59 lines
3.2 KiB
Python
59 lines
3.2 KiB
Python
|
|
||
|
|
import json, torch, numpy as np
|
||
|
|
from sentence_transformers import SentenceTransformer, CrossEncoder
|
||
|
|
import faiss
|
||
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||
|
|
|
||
|
|
class PapersRAG:
|
||
|
|
def __init__(self, model_dir="."):
|
||
|
|
with open(f"{model_dir}/rag_config.json") as f:
|
||
|
|
config = json.load(f)
|
||
|
|
self.embedder = SentenceTransformer(config["embedder_model"])
|
||
|
|
self.index = faiss.read_index(f"{model_dir}/papersrag_index.faiss")
|
||
|
|
with open(f"{model_dir}/chunks.txt", "r", encoding="utf-8") as f:
|
||
|
|
raw = f.read().split("<|CHUNK_END|>")
|
||
|
|
self.chunks = [c.strip() for c in raw if c.strip()]
|
||
|
|
self.reranker = CrossEncoder(f"{model_dir}/cross_encoder_model")
|
||
|
|
self.tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
|
||
|
|
self.model = AutoModelForCausalLM.from_pretrained(
|
||
|
|
model_dir,
|
||
|
|
torch_dtype=torch.float16,
|
||
|
|
device_map="auto",
|
||
|
|
trust_remote_code=True
|
||
|
|
)
|
||
|
|
|
||
|
|
def ask(self, question, max_tokens=400):
|
||
|
|
q = question.strip().lower().rstrip('?!.')
|
||
|
|
greetings = ["hi", "hello", "hey", "yo", "sup", "good morning", "how are you"]
|
||
|
|
if any(q == g or q.startswith(g) for g in greetings):
|
||
|
|
return "Hello! I'm PapersRAG, your AI research assistant. I have 50 recent arXiv papers on computational linguistics and NLP. Ask me anything about them!"
|
||
|
|
identity_qs = ["who are you", "what is your name", "what are you", "what do you do", "tell me about yourself"]
|
||
|
|
if any(idq in q for idq in identity_qs):
|
||
|
|
return "I'm PapersRAG 🧪, a research assistant that can answer questions about the latest 50 arXiv papers in cs.CL. I'll cite the paper titles in my answers. Ask me anything about the papers!"
|
||
|
|
|
||
|
|
q_emb = self.embedder.encode([question]).astype("float32")
|
||
|
|
_, indices = self.index.search(q_emb, 10)
|
||
|
|
candidates = [self.chunks[i] for i in indices[0]]
|
||
|
|
pairs = [(question, c) for c in candidates]
|
||
|
|
scores = self.reranker.predict(pairs)
|
||
|
|
if max(scores) < -4.5:
|
||
|
|
return "I don't have enough information from my arXiv papers to answer that accurately. Try asking about specific NLP or computational linguistics papers."
|
||
|
|
best = sorted(zip(scores, candidates), reverse=True)[:4]
|
||
|
|
context = "\\n\\n".join([c for _, c in best])
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{"role": "system", "content": "You are PapersRAG, a scientific research assistant. Use ONLY the provided paper abstracts to answer. Always mention the paper title when you use information from it. If unsure, say you don't have that information."},
|
||
|
|
{"role": "user", "content": f"Context:\\n{context}\\n\\nQuestion: {question}"}
|
||
|
|
]
|
||
|
|
prompt = self.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||
|
|
inputs = self.tokenizer(prompt, return_tensors="pt").to(self.model.device)
|
||
|
|
outputs = self.model.generate(
|
||
|
|
**inputs,
|
||
|
|
max_new_tokens=max_tokens,
|
||
|
|
temperature=0.7,
|
||
|
|
do_sample=True,
|
||
|
|
pad_token_id=self.tokenizer.eos_token_id
|
||
|
|
)
|
||
|
|
answer = self.tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
|
||
|
|
return answer.strip()
|