| 10 | logger = logging.getLogger(__name__) |
| 11 | |
| 12 | class LEAP: |
| 13 | def __init__(self, system_prompt: str, client, model: str, request_config: dict = None, request_id: str = None): |
| 14 | self.system_prompt = system_prompt |
| 15 | self.client = client |
| 16 | self.model = model |
| 17 | self.request_id = request_id |
| 18 | self.low_level_principles = [] |
| 19 | self.high_level_principles = [] |
| 20 | self.leap_completion_tokens = 0 |
| 21 | |
| 22 | # Extract max_tokens from request_config with default |
| 23 | self.max_tokens = 4096 |
| 24 | if request_config: |
| 25 | self.max_tokens = request_config.get('max_tokens', self.max_tokens) |
| 26 | |
| 27 | def extract_output(self, text: str) -> str: |
| 28 | match = re.search(r'<output>(.*?)(?:</output>|$)', text, re.DOTALL) |
| 29 | return match.group(1).strip() if match else "" |
| 30 | |
| 31 | def _extract_content(self, response, context: str) -> str: |
| 32 | """Validate a provider response and return its message content. |
| 33 | |
| 34 | Guards against the empty / None / length-truncated responses that would |
| 35 | otherwise raise IndexError/TypeError downstream (e.g. when the content |
| 36 | is fed to extract_output or split). Mirrors the response-validation |
| 37 | idiom already used in moa/bon/plansearch. |
| 38 | """ |
| 39 | if (response is None or |
| 40 | not response.choices or |
| 41 | response.choices[0].message.content is None or |
| 42 | response.choices[0].finish_reason == "length"): |
| 43 | raise Exception(f"LEAP: provider returned an empty, None, or truncated response while {context}") |
| 44 | return response.choices[0].message.content |
| 45 | |
| 46 | def extract_examples_from_query(self, initial_query: str) -> List[Tuple[str, str]]: |
| 47 | logger.info("Extracting examples from initial query") |
| 48 | |
| 49 | # Prepare request for logging |
| 50 | provider_request = { |
| 51 | "model": self.model, |
| 52 | "max_tokens": self.max_tokens, |
| 53 | "messages": [ |
| 54 | {"role": "system", "content": self.system_prompt}, |
| 55 | {"role": "user", "content": f""" |
| 56 | Analyze the following query and determine if it contains few-shot examples. |
| 57 | If it does, extract the examples and their corresponding answers. |
| 58 | Format the examples as a JSON array of objects, where each object has "question" and "answer" fields. |
| 59 | If there are no examples, return an empty array. |
| 60 | Enclose your response within <output></output> tags. |
| 61 | Do not put any explanation or any other reponse other than the JSON array within the <output></output> tags. |
| 62 | |
| 63 | Example output format: |
| 64 | <output> |
| 65 | [ |
| 66 | {{"question": "What is 2+2?", "answer": "4"}}, |
| 67 | {{"question": "What is the capital of France?", "answer": "Paris"}} |
| 68 | ] |
| 69 | </output> |