Spaces:
Sleeping
Sleeping
File size: 25,476 Bytes
2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb 5e65a8e b7c72d9 c2c2b86 b7c72d9 c2c2b86 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 2f9b6eb b7c72d9 c2c2b86 2f9b6eb b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 fd47d39 b7c72d9 6834be5 b7c72d9 2f9b6eb 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 da34b65 c2c2b86 2f9b6eb 6834be5 b7c72d9 3d54fb2 6834be5 c2c2b86 b7c72d9 c2c2b86 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 3d54fb2 6834be5 3d54fb2 6834be5 3d54fb2 6834be5 b7c72d9 3d54fb2 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 c8e5f2d 0143b47 c8e5f2d b7c72d9 6834be5 c8e5f2d b7c72d9 6834be5 2f9b6eb 6834be5 b7c72d9 c8e5f2d 6834be5 b7c72d9 c8e5f2d 6834be5 b7c72d9 6834be5 c8e5f2d 6834be5 c8e5f2d 6834be5 b7c72d9 6834be5 7c96c51 fc8d210 7c96c51 b7c72d9 6834be5 b7c72d9 7c96c51 0aeacb8 7c96c51 6834be5 7c96c51 cc4a4e5 7c96c51 6834be5 2f9b6eb 6834be5 b7c72d9 6834be5 b7c72d9 da34b65 b7c72d9 3d54fb2 b7c72d9 6834be5 b44c007 6834be5 b44c007 6834be5 3d54fb2 b7c72d9 6834be5 3d54fb2 25474c3 3d54fb2 6834be5 b7c72d9 6834be5 b7c72d9 3d54fb2 25474c3 6834be5 b7c72d9 6834be5 b7c72d9 3d54fb2 b7c72d9 6834be5 25474c3 b7c72d9 0143b47 3d54fb2 b7c72d9 6834be5 3d54fb2 6834be5 3d54fb2 b7c72d9 c2c2b86 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 c8e5f2d 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 6834be5 b7c72d9 f70a184 0143b47 f70a184 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 |
"""
Full Conversational RAG Pipeline for Agri-Critique
Includes: Session management, context-aware retrieval, memory management
Loads everything from HuggingFace Hub
"""
import os
import json
import sqlite3
import uuid
from datetime import datetime
from typing import List, Dict, Any
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
from sentence_transformers import SentenceTransformer
import faiss
import numpy as np
from huggingface_hub import hf_hub_download
import warnings
# Suppress PEFT warnings about unexpected keys in LoraConfig
warnings.filterwarnings("ignore", category=UserWarning, module="peft")
class ConversationManager:
"""Manages conversation sessions with persistent storage"""
def __init__(self, db_path="conversations.db"):
self.db_path = db_path
self.conn = sqlite3.connect(db_path, check_same_thread=False)
self.cursor = self.conn.cursor()
self._init_db()
def _init_db(self):
"""Initialize session database"""
self.cursor.execute("""
CREATE TABLE IF NOT EXISTS sessions (
session_id TEXT PRIMARY KEY,
created_at TEXT,
last_updated TEXT,
metadata TEXT
)
""")
self.cursor.execute("""
CREATE TABLE IF NOT EXISTS messages (
message_id INTEGER PRIMARY KEY AUTOINCREMENT,
session_id TEXT,
role TEXT,
content TEXT,
timestamp TEXT,
evidence TEXT,
FOREIGN KEY (session_id) REFERENCES sessions(session_id)
)
""")
self.conn.commit()
def create_session(self, metadata=None):
"""Create a new conversation session"""
session_id = str(uuid.uuid4())
now = datetime.utcnow().isoformat()
self.cursor.execute("""
INSERT INTO sessions (session_id, created_at, last_updated, metadata)
VALUES (?, ?, ?, ?)
""", (session_id, now, now, json.dumps(metadata or {})))
self.conn.commit()
return session_id
def add_message(self, session_id, role, content, evidence=None):
"""Add a message to a session"""
now = datetime.utcnow().isoformat()
self.cursor.execute("""
INSERT INTO messages (session_id, role, content, timestamp, evidence)
VALUES (?, ?, ?, ?, ?)
""", (session_id, role, content, now, json.dumps(evidence) if evidence else None))
# Update session timestamp
self.cursor.execute("""
UPDATE sessions SET last_updated = ? WHERE session_id = ?
""", (now, session_id))
self.conn.commit()
def get_session_history(self, session_id, limit=None):
"""Get conversation history for a session"""
query = """
SELECT role, content, timestamp, evidence
FROM messages
WHERE session_id = ?
ORDER BY timestamp ASC
"""
if limit:
query += f" LIMIT {limit}"
self.cursor.execute(query, (session_id,))
messages = []
for row in self.cursor.fetchall():
messages.append({
'role': row[0],
'content': row[1],
'timestamp': row[2],
'evidence': json.loads(row[3]) if row[3] else None
})
return messages
def summarize_old_messages(self, session_id, keep_recent=4):
"""Summarize old messages to save context window"""
messages = self.get_session_history(session_id)
if len(messages) <= keep_recent:
return messages
# Keep recent messages
recent = messages[-keep_recent:]
old = messages[:-keep_recent]
# Create summary of old messages
summary = "Previous conversation summary:\n"
for msg in old[::2]: # Sample every other message
summary += f"- {msg['role']}: {msg['content'][:100]}...\n"
# Return summary + recent messages
return [{'role': 'system', 'content': summary}] + recent
class AgriCritiqueRAG:
"""Full RAG system with conversational capabilities"""
def __init__(self):
print("π Initializing Agri-Critique Conversational RAG System...")
# Model paths
self.model_id = "sayande/AgriScholarQA-CoT"
self.base_model_id = "Qwen/Qwen3-4B-Thinking-2507"
self.index_repo = "sayande/agri-critique-index"
# Conversation manager
self.conversation_manager = ConversationManager()
self.current_session = None
# Load retriever
print("π₯ Loading retriever...")
self.retriever = SentenceTransformer("all-mpnet-base-v2")
# ------------------------------------------------------------------
# Load FAISS indices (local first, then HF fallback)
# ------------------------------------------------------------------
print("π₯ Loading FAISS indices...")
self.chunk_index = None
self.paper_index = None
self.index = None # alias kept for backward compatibility
base_dir = os.path.dirname(__file__) if "__file__" in globals() else os.getcwd()
local_chunk_path = os.path.join(base_dir, "faiss.index")
local_paper_path = os.path.join(base_dir, "faiss_papers.index")
local_meta_path = os.path.join(base_dir, "meta.json")
# ---- Try LOCAL chunk index ----
try:
if os.path.exists(local_chunk_path):
print(f"π Found local chunk index: {local_chunk_path}")
self.chunk_index = faiss.read_index(local_chunk_path)
self.index = self.chunk_index
print(f"β
Loaded local chunk FAISS index with {self.chunk_index.ntotal} vectors")
else:
print("βΉοΈ Local chunk index 'faiss.index' not found, will try HuggingFace Hub...")
except Exception as e:
print(f"β οΈ Could not load local chunk index: {e}")
self.chunk_index = None
self.index = None
# ---- If no local chunk index, fall back to HF ----
if self.chunk_index is None:
print("π₯ Loading FAISS index from HuggingFace dataset...")
try:
index_path = hf_hub_download(
repo_id=self.index_repo,
filename="faiss.index",
repo_type="dataset"
)
self.chunk_index = faiss.read_index(index_path)
self.index = self.chunk_index
print(f"β
Loaded HF FAISS index with {self.chunk_index.ntotal} vectors")
except Exception as e:
print(f"β οΈ Could not load FAISS index from HF: {e}")
self.chunk_index = None
self.index = None
# ---- Optional: paper-level index (not strictly required) ----
try:
if os.path.exists(local_paper_path):
print(f"π Found local paper index: {local_paper_path}")
self.paper_index = faiss.read_index(local_paper_path)
print(f"β
Loaded local paper FAISS index with {self.paper_index.ntotal} vectors")
else:
print("βΉοΈ Local paper index 'faiss_papers.index' not found (this is optional).")
except Exception as e:
print(f"β οΈ Could not load local paper index: {e}")
self.paper_index = None
# ------------------------------------------------------------------
# Load metadata (local first, then HF)
# ------------------------------------------------------------------
print("π₯ Loading metadata...")
self.metadata = []
# Try local meta.json
try:
if os.path.exists(local_meta_path):
print(f"π Found local metadata: {local_meta_path}")
with open(local_meta_path, "r", encoding="utf-8") as f:
self.metadata = json.load(f)
print(f"β
Loaded local metadata for {len(self.metadata)} chunks")
else:
print("βΉοΈ Local 'meta.json' not found, will try HuggingFace Hub...")
except Exception as e:
print(f"β οΈ Could not load local metadata: {e}")
self.metadata = []
# If still empty, try HF
if not self.metadata:
print("π₯ Loading metadata from HuggingFace dataset...")
try:
meta_path = hf_hub_download(
repo_id=self.index_repo,
filename="meta.json",
repo_type="dataset"
)
with open(meta_path, "r", encoding="utf-8") as f:
self.metadata = json.load(f)
print(f"β
Loaded HF metadata for {len(self.metadata)} chunks")
except Exception as e:
print(f"β οΈ Could not load metadata from HF: {e}")
self.metadata = []
# Model will be loaded lazily on first use
self.model = None
self.tokenizer = None
self.model_loaded = False
print("β
Agri-Critique Conversational RAG System initialized!")
print("βΉοΈ Model will load on first query (Qwen3-4B with INT8 quantization)")
def _ensure_model_loaded(self):
"""Lazy load model on first use"""
if self.model_loaded:
return
print("π₯ Loading Agri-Critique model (this may take 1-2 minutes)...")
# Get HF token from environment
hf_token = os.getenv("HF_TOKEN")
if not hf_token:
raise ValueError("HF_TOKEN not found. Please add it to Space secrets.")
self.tokenizer = AutoTokenizer.from_pretrained(
self.base_model_id,
token=hf_token
)
from transformers import AutoConfig
config = AutoConfig.from_pretrained(self.base_model_id, token=hf_token)
# Qwen models work well with default config
# No special rope_scaling adjustments needed
print("π₯οΈ Loading Qwen3-4B model with INT4 quantization for speed")
# Try to use INT4 quantization for faster inference (better for 4B models on CPU)
try:
from transformers import BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(
load_in_4bit=True, # INT4 is better for larger models on CPU
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4"
)
print("β
Using INT4 (NF4) quantization - optimized for Qwen 4B on CPU")
except ImportError:
print("β οΈ bitsandbytes not available, using float32")
quantization_config = None
base_model = AutoModelForCausalLM.from_pretrained(
self.base_model_id,
config=config,
quantization_config=quantization_config,
torch_dtype=torch.float32 if quantization_config is None else None,
device_map="auto" if quantization_config else "cpu",
low_cpu_mem_usage=True,
token=hf_token,
)
self.model = PeftModel.from_pretrained(
base_model,
self.model_id,
token=hf_token,
)
self.model.eval()
self.model_loaded = True
print("β
Model loaded successfully!")
def _refine_query_with_llm(self, query):
"""Use LLM to extract core search terms (Query Understanding/NER)"""
if not self.model_loaded:
return query # Can't refine if model not loaded yet
prompt = [
{"role": "system", "content": "You are a search query optimizer. Extract ONLY the most important agricultural keywords, entities (crops, diseases, chemicals), and timeframes from the user's question. Return a concise string of keywords."},
{"role": "user", "content": f"Query: {query}"}
]
try:
input_text = self.tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=True)
inputs = self.tokenizer(input_text, return_tensors="pt").to(self.model.device)
with torch.no_grad():
outputs = self.model.generate(
**inputs,
max_new_tokens=256, # Short output
temperature=0.3
)
refined = self.tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()
# print(f"DEBUG: Refined '{query}' -> '{refined}'")
return refined
except Exception:
# Fallback
return query
def start_session(self, metadata=None):
"""Start a new conversation session"""
self.current_session = self.conversation_manager.create_session(metadata)
print(f"π Started new session: {self.current_session[:8]}...")
return self.current_session
def retrieve_with_context(self, query, conversation_history, top_k=5):
"""Context-aware retrieval: considers conversation history"""
# Use chunk_index (local or HF). If missing, no retrieval.
if self.chunk_index is None or not self.metadata:
return []
# Combine current query with recent context
context_queries = [query]
# EXPERIMENTAL: Query Understanding / Refinement
# If the model is already loaded, we can use it to "understand" the query
# and extract better search terms (NER-lite).
if self.model_loaded:
refined = self._refine_query_with_llm(query)
if refined and refined != query:
# Add refined keywords with high importance
context_queries.append(refined)
# Add recent user questions for context
for msg in conversation_history[-4:] if conversation_history else []:
if msg["role"] == "user":
context_queries.append(msg["content"])
# Encode all queries
embeddings = self.retriever.encode(context_queries, convert_to_numpy=True)
# Weighting: Original Query (high), Refined Query (med), History (low)
# Dynamic weighting based on what we have
num_q = len(context_queries)
if num_q == 1:
weights = [1.0]
else:
# Simple heuristic: First item (Original) gets 0.6
# Others share the remaining 0.4
weights = [0.6] + [0.4 / (num_q - 1)] * (num_q - 1)
weighted_embedding = np.average(embeddings, axis=0, weights=weights).reshape(1, -1).astype("float32")
faiss.normalize_L2(weighted_embedding)
# Extract year from query (e.g., 2024, 2025)
import re
year_match = re.search(r"\b(20\d{2})\b", query)
target_year = year_match.group(1) if year_match else None
# Search over chunk index
# Fetch more candidates to allow for temporal re-ranking
# If year detected, fetch deep (e.g. 100) to find the year-match chunks
if target_year:
initial_k = 100
else:
initial_k = top_k * 3
distances, indices = self.chunk_index.search(weighted_embedding.astype("float32"), initial_k)
candidates = []
for idx, dist in zip(indices[0], distances[0]):
if 0 <= idx < len(self.metadata):
chunk_info = self.metadata[idx]
# Check for year match in paper_id
is_year_match = False
if target_year and target_year in chunk_info.get("paper_id", ""):
is_year_match = True
candidates.append({
"data": chunk_info,
"dist": float(dist),
"is_year_match": is_year_match
})
# Soft Boost Logic:
# Instead of force-sorting year matches to the top (which brings in irrelevant junk),
# we improve their distance score by a fixed amount (e.g., 0.5).
# Assuming L2 distance (smaller is better): new_dist = old_dist - 0.5
# This lets a "Relevant Year-Match" beat "Relevant Non-Match",
# but a "Totally Irrelevant Year-Match" will still lose to "Relevant Content".
for cand in candidates:
if cand["is_year_match"]:
cand["effective_dist"] = cand["dist"] - 0.5
else:
cand["effective_dist"] = cand["dist"]
# Sort by effective distance (ascending)
candidates.sort(key=lambda x: x["effective_dist"])
# Select top_k
final_candidates = candidates[:top_k]
evidence = []
for cand in final_candidates:
ev = dict(cand["data"])
ev["score"] = cand["dist"]
# FALLBACK: If 'text' is missing in metadata (common issue with this dataset version),
# construct a proxy text from the section and paper ID so the RAG doesn't see empty strings.
if "text" not in ev or not ev["text"]:
paper = ev.get("paper_id", "Unknown Paper")
sect = ev.get("section", "General")
ev["text"] = f"[Note: Full text missing in metadata] Section '{sect}' from paper '{paper}'."
evidence.append(ev)
return evidence
def _clean_paper_id(self, paper_id):
"""Clean paper ID for display"""
if not isinstance(paper_id, str):
return str(paper_id)
clean = paper_id.strip("-_")
clean = clean.replace("_", " ").replace("-", " ")
return clean.title()
def validate_and_answer(self, question, evidence, conversation_history):
"""Generate validated answer with reasoning - OPTIMIZED single-call version"""
self._ensure_model_loaded()
# Format evidence text for the model
# Include title/paper_id so the model knows the source date/context
evidence_text = "\n\n".join(
[
f"[{i+1}] {ev.get('paper_title') or ev.get('paper_id')}\n{ev.get('text', '')}"
for i, ev in enumerate(evidence)
]
)
# OPTIMIZED: Single model call for both validation and answer
# This reduces inference time by ~50%
combined_messages = [
{
"role": "system",
"content": (
"You are an agricultural research assistant. Your task is to:\n"
"1. Validate the question against the evidence\n"
"2. Provide a clear, comprehensive answer based ONLY on the evidence\n"
"3. Cite sources as [1], [2], etc.\n\n"
"Check: Is the question relevant? Are there conflicting facts? "
"Is there enough information?"
),
},
{
"role": "user",
"content": f"""EVIDENCE:
{evidence_text}
QUESTION: {question}
TASK: Provide a validated answer to the question. First briefly explain your reasoning, then give the final answer. Be detailed and thorough.""",
},
]
input_text = self.tokenizer.apply_chat_template(
combined_messages, tokenize=False, add_generation_prompt=True
)
inputs = self.tokenizer(input_text, return_tensors="pt").to(self.model.device)
# UPDATED: Increased token limit for more detailed answers
with torch.no_grad():
outputs = self.model.generate(
**inputs,
max_new_tokens=2048, # Increased from 256 for detail
temperature=0.5, # Balanced temperature
do_sample=True,
pad_token_id=self.tokenizer.pad_token_id,
)
full_response = self.tokenizer.decode(
outputs[0][inputs["input_ids"].shape[1] :],
skip_special_tokens=True,
).strip()
# Split response into reasoning and answer
# The model should naturally provide reasoning first, then answer
if "\n\n" in full_response:
parts = full_response.split("\n\n", 1)
reasoning = parts[0]
answer = parts[1] if len(parts) > 1 else full_response
else:
# Fallback: use entire response as answer
reasoning = "Validated against evidence."
answer = full_response
return reasoning, answer
def ask(self, question):
"""Full RAG pipeline with validation"""
if not self.current_session:
self.start_session()
# Get conversation history
history = self.conversation_manager.get_session_history(self.current_session)
# Context-aware retrieval
evidence = self.retrieve_with_context(question, history, top_k=5)
# Generate validated answer
reasoning, answer = self.validate_and_answer(question, evidence, history)
# Save to session
self.conversation_manager.add_message(
self.current_session, "user", question, evidence
)
self.conversation_manager.add_message(
self.current_session, "assistant", answer
)
# Format evidence for return (include all useful fields)
formatted_evidence = []
for ev in evidence:
paper_id = ev.get("paper_id", "unknown")
display_paper = ev.get("paper_title") or self._clean_paper_id(paper_id)
formatted_evidence.append(
{
"paper_id": display_paper,
"raw_paper_id": paper_id,
"text": ev.get("text", ""),
"score": ev.get("score", 0.0),
}
)
return {
"answer": answer,
"reasoning": reasoning,
"evidence": formatted_evidence,
"session_id": self.current_session,
}
def validate_answer(self, question: str, proposed_answer: str, evidence: List[Dict[str, Any]]) -> str:
"""
Validate a proposed answer against evidence and return critique.
This method uses the fine-tuned Llama model to critique the answer by checking:
- Are all claims supported by the evidence?
- Are there any hallucinations or fake findings?
- Are citations accurate?
- Are there temporal or causal errors?
Args:
question: The original question
proposed_answer: The answer to validate
evidence: List of evidence chunks with 'text' field
Returns:
Critique string identifying issues or confirming validity
"""
self._ensure_model_loaded()
# Format evidence text for validation
evidence_text = "\n\n".join([
f"[{i+1}] {ev.get('text', '')}"
for i, ev in enumerate(evidence[:5]) # Limit to top 5 for context window
])
if not evidence_text.strip():
evidence_text = "(no evidence provided)"
# Create validation prompt
validation_messages = [
{
"role": "system",
"content": (
"You are a strict agricultural research validator. "
"Your job is to critique the proposed answer by checking:\n"
"1. Are all claims supported by the evidence?\n"
"2. Are there any hallucinations or fake findings?\n"
"3. Are citations accurate and properly used?\n"
"4. Are there temporal or causal errors?\n"
"5. Are there any unsupported extrapolations?\n\n"
"Provide a concise critique. If the answer is well-supported, say so. "
"If there are issues, clearly identify them."
),
},
{
"role": "user",
"content": f"""QUESTION: {question}
EVIDENCE:
{evidence_text}
PROPOSED ANSWER:
{proposed_answer}
TASK: Critique this answer. Identify any unsupported claims, hallucinations, citation errors, or other issues.""",
},
]
# Generate critique using the model
input_text = self.tokenizer.apply_chat_template(
validation_messages, tokenize=False, add_generation_prompt=True
)
inputs = self.tokenizer(input_text, return_tensors="pt").to(self.model.device)
with torch.no_grad():
outputs = self.model.generate(
**inputs,
max_new_tokens=1024,
temperature=0.3,
do_sample=True,
pad_token_id=self.tokenizer.pad_token_id,
)
critique = self.tokenizer.decode(
outputs[0][inputs["input_ids"].shape[1]:],
skip_special_tokens=True,
).strip()
return critique
|