DSPy: The framework for programming—not prompting—language models
Python
minor react fixes (#1880)
Chen Qian97ca14cdae61
diff --git a/dspy/predict/react.py b/dspy/predict/react.pyindex c6fcb6130..b626f1dc3 100644--- a/dspy/predict/react.py+++ b/dspy/predict/react.py@@ -1,23 +1,27 @@-import dspyimport inspect+from typing import Any, Callable, Literal, get_origin, get_type_hintsfrom pydantic import BaseModel++import dspy+from dspy.adapters.json_adapter import get_annotation_namefrom dspy.primitives.program import Modulefrom dspy.signatures.signature import ensure_signature-from dspy.adapters.json_adapter import get_annotation_namefrom dspy.utils.callback import with_callbacks-from typing import Callable, Any, get_type_hints, get_origin, Literal+class Tool:def __init__(self, func: Callable, name: str = None, desc: str = None, args: dict[str, Any] = None):annotations_func = func if inspect.isfunction(func) or inspect.ismethod(func) else func.__call__self.func = func- self.name = name or getattr(func, '__name__', type(func).__name__)- self.desc = desc or getattr(func, '__doc__', None) or getattr(annotations_func, '__doc__', "")+ self.name = name or getattr(func, "__name__", type(func).__name__)+ self.desc = desc or getattr(func, "__doc__", None) or getattr(annotations_func, "__doc__", "")self.args = {- k: v.schema() if isinstance((origin := get_origin(v) or v), type) and issubclass(origin, BaseModel)+ k: v.schema()+ if isinstance((origin := get_origin(v) or v), type) and issubclass(origin, BaseModel)else get_annotation_name(v)- for k, v in (args or get_type_hints(annotations_func)).items() if k != 'return'+ for k, v in (args or get_type_hints(annotations_func)).items()+ if k != "return"}@with_callbacks@@ -28,36 +32,40 @@ class Tool:class ReAct(Module):def __init__(self, signature, tools: list[Callable], max_iters=5):"""- Tools is either a list of functions, callable classes, or dspy.Tool instances.+ `tools` is either a list of functions, callable classes, or `dspy.Tool` instances."""self.signature = signature = ensure_signature(signature)self.max_iters = max_iters- tools = [t if isinstance(t, Tool) or hasattr(t, 'input_variable') else Tool(t) for t in tools]+ tools = [t if isinstance(t, Tool) or hasattr(t, "input_variable") else Tool(t) for t in tools]tools = {tool.name: tool for tool in tools}- inputs_ = ", ".join([f"`{k}`" for k in signature.input_fields.keys()])- outputs_ = ", ".join([f"`{k}`" for k in signature.output_fields.keys()])+ inputs = ", ".join([f"`{k}`" for k in signature.input_fields.keys()])+ outputs = ", ".join([f"`{k}`" for k in signature.output_fields.keys()])instr = [f"{signature.instructions}\n"] if signature.instructions else []- instr.extend([- f"You will be given {inputs_} and your goal is to finish with {outputs_}.\n",- "To do this, you will interleave Thought, Tool Name, and Tool Args, and receive a resulting Observation.\n",- "Thought can reason about the current situation, and Tool Name can be the following types:\n",- ])+ instr.extend(+ [+ f"You will be given {inputs} and your goal is to finish with {outputs}.\n",+ "To do this, you will interleave Thought, Tool Name, and Tool Args, and receive a resulting Observation.\n",+ "Thought can reason about the current situation, and Tool Name can be the following types:\n",+ ]+ )- finish_desc = f"Signals that the final outputs, i.e. {outputs_}, are now available and marks the task as complete."- finish_args = {} #k: v.annotation for k, v in signature.output_fields.items()}+ finish_desc = (+ f"Signals that the final outputs, i.e. {outputs}, are now available and marks the task as complete."+ )+ finish_args = {} # k: v.annotation for k, v in signature.output_fields.items()}tools["finish"] = Tool(func=lambda **kwargs: "Completed.", name="finish", desc=finish_desc, args=finish_args)for idx, tool in enumerate(tools.values()):- args = tool.args if hasattr(tool, 'args') else str({tool.input_variable: str})- desc = (f", whose description is <desc>{tool.desc}</desc>." if tool.desc else ".").replace('\n', " ")+ args = tool.args if hasattr(tool, "args") else str({tool.input_variable: str})+ desc = (f", whose description is <desc>{tool.desc}</desc>." if tool.desc else ".").replace("\n", " ")desc += f" It takes arguments {args} in JSON format."instr.append(f"({idx+1}) {tool.name}{desc}")- signature_ = (+ react_signature = (dspy.Signature({**signature.input_fields}, "\n".join(instr)).append("trajectory", dspy.InputField(), type_=str).append("next_thought", dspy.OutputField(), type_=str)@@ -65,25 +73,23 @@ class ReAct(Module):.append("next_tool_args", dspy.OutputField(), type_=dict[str, Any]))- fallback_signature = (- dspy.Signature({**signature.input_fields, **signature.output_fields}, signature.instructions)- .append("trajectory", dspy.InputField(), type_=str)- )+ fallback_signature = dspy.Signature(+ {**signature.input_fields, **signature.output_fields}, signature.instructions+ ).append("trajectory", dspy.InputField(), type_=str)self.tools = tools- self.react = dspy.Predict(signature_)+ self.react = dspy.Predict(react_signature)self.extract = dspy.ChainOfThought(fallback_signature)-- def forward(self, **input_args):- trajectory = {}- def format(trajectory_: dict[str, Any], last_iteration: bool):+ def forward(self, **input_args):+ def format(trajectory: dict[str, Any], last_iteration: bool):adapter = dspy.settings.adapter or dspy.ChatAdapter()- signature_ = dspy.Signature(f"{', '.join(trajectory_.keys())} -> x")- return adapter.format_fields(signature_, trajectory_, role='user')+ trajectory_signature = dspy.Signature(f"{', '.join(trajectory.keys())} -> x")+ return adapter.format_fields(trajectory_signature, trajectory, role="user")+ trajectory = {}for idx in range(self.max_iters):- pred = self.react(**input_args, trajectory=format(trajectory, last_iteration=(idx == self.max_iters-1)))+ pred = self.react(**input_args, trajectory=format(trajectory, last_iteration=(idx == self.max_iters - 1)))trajectory[f"thought_{idx}"] = pred.next_thoughttrajectory[f"tool_name_{idx}"] = pred.next_tool_name@@ -96,7 +102,7 @@ class ReAct(Module):if pred.next_tool_name == "finish":break-+extract = self.extract(**input_args, trajectory=format(trajectory, last_iteration=False))return dspy.Prediction(trajectory=trajectory, **extract)