvkuwKAuY1ao1BbVLI0,15676
 dyff/schema/v0/r1/test.py,sha256=X6dUyVd5svcPCI-PBMOAqEfK9jv3bRDvkQTJzwS96c0,10720
 dyff/schema/v0/r1/version.py,sha256=isKAGuGxsdru8vDaYmI4YiZdJOu_wNxXK7u6QzD6FE4,392
@@ -37,9 +37,9 @@ dyff/schema/v0/r1/dataset/text.py,sha256=nLIn91Zlt0tNdXUklSgjJ-kEDxoPX32ISLkiv2D
 dyff/schema/v0/r1/dataset/vision.py,sha256=aIe0fbfM_g3DsrDTdg2K803YKLjZBpurM_VJcJFuZLc,369
 dyff/schema/v0/r1/io/__init__.py,sha256=L5y8UhRnojerPYHumsxQJRcHCNz8Hj9NM8b47mewMNs,92
 dyff/schema/v0/r1/io/vllm.py,sha256=CUE9y8KthtUI7sD49S875rDmPvKotSXVIRaBS79aBZs,5320
-dyff_schema-0.23.0.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-dyff_schema-0.23.0.dist-info/METADATA,sha256=D_F0_u21Cwr_Wh1n8KGZtOOKlQCYk52gkL-Z6EqbC5o,3482
-dyff_schema-0.23.0.dist-info/NOTICE,sha256=YONACu0s_Ui6jNi-wtEsVQbTU1JIkh8wvLH6d1-Ni_w,43
-dyff_schema-0.23.0.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
-dyff_schema-0.23.0.dist-info/top_level.txt,sha256=9e3VVdeX73t_sUJOPQPCcGtYO1JhoErhHIi3WoWGcFI,5
-dyff_schema-0.23.0.dist-info/RECORD,,
+dyff_schema-0.24.1.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+dyff_schema-0.24.1.dist-info/METADATA,sha256=joonUGA6du1G3HJKOTBQE7HuOivJW33vIMw3zNrl3bg,3482
+dyff_schema-0.24.1.dist-info/NOTICE,sha256=YONACu0s_Ui6jNi-wtEsVQbTU1JIkh8wvLH6d1-Ni_w,43
+dyff_schema-0.24.1.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
+dyff_schema-0.24.1.dist-info/top_level.txt,sha256=9e3VVdeX73t_sUJOPQPCcGtYO1JhoErhHIi3WoWGcFI,5
+dyff_schema-0.24.1.dist-info/RECORD,,