synth-ai 0.2.0__py3-none-any.whl → 0.2.1.dev0__py3-none-any.whl

synth_ai/environments/examples/enron/environment.py

@@ -0,0 +1,165 @@
+# environment.py
+from __future__ import annotations
+from typing import List, Optional, Dict, Any, Union
+from pydantic import BaseModel, Field
+
+from synth_ai.environments.environment.tools import (
+    EnvToolCall,
+    ToolResult,
+    TOOL_REGISTRY,
+    register_tool,
+)
+from synth_ai.environments.environment.shared_engine import (
+    GetObservationCallable,
+    InternalObservation,
+)
+from synth_ai.environments.stateful.core import StatefulEnvironment
+from synth_ai.environments.examples.enron.engine import (
+    EnronEngine,
+    ACTION_SEARCH,
+    ACTION_READ,
+    ACTION_ANSWER,
+)
+from synth_ai.environments.examples.enron.taskset import EnronTaskInstance
+
+
+# -------- pydantic schemas (used by agent / LLM function calls)
+class SearchEmailsArgs(BaseModel):
+    inbox: str = Field(..., description="Email address performing the search (used by tool logic)")
+    keywords: List[str] = Field(..., description="Keywords to AND-search for")
+    from_addr: Optional[str] = None
+    to_addr: Optional[str] = None
+    sent_after: Optional[str] = None
+    sent_before: Optional[str] = None
+    max_results: int = Field(10, le=10)
+
+
+class ReadEmailArgs(BaseModel):
+    message_id: str
+
+
+class AnswerQuestionArgs(BaseModel):
+    answer: str
+
+
+# --------------------------------------------------------------------------- tool wrappers
+class SearchEmails(EnvToolCall):
+    def __init__(self, **kwargs):
+        self.action = (ACTION_SEARCH, kwargs)
+
+
+class ReadEmail(EnvToolCall):
+    def __init__(self, message_id: str):
+        self.action = (ACTION_READ, message_id)
+
+
+class AnswerQuestion(EnvToolCall):
+    def __init__(self, answer: str):
+        self.action = (ACTION_ANSWER, answer)
+
+
+# -- terminate wrapper (maps to an empty-answer ACTION_ANSWER) --------------
+class Terminate(EnvToolCall):
+    def __init__(self):
+        self.action = (ACTION_ANSWER, "")
+
+
+# -------- observation callable (optional for formatted observations)
+class SynthEnronObservationCallable(GetObservationCallable):
+    async def get_observation(
+        self, pub: Dict[str, Any], priv: Dict[str, Any]
+    ) -> InternalObservation:
+        """Format observation as a human-readable string."""
+        q = pub.get("question")
+        rwd = priv.get("reward_last")
+        return f"Q: {q}\nTools: {pub.get('tools')}\nAnswered: {pub.get('already_answered')}\nSearch Res: {len(pub.get('search_results', []))} items\nEmail Loaded: {pub.get('email') is not None}\nTool Error: {pub.get('tool_error')}\nReward Δ: {rwd}"
+
+
+# --------------------------------------------------------------------------- environment
+class EnronEnvironment(StatefulEnvironment):
+    def __init__(
+        self,
+        task_instance: EnronTaskInstance,
+        custom_obs: Optional[GetObservationCallable] = None,
+    ):
+        self.engine = EnronEngine(task_instance)
+        self.custom_obs = custom_obs or SynthEnronObservationCallable()
+        self.name = "Enron-QA-Env"
+
+        # Store tool instances on self for reliable access
+        self._tools_instances = {
+            "search_emails": SearchEmailsTool(self.engine),
+            "read_email": ReadEmailTool(self.engine),
+            "answer_question": AnswerQuestionTool(self.engine),
+            "terminate": TerminateTool(self.engine),
+        }
+        for tool_name, tool_instance in self._tools_instances.items():
+            if tool_name not in TOOL_REGISTRY:
+                register_tool(tool_instance)
+            elif TOOL_REGISTRY[tool_name].engine is not self.engine:
+                register_tool(tool_instance)
+
+    async def initialize(self) -> InternalObservation:
+        priv, pub = await self.engine._reset_engine()
+        return await self._obs(priv, pub)
+
+    async def step(
+        self,
+        calls: Union[EnvToolCall, List[EnvToolCall], List[List[EnvToolCall]]],
+    ) -> InternalObservation:
+        # normalise → always [[EnvToolCall]]
+        if isinstance(calls, EnvToolCall):
+            calls = [[calls]]
+        elif calls and isinstance(calls[0], EnvToolCall):
+            calls = [calls]
+
+        if not isinstance(calls[0][0], EnvToolCall):
+            raise TypeError(f"Processed call is not EnvToolCall: {type(calls[0][0])}")
+
+        tool_name = calls[0][0].tool
+        tool_to_execute = self._tools_instances.get(tool_name)
+
+        if not tool_to_execute:
+            tool_to_execute = TOOL_REGISTRY.get(tool_name)
+            if not tool_to_execute:
+                raise ValueError(f"Tool '{tool_name}' not found.")
+
+        tool_result: ToolResult = await tool_to_execute(calls[0][0])
+
+        public_payload_for_engine = (
+            tool_result.payload if tool_result.ok and tool_result.payload else {}
+        )
+        if not tool_result.ok:
+            public_payload_for_engine["tool_error"] = tool_result.error
+
+        priv, pub = await self.engine._step_engine(public_payload_for_engine)
+        return await self._obs(priv, pub)
+
+    async def terminate(self) -> InternalObservation:
+        self.engine.close_db()
+        priv_state_on_terminate = {
+            "reward_last": 0,
+            "total_reward": self.engine.total_reward,
+            "terminated": True,
+            "truncated": False,
+            "gold_answer": self.engine._sample()["answer"],
+        }
+        pub_state_on_terminate = {
+            "question": self.engine._sample()["question"],
+            "tools": [],
+            "already_answered": self.engine.answered,
+            "status": "terminated_by_env",
+        }
+        return await self._obs(priv_state_on_terminate, pub_state_on_terminate)
+
+    async def checkpoint(self) -> InternalObservation:
+        snapshot = await self.engine._serialize_engine()
+        return {
+            "engine_snapshot": snapshot.model_dump(),
+            "message": "Checkpoint created",
+        }
+
+    async def _obs(self, priv: Dict[str, Any], pub: Dict[str, Any]):
+        if self.custom_obs:
+            return await self.custom_obs.get_observation(pub, priv)
+        return {**pub, **priv}

synth_ai/environments/examples/enron/taskset.py

@@ -0,0 +1,112 @@
+# taskset.py
+from __future__ import annotations
+import asyncio
+from uuid import uuid4
+import os
+
+from datasets import load_dataset
+from dataclasses import dataclass, asdict
+
+from synth_ai.environments.tasks.core import (
+    Task,
+    TaskInstance,
+    TaskInstanceSet,
+    TaskInstanceMetadata,
+    SplitInfo,
+    Impetus,
+    Intent,
+)
+
+enron_task = Task(
+    global_premises="Answer factual questions by reading Enron e-mails",
+    global_constraints="",
+    global_objectives="Provide the correct answer; minimise queries",
+    shared_env_params={},
+)
+
+
+# --------------------------------------------------------------------------- metadata
+@dataclass
+class EnronTaskInstanceMetadata(TaskInstanceMetadata):
+    split: str
+    email_count: int
+    message_ids: list[str]
+
+
+@dataclass
+class EnronTaskInstance(TaskInstance):
+    async def serialize(self):
+        data = asdict(self)
+        if isinstance(data.get("id"), uuid4().__class__):
+            data["id"] = str(data["id"])
+        return data
+
+    @classmethod
+    async def deserialize(cls, data: dict) -> "EnronTaskInstance":
+        return cls(**data)
+
+
+# --------------------------------------------------------------------------- task-set builder
+# Use a local dataset cache under examples/enron/dataset
+CACHE_DIR = os.path.join(os.path.dirname(__file__), "dataset")
+os.makedirs(CACHE_DIR, exist_ok=True)
+
+
+async def create_enron_taskset() -> TaskInstanceSet:
+    ds_train = load_dataset(
+        "corbt/enron_emails_sample_questions",
+        split="train",
+        cache_dir=CACHE_DIR,
+    )
+    ds_test = load_dataset(
+        "corbt/enron_emails_sample_questions",
+        split="test",
+        cache_dir=CACHE_DIR,
+    )
+
+    def to_instance(row: dict, split: str) -> EnronTaskInstance:
+        impetus = Impetus(instructions=row["question"])
+        intent = Intent(
+            rubric={"goal": "Answer the question using the Enron emails."},
+            gold_trajectories=None,
+            gold_state_diff={"answer": row["answer"]},
+        )
+        metadata = EnronTaskInstanceMetadata(
+            split=split,
+            email_count=len(row["message_ids"]),
+            message_ids=row["message_ids"],
+        )
+        return EnronTaskInstance(
+            id=uuid4(),
+            impetus=impetus,
+            intent=intent,
+            metadata=metadata,
+            is_reproducible=True,
+            initial_engine_snapshot=row,
+        )
+
+    train_instances = [to_instance(r, "train") for r in ds_train]
+    test_instances = [to_instance(r, "test") for r in ds_test]
+
+    split_info = SplitInfo(
+        val_instance_ids=set(),
+        test_instance_ids={inst.id for inst in test_instances},
+        _is_split_defined=True,
+    )
+
+    return TaskInstanceSet(
+        name="Enron-QA",
+        description="QA over Enron email dataset sample.",
+        instances=train_instances + test_instances,
+        split_info=split_info,
+    )
+
+
+# quick sanity check ----------------------------------------------------------
+if __name__ == "__main__":
+
+    async def _main():
+        ts = await create_enron_taskset()
+        print(f"{len(ts.instances)} instances built.")
+
+    asyncio.run(_main())

synth_ai/environments/examples/enron/units/keyword_stats.py

@@ -0,0 +1,111 @@
+"""
+Script: enron_keyword_logging.py
+Purpose: Iterate over a sample of Enron-QA tasks and compare the hit-rate of the
+full keyword list extracted from the natural-language question with the hit-rate
+when the *final* keyword is dropped (the heuristic your current agent uses).
+
+It logs the result counts side-by-side so you can see whether the heuristic is
+generally helpful or not.
+
+Run with:
+    python enron_keyword_logging.py --n 50  # test 50 random tasks
+Outputs a CSV "keyword_stats.csv" for easy inspection in Excel/Sheets.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import re
+import random
+import asyncio
+from pathlib import Path
+from synth_ai.environments.examples.enron.taskset import create_enron_taskset
+from synth_ai.environments.examples.enron.art_helpers import email_search_tools  # low-level search
+
+# --- simple helpers ---------------------------------------------------------
+STOPWORDS = {
+    "the",
+    "a",
+    "an",
+    "and",
+    "or",
+    "of",
+    "at",
+    "in",
+    "on",
+    "for",
+    "to",
+    "with",
+    "my",
+    "your",
+    "our",
+    "did",
+    "do",
+    "is",
+    "was",
+    "were",
+    "be",
+    "been",
+    "am",
+    "when",
+    "what",
+    "which",
+    "who",
+}
+
+TOKEN_RE = re.compile(r"[A-Za-z0-9']+")
+
+
+def extract_keywords(question: str) -> list[str]:
+    """Very naive keyword extractor: tokens minus stop-words."""
+    tokens = [t.lower() for t in TOKEN_RE.findall(question)]
+    return [t for t in tokens if t not in STOPWORDS]
+
+
+# ---------------------------------------------------------------------------
+async def main(n: int):
+    taskset = await create_enron_taskset()
+    sample = random.sample(taskset.instances, k=min(n, len(taskset.instances)))
+
+    rows: list[dict[str, str | int]] = []
+    for inst in sample:
+        q = inst.impetus.instructions
+        kws_full = extract_keywords(q)
+        if not kws_full:
+            continue
+
+        # search using the low-level helper once so we don't need a whole env
+        hits_full = email_search_tools.search_emails(inbox="user", keywords=kws_full, max_results=5)
+
+        hits_trim = (
+            email_search_tools.search_emails(inbox="user", keywords=kws_full[:-1], max_results=5)
+            if len(kws_full) > 1
+            else []
+        )
+
+        rows.append(
+            {
+                "question": q,
+                "keywords_full": " ".join(kws_full),
+                "hits_full": len(hits_full),
+                "keywords_trim": " ".join(kws_full[:-1]),
+                "hits_trim": len(hits_trim),
+            }
+        )
+
+    # write CSV
+    out_path = Path("keyword_stats.csv")
+    with out_path.open("w", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=list(rows[0].keys()))
+        writer.writeheader()
+        writer.writerows(rows)
+
+    print(f"Wrote {len(rows)} rows to {out_path.resolve()}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--n", type=int, default=30, help="number of tasks to sample")
+    args = parser.parse_args()
+    asyncio.run(main(args.n))

synth_ai/environments/examples/enron/units/test_email_index.py

@@ -0,0 +1,8 @@
+import pytest
+from synth_ai.environments.examples.enron.art_helpers.email_search_tools import search_emails
+
+
+@pytest.mark.parametrize("kw", [["enron"]])  # , ["meeting"], ["energy"]
+def test_index_has_hits(kw):
+    hits = search_emails(inbox="john.lavorato@enron.com", keywords=kw)
+    assert len(hits) > 0, f"no hits for {kw}"

synth_ai/environments/examples/minigrid/__init__.py

@@ -0,0 +1,48 @@
+"""MiniGrid environment example for synth_env.
+
+This module provides a comprehensive implementation of MiniGrid environments
+with full state management, tool-based interaction, and task generation.
+"""
+
+from synth_ai.environments.examples.minigrid.engine import (
+    MiniGridEngine,
+    MiniGridPublicState,
+    MiniGridPrivateState,
+    MiniGridGoalReachedComponent,
+    MiniGridStepPenaltyComponent,
+    MiniGridObservationCallable,
+    MiniGridCheckpointObservationCallable,
+)
+from synth_ai.environments.examples.minigrid.environment import (
+    MiniGridEnvironment,
+    MiniGridInteractTool,
+    MiniGridActionInput,
+)
+from synth_ai.environments.examples.minigrid.taskset import (
+    MiniGridTaskInstance,
+    MiniGridTaskInstanceMetadata,
+    DEFAULT_MINIGRID_TASK,
+    create_minigrid_taskset,
+    taskset,
+)
+
+__all__ = [
+    # Engine
+    "MiniGridEngine",
+    "MiniGridPublicState",
+    "MiniGridPrivateState",
+    "MiniGridGoalReachedComponent",
+    "MiniGridStepPenaltyComponent",
+    "MiniGridObservationCallable",
+    "MiniGridCheckpointObservationCallable",
+    # Environment
+    "MiniGridEnvironment",
+    "MiniGridInteractTool",
+    "MiniGridActionInput",
+    # TaskSet
+    "MiniGridTaskInstance",
+    "MiniGridTaskInstanceMetadata",
+    "DEFAULT_MINIGRID_TASK",
+    "create_minigrid_taskset",
+    "taskset",
+]