-
Notifications
You must be signed in to change notification settings - Fork 380
Expand file tree
/
Copy pathreread.py
More file actions
70 lines (59 loc) · 2.81 KB
/
Copy pathreread.py
File metadata and controls
70 lines (59 loc) · 2.81 KB
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
import logging
import optillm
from optillm import conversation_logger
logger = logging.getLogger(__name__)
def re2_approach(system_prompt, initial_query, client, model, n=1, request_config: dict = None, request_id: str = None):
"""
Implement the RE2 (Re-Reading) approach for improved reasoning in LLMs.
Args:
system_prompt (str): The system prompt to be used.
initial_query (str): The initial user query.
client: The OpenAI client object.
model (str): The name of the model to use.
n (int): Number of completions to generate.
request_config (dict): Optional configuration including max_tokens.
Returns:
str or list: The generated response(s) from the model.
"""
logger.info("Using RE2 approach for query processing")
re2_completion_tokens = 0
# Extract max_tokens from request_config if provided
max_tokens = None
if request_config:
max_tokens = request_config.get('max_tokens')
# Construct the RE2 prompt
re2_prompt = f"{initial_query}\nRead the question again: {initial_query}"
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": re2_prompt}
]
try:
provider_request = {
"model": model,
"messages": messages,
"n": n
}
if max_tokens is not None:
provider_request["max_tokens"] = max_tokens
response = client.chat.completions.create(**provider_request)
# Log provider call
if hasattr(optillm, 'conversation_logger') and optillm.conversation_logger and request_id:
response_dict = response.model_dump() if hasattr(response, 'model_dump') else response
optillm.conversation_logger.log_provider_call(request_id, provider_request, response_dict)
if response is None or not response.choices:
raise Exception("re2_approach: provider returned an empty or None response")
re2_completion_tokens += response.usage.completion_tokens
if n == 1:
if (response.choices[0].message.content is None or
response.choices[0].finish_reason == "length"):
raise Exception("re2_approach: provider returned a None or truncated response")
return response.choices[0].message.content.strip(), re2_completion_tokens
else:
# Keep only choices that produced usable, non-truncated content,
# consistent with how the single-choice path treats a truncation.
return [choice.message.content.strip() for choice in response.choices
if choice.message.content is not None
and choice.finish_reason != "length"], re2_completion_tokens
except Exception as e:
logger.error(f"Error in RE2 approach: {str(e)}")
raise