forked from srush/MiniChain
-
Notifications
You must be signed in to change notification settings - Fork 0
/
base.py
231 lines (191 loc) · 6.53 KB
/
base.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
import json
from dataclasses import asdict, dataclass, fields, is_dataclass
from enum import Enum
from itertools import count
from typing import (
Any,
Callable,
Dict,
Generic,
List,
Optional,
Tuple,
Type,
TypeVar,
Union,
get_args,
get_origin,
)
from eliot import start_action
from jinja2 import (
Environment,
FileSystemLoader,
PackageLoader,
Template,
select_autoescape,
)
from .backend import Backend, MinichainContext, Request
env = Environment(
loader=PackageLoader("minichain"),
autoescape=select_autoescape(),
extensions=["jinja2_highlight.HighlightExtension"],
)
def _prompt(r: Union[str, Request]) -> Request:
if isinstance(r, str):
return Request(r)
else:
return r
Input = TypeVar("Input")
Output = TypeVar("Output")
FnOutput = TypeVar("FnOutput")
def enum(x: Type[Enum]) -> Dict[str, int]:
d = {e.name: e.value for e in x}
return d
def walk(x: Any) -> Any:
if issubclass(x if get_origin(x) is None else get_origin(x), List):
return {"_t_": "list", "t": walk(get_args(x)[0])}
if issubclass(x, Enum):
return enum(x)
if is_dataclass(x):
return {y.name: walk(y.type) for y in fields(x)}
return x.__name__
def type_to_prompt(Out: type) -> str:
tmp = env.get_template("type_prompt.pmpt.tpl")
d = walk(Out)
return tmp.render({"typ": d})
def simple(model, **kwargs): # type: ignore
return model(kwargs)
@dataclass
class History:
prompt: "Prompt[Input, Output, FnOutput]"
inputs: List[Any]
@dataclass
class Fail:
arg_num: int
data: Any
@dataclass
class Chain:
# TODO: Add caching
history: History
def run(self, trial: int = 0, data: Any = None) -> Any:
args = []
count = []
for inp in self.history.inputs:
if isinstance(inp, Chain):
inp = inp.run()
args.append(inp)
count.append(0)
out = self.history.prompt.expand(args, trial, data)
while isinstance(out, Fail):
count[out.arg_num] += 1
inp = self.history.inputs[out.arg_num]
assert isinstance(inp, Chain)
args[out.arg_num] = inp.run(trial=count[out.arg_num], data=out.data)
out = self.history.prompt.expand(args, trial, data)
return out
class Prompt(Generic[Input, Output, FnOutput]):
counter = count()
def __init__(
self,
backend: Backend,
parser: Union[str, Callable[[str], Output]],
template_file: Optional[str],
template: Optional[str],
stop_template: Optional[str],
fn: Callable[[Callable[[Input], Output]], FnOutput] = simple,
):
self.backend: Backend = backend
self.parser: Union[str, Callable[[str], Output]] = parser
self.template_file: Optional[str] = template_file
self.template: Optional[str] = template
self.stop_template: Optional[str] = stop_template
self.fn = fn
self._fn: str = fn.__name__
self._id: int = Prompt.counter.__next__()
def parse(self, response: str) -> Any:
"""
Convert from the string response of the function
to the output type.
"""
if isinstance(self.parser, str):
if self.parser == "str":
return response
elif self.parser == "json":
return json.loads(response)
else:
return self.parser(response)
def run_verbose(self, r: Union[str, Request]) -> Tuple[Request, str, Output]:
# assert self.backend is not None
with start_action(action_type=str(self.fn)):
request = _prompt(r)
with start_action(action_type="Prompted", prompt=request.prompt):
response: Union[str, Any] = self.backend.run(request)
with start_action(action_type="Response", result=response):
output = self.parse(response)
if not isinstance(response, str):
response = "(data)"
return (request, response, output)
def template_fill(self, inp: Any) -> Request:
kwargs = inp
if self.template_file:
tmp = Environment(loader=FileSystemLoader(".")).get_template(
name=self.template_file
)
elif self.template:
tmp = Template(self.template)
x = tmp.render(**kwargs)
if self.stop_template:
stop = [Template(self.stop_template).render(**kwargs)]
else:
stop = None
return Request(x, stop)
def __call__(self, *args: Any) -> FnOutput:
return Chain(History(self, args))
class Model:
def __init__(self, prompt, trial, data):
self.prompt = prompt
self.trial = trial
self.data = data
self.run_log = None
def fail(self, argnum: int, data: Any = None) -> Fail:
return Fail(argnum - 1, data)
def __call__(self, input_):
assert self.run_log is None, "Only call `model` once per function"
if (
self.prompt.template is not None
or self.prompt.template_file is not None
):
if not isinstance(input_, dict):
input_ = asdict(input_)
input_ = dict(**input_)
input_["_trial"] = self.trial
input_["_fail_data"] = self.data
result = self.prompt.template_fill(input_)
else:
result = input_
self.run_log = self.prompt.run_verbose(result)
return self.run_log[-1]
def expand(self, args, trial=0, data=None):
model = self.Model(self, trial, data)
output = self.fn(model, *args)
if not isinstance(output, Fail):
t = model.run_log
MinichainContext.prompt_count.setdefault(self._id, -1)
if trial == 0:
MinichainContext.prompt_count[self._id] += 1
count = MinichainContext.prompt_count[self._id]
MinichainContext.prompt_store.setdefault((self._id, count), [])
MinichainContext.prompt_store[self._id, count].append(
(args, t[0], t[1], output)
)
return output
def prompt(
backend: Backend,
parser: Union[str, Any] = "str",
template_file: Optional[str] = None,
template: Optional[str] = None,
stop_template: Optional[str] = None,
) -> Callable[[Any], Prompt[Input, Output, FnOutput]]:
return lambda fn: Prompt(
backend, parser, template_file, template, stop_template, fn
)