File size: 13,004 Bytes
46df5f0 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 |
"""
LLM-based citation relevance evaluator.
Supports OpenAI, Anthropic, DeepSeek, Gemini, vLLM, and Ollama backends.
"""
import json
import re
from dataclasses import dataclass
from typing import Optional, Dict, Any
from enum import Enum
import os
import requests
class LLMBackend(Enum):
OPENAI = "openai"
ANTHROPIC = "anthropic"
GEMINI = "gemini"
VLLM = "vllm"
OLLAMA = "ollama"
DEEPSEEK = "deepseek"
@dataclass
class EvaluationResult:
"""Result of LLM citation evaluation."""
entry_key: str
relevance_score: int # 1-5
is_relevant: bool
explanation: str
context_used: str
abstract_used: str
line_number: Optional[int] = None
file_path: Optional[str] = None
error: Optional[str] = None
@property
def score_label(self) -> str:
labels = {
1: "Not Relevant",
2: "Marginally Relevant",
3: "Somewhat Relevant",
4: "Relevant",
5: "Highly Relevant"
}
return labels.get(self.relevance_score, "Unknown")
class LLMEvaluator:
"""Evaluates citation relevance using LLM."""
PROMPT_TEMPLATE = """You are an expert academic reviewer. Given a citation context from a LaTeX document and the cited paper's abstract, evaluate whether this citation is appropriate and relevant.
## Citation Context (from the manuscript):
{context}
## Cited Paper's Abstract:
{abstract}
## Task:
Evaluate the relevance and appropriateness of this citation. Consider:
1. Does the citation support the claim being made in the context?
2. Is the cited paper's topic related to the discussion?
3. Is this citation necessary, or could it be replaced with a more relevant one?
## Response Format:
Provide your response in the following JSON format:
{{
"relevance_score": <1-5 integer>,
"is_relevant": <true/false>,
"explanation": "<brief explanation in 1-2 sentences>"
}}
Score guide:
- 1: Not relevant at all
- 2: Marginally relevant
- 3: Somewhat relevant
- 4: Relevant and appropriate
- 5: Highly relevant and essential
STRICTLY FOLLOW THE JSON FORMAT. Respond ONLY with the JSON object, no other text."""
def __init__(
self,
backend: LLMBackend = LLMBackend.GEMINI,
endpoint: Optional[str] = None,
model: Optional[str] = None,
api_key: Optional[str] = None
):
self.backend = backend
self.api_key = api_key or os.environ.get(f"{backend.name}_API_KEY")
# Set defaults based on backend
if backend == LLMBackend.OPENAI:
self.endpoint = endpoint or "https://api.openai.com/v1/chat/completions"
self.model = model or "gpt-5-mini"
elif backend == LLMBackend.ANTHROPIC:
self.endpoint = endpoint or "https://api.anthropic.com/v1/messages"
self.model = model or "claude-4.5-haiku"
elif backend == LLMBackend.DEEPSEEK:
self.endpoint = endpoint or "https://api.deepseek.com/chat/completions"
self.model = model or "deepseek-chat"
elif backend == LLMBackend.OLLAMA:
self.endpoint = endpoint or "http://localhost:11434/api/generate"
self.model = model or "Qwen/qwen3-4B-Instruct-2507"
elif backend == LLMBackend.VLLM:
self.endpoint = endpoint or "http://localhost:8000/v1/chat/completions"
self.model = model or "Qwen/qwen3-4B-Instruct-2507"
elif backend == LLMBackend.GEMINI:
self.endpoint = endpoint or "https://generativelanguage.googleapis.com/v1beta/models"
self.model = model or "gemini-2.5-flash-lite"
def evaluate(self, entry_key: str, context: str, abstract: str) -> EvaluationResult:
"""Evaluate citation relevance."""
if not context or not abstract:
return EvaluationResult(
entry_key=entry_key,
relevance_score=0,
is_relevant=False,
explanation="Missing context or abstract",
context_used=context,
abstract_used=abstract,
error="Missing context or abstract for evaluation"
)
# Don't truncate - preserve full context and abstract
prompt = self.PROMPT_TEMPLATE.format(context=context, abstract=abstract)
try:
if self.backend in (LLMBackend.OPENAI, LLMBackend.DEEPSEEK, LLMBackend.VLLM):
response = self._call_openai_compatible(prompt)
elif self.backend == LLMBackend.ANTHROPIC:
response = self._call_anthropic(prompt)
elif self.backend == LLMBackend.OLLAMA:
response = self._call_ollama(prompt)
elif self.backend == LLMBackend.GEMINI:
response = self._call_gemini(prompt)
else:
raise ValueError(f"Unknown backend: {self.backend}")
return self._parse_response(entry_key, response, context, abstract)
except Exception as e:
return EvaluationResult(
entry_key=entry_key,
relevance_score=0,
is_relevant=False,
explanation="",
context_used=context,
abstract_used=abstract,
error=str(e)
)
def _call_openai_compatible(self, prompt: str) -> str:
"""Call OpenAI-compatible API (OpenAI, DeepSeek, vLLM)."""
headers = {
"Content-Type": "application/json",
"Authorization": f"Bearer {self.api_key}"
}
payload = {
"model": self.model,
"messages": [
{"role": "user", "content": prompt}
],
"temperature": 0.1,
"max_tokens": 2000,
"response_format": {"type": "json_object"} if self.backend == LLMBackend.OPENAI else None
}
response = requests.post(
self.endpoint,
json=payload,
headers=headers,
timeout=60
)
response.raise_for_status()
data = response.json()
choices = data.get("choices", [])
if choices:
return choices[0].get("message", {}).get("content", "")
return ""
def _call_anthropic(self, prompt: str) -> str:
"""Call Anthropic API."""
headers = {
"x-api-key": self.api_key,
"anthropic-version": "2023-06-01",
"content-type": "application/json"
}
payload = {
"model": self.model,
"max_tokens": 2000,
"temperature": 0.1,
"messages": [
{"role": "user", "content": prompt}
]
}
response = requests.post(
self.endpoint,
json=payload,
headers=headers,
timeout=60
)
response.raise_for_status()
data = response.json()
content = data.get("content", [])
if content and content[0].get("type") == "text":
return content[0].get("text", "")
return ""
def _call_ollama(self, prompt: str) -> str:
"""Call Ollama API."""
payload = {
"model": self.model,
"prompt": prompt,
"stream": False,
"options": {
"temperature": 0.1,
"num_predict": 2000
},
"format": "json"
}
response = requests.post(
self.endpoint,
json=payload,
timeout=60
)
response.raise_for_status()
return response.json().get("response", "")
def _call_gemini(self, prompt: str) -> str:
"""Call Gemini API."""
# Build URL with model
url = f"{self.endpoint}/{self.model}:generateContent"
if self.api_key:
url += f"?key={self.api_key}"
payload = {
"contents": [
{
"parts": [
{"text": prompt}
]
}
],
"generationConfig": {
"temperature": 0.1,
"maxOutputTokens": 2000,
"responseMimeType": "application/json"
}
}
response = requests.post(
url,
json=payload,
timeout=60
)
response.raise_for_status()
candidates = response.json().get("candidates", [])
if candidates:
content = candidates[0].get("content", {})
parts = content.get("parts", [])
if parts:
return parts[0].get("text", "")
return ""
def _parse_response(self, entry_key: str, response: str, context: str, abstract: str) -> EvaluationResult:
"""Parse LLM response."""
# Try to extract JSON from response
json_match = re.search(r'\{[^{}]*\}', response, re.DOTALL)
data = {}
if not json_match:
# Try to parse the whole response as JSON
try:
data = json.loads(response.strip())
except json.JSONDecodeError:
pass
else:
try:
data = json.loads(json_match.group())
except json.JSONDecodeError:
pass
if not data:
return EvaluationResult(
entry_key=entry_key,
relevance_score=0,
is_relevant=False,
explanation=response,
context_used=context,
abstract_used=abstract,
error="Failed to parse LLM response as JSON"
)
# Extract fields
relevance_score = data.get("relevance_score", 0)
if isinstance(relevance_score, str):
try:
relevance_score = int(relevance_score)
except ValueError:
relevance_score = 0
is_relevant = data.get("is_relevant", False)
if isinstance(is_relevant, str):
is_relevant = is_relevant.lower() in ("true", "yes", "1")
explanation = data.get("explanation", "")
return EvaluationResult(
entry_key=entry_key,
relevance_score=relevance_score,
is_relevant=is_relevant,
explanation=explanation,
context_used=context,
abstract_used=abstract
)
def test_connection(self) -> bool:
"""Test if LLM backend is accessible."""
try:
if self.backend == LLMBackend.OLLAMA:
response = requests.get(
self.endpoint.replace("/api/generate", "/api/tags"),
timeout=5
)
return response.status_code == 200
elif self.backend in (LLMBackend.OPENAI, LLMBackend.DEEPSEEK, LLMBackend.VLLM):
# Test with a simple model list or empty completion
headers = {"Authorization": f"Bearer {self.api_key}"}
# Try listing models if possible, otherwise simple completion
if "chat/completions" in self.endpoint:
# Try a minimal completion
payload = {
"model": self.model,
"messages": [{"role": "user", "content": "hi"}],
"max_tokens": 1
}
response = requests.post(self.endpoint, json=payload, headers=headers, timeout=10)
return response.status_code == 200
else:
return False
elif self.backend == LLMBackend.ANTHROPIC:
headers = {
"x-api-key": self.api_key,
"anthropic-version": "2023-06-01",
"content-type": "application/json"
}
payload = {
"model": self.model,
"max_tokens": 1,
"messages": [{"role": "user", "content": "hi"}]
}
response = requests.post(self.endpoint, json=payload, headers=headers, timeout=10)
return response.status_code == 200
elif self.backend == LLMBackend.GEMINI:
if not self.api_key:
return False
url = f"{self.endpoint}/{self.model}:generateContent?key={self.api_key}"
payload = {
"contents": [{"parts": [{"text": "test"}]}],
"generationConfig": {"maxOutputTokens": 10}
}
response = requests.post(url, json=payload, timeout=10)
return response.status_code == 200
except Exception:
return False
return False
|