Merge pull request #490 from julep-ai/f/parallel-map-reduce

agents-api/agents_api/activities/task_steps/base_evaluate.py

@@ -1,3 +1,4 @@
+import ast
 from typing import Any
 
 from beartype import beartype
@@ -13,10 +14,27 @@
 async def base_evaluate(
     exprs: str | list[str] | dict[str, str],
     values: dict[str, Any] = {},
+    extra_lambda_strs: dict[str, str] | None = None,
 ) -> Any | list[Any] | dict[str, Any]:
     input_len = 1 if isinstance(exprs, str) else len(exprs)
     assert input_len > 0, "exprs must be a non-empty string, list or dict"
 
+    extra_lambdas = {}
+    if extra_lambda_strs:
+        for k, v in extra_lambda_strs.items():
+            v = v.strip()
+
+            # Check that all extra lambdas are valid
+            assert v.startswith("lambda "), "All extra lambdas must start with 'lambda'"
+
+            try:
+                ast.parse(v)
+            except Exception as e:
+                raise ValueError(f"Invalid lambda: {v}") from e
+
+            # Eval the lambda and add it to the extra lambdas
+            extra_lambdas[k] = eval(v)
+
     # Turn the nested dict values from pydantic to dicts where possible
     values = {
         k: v.model_dump() if isinstance(v, BaseModel) else v for k, v in values.items()
@@ -25,7 +43,7 @@ async def base_evaluate(
     # frozen_box doesn't work coz we need some mutability in the values
     values = Box(values, frozen_box=False, conversion_box=True)
 
-    evaluator = get_evaluator(names=values)
+    evaluator = get_evaluator(names=values, extra_functions=extra_lambdas)
 
     try:
         match exprs:

agents-api/agents_api/activities/utils.py

@@ -1,5 +1,9 @@
 import json
-from typing import Any
+from functools import reduce
+from itertools import accumulate
+from random import random
+from time import time
+from typing import Any, Callable
 
 import re2
 import yaml
@@ -10,6 +14,7 @@
 # TODO: We need to make sure that we dont expose any security issues
 ALLOWED_FUNCTIONS = {
     "abs": abs,
+    "accumulate": accumulate,
     "all": all,
     "any": any,
     "bool": bool,
@@ -22,22 +27,32 @@
     "list": list,
     "load_json": json.loads,
     "load_yaml": lambda string: yaml.load(string, Loader=CSafeLoader),
+    "map": map,
     "match_regex": lambda pattern, string: bool(re2.fullmatch(pattern, string)),
     "max": max,
     "min": min,
+    "random": random,
+    "range": range,
+    "reduce": reduce,
     "round": round,
     "search_regex": lambda pattern, string: re2.search(pattern, string),
     "set": set,
     "str": str,
     "sum": sum,
+    "time": time,
     "tuple": tuple,
     "zip": zip,
 }
 
 
 @beartype
-def get_evaluator(names: dict[str, Any]) -> SimpleEval:
-    evaluator = EvalWithCompoundTypes(names=names, functions=ALLOWED_FUNCTIONS)
+def get_evaluator(
+    names: dict[str, Any], extra_functions: dict[str, Callable] | None = None
+) -> SimpleEval:
+    evaluator = EvalWithCompoundTypes(
+        names=names, functions=ALLOWED_FUNCTIONS | (extra_functions or {})
+    )
+
     return evaluator
 
 

agents-api/agents_api/autogen/Tasks.py

@@ -347,6 +347,13 @@ class Main(BaseModel):
     A special parameter named `results` is the accumulator and `_` is the current value.
     """
     initial: Any = []
+    """
+    The initial value of the reduce expression
+    """
+    parallelism: Annotated[int | None, Field(None, ge=1, le=100)]
+    """
+    Whether to run the reduce expression in parallel and how many items to run in each batch
+    """
 
 
 class MainModel(BaseModel):
@@ -381,6 +388,13 @@ class MainModel(BaseModel):
     A special parameter named `results` is the accumulator and `_` is the current value.
     """
     initial: Any = []
+    """
+    The initial value of the reduce expression
+    """
+    parallelism: Annotated[int | None, Field(None, ge=1, le=100)]
+    """
+    Whether to run the reduce expression in parallel and how many items to run in each batch
+    """
 
 
 class ParallelStep(BaseModel):

agents-api/agents_api/env.py

@@ -20,6 +20,10 @@
 api_prefix: str = env.str("AGENTS_API_PREFIX", default="")
 
 
+# Tasks
+# -----
+task_max_parallelism: int = env.int("AGENTS_API_TASK_MAX_PARALLELISM", default=100)
+
 # Debug
 # -----
 debug: bool = env.bool("AGENTS_API_DEBUG", default=False)

agents-api/agents_api/models/execution/create_execution_transition.py

@@ -82,12 +82,16 @@ def create_execution_transition(
     # Only required for updating the execution status as well
     update_execution_status: bool = False,
     task_id: UUID | None = None,
-) -> tuple[list[str], dict]:
+) -> tuple[list[str | None], dict]:
     transition_id = transition_id or uuid4()
 
     data.metadata = data.metadata or {}
     data.execution_id = execution_id
 
+    # TODO: This is a hack to make sure the transition is valid
+    #       (parallel transitions are whack, we should do something better)
+    is_parallel = data.current.workflow.startswith("PAR:")
+
     # Prepare the transition data
     transition_data = data.model_dump(exclude_unset=True, exclude={"id"})
 
@@ -184,9 +188,9 @@ def create_execution_transition(
             execution_id=execution_id,
             parents=[("agents", "agent_id"), ("tasks", "task_id")],
         ),
-        validate_status_query,
-        update_execution_query,
-        check_last_transition_query,
+        validate_status_query if not is_parallel else None,
+        update_execution_query if not is_parallel else None,
+        check_last_transition_query if not is_parallel else None,
         insert_query,
     ]
 

agents-api/agents_api/models/session/get_session.py

@@ -94,7 +94,7 @@ def get_session(
             render_templates,
             token_budget,
             context_overflow,
-            @ "NOW"
+            @ "END"
         }, updated_at = to_int(validity)
     """
 

agents-api/agents_api/models/session/list_sessions.py

@@ -101,7 +101,7 @@ def list_sessions(
                 metadata,
                 token_budget,
                 context_overflow,
-                @ "NOW"
+                @ "END"
             }},
             users_p[users, id],
             participants[agents, "agent", id],

agents-api/agents_api/models/session/patch_session.py

@@ -95,7 +95,7 @@ def patch_session(
             input[{session_update_cols}],
             ids[session_id, developer_id],
             *sessions{{
-                {rest_fields}, metadata: md, @ "NOW"
+                {rest_fields}, metadata: md, @ "END"
             }},
             updated_at = 'ASSERT',
             metadata = concat(md, $metadata),

agents-api/agents_api/models/session/prepare_session_data.py

@@ -159,7 +159,7 @@ def prepare_session_data(
             render_templates,
             token_budget,
             context_overflow,
-            @ "NOW"
+            @ "END"
         },
         updated_at = to_int(validity),
         record = {

agents-api/agents_api/models/session/update_session.py

@@ -84,7 +84,7 @@ def update_session(
             input[{session_update_cols}],
             ids[session_id, developer_id],
             *sessions{{
-                {rest_fields}, @ "NOW"
+                {rest_fields}, @ "END"
             }},
             updated_at = 'ASSERT'
 

agents-api/agents_api/models/task/get_task.py

@@ -64,7 +64,7 @@ def get_task(
             workflows,
             created_at,
             metadata,
-            @ 'NOW'
+            @ 'END'
         },
         updated_at = to_int(updated_at_ms) / 1000
 

agents-api/agents_api/models/task/list_tasks.py

@@ -70,7 +70,7 @@ def list_tasks(
             workflows,
             created_at,
             metadata,
-            @ 'NOW'
+            @ 'END'
         }},
         updated_at = to_int(updated_at_ms) / 1000
 

agents-api/agents_api/models/utils.py

@@ -97,7 +97,7 @@ def mark_session_updated_query(developer_id: UUID | str, session_id: UUID | str)
             render_templates,
             token_budget,
             context_overflow,
-            @ 'NOW'
+            @ 'END'
         }},
         updated_at = [floor(now()), true]
 
@@ -173,7 +173,7 @@ def make_cozo_json_query(fields):
 
 
 def cozo_query(
-    func: Callable[P, tuple[str | list[str], dict]] | None = None,
+    func: Callable[P, tuple[str | list[str | None], dict]] | None = None,
     debug: bool | None = None,
 ):
     def cozo_query_dec(func: Callable[P, tuple[str | list[Any], dict]]):