# ============================================================ # app.py – AI Knowledge Assistant (RAG) Flask Backend # ============================================================ import os, re, textwrap, shutil, csv, time import numpy as np from PyPDF2 import PdfReader import torch from transformers import AutoTokenizer, AutoModel, AutoModelForCausalLM from openai import OpenAI from docx import Document from flask import Flask, request, jsonify, send_from_directory app = Flask(__name__, static_folder='static', static_url_path='/') OPENROUTER_KEY = os.environ["OPENROUTER_KEY"] # ---------- Embedding Model ---------- model_name = "sentence-transformers/all-MiniLM-L6-v2" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModel.from_pretrained(model_name) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") model = model.to(device) def mean_pooling(model_output, attention_mask): token_embeddings = model_output[0] input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float() return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9) def embed_sentences(sentences, batch_size=32): all_emb = [] for i in range(0, len(sentences), batch_size): batch = sentences[i:i+batch_size] encoded = tokenizer(batch, padding=True, truncation=True, max_length=128, return_tensors='pt').to(device) with torch.no_grad(): out = model(**encoded) emb = mean_pooling(out, encoded['attention_mask']) all_emb.append(emb.cpu().numpy()) return np.concatenate(all_emb, axis=0) class CustomEmbedder: def encode(self, texts, **kwargs): if isinstance(texts, str): texts = [texts] return embed_sentences(texts) embedder = CustomEmbedder() # ---------- Vector Store ---------- class SimpleVectorStore: def __init__(self): self.documents = [] self.embeddings = [] def add(self, embeddings, documents, ids): for emb, doc in zip(embeddings, documents): self.embeddings.append(np.array(emb)) self.documents.append(doc) def query(self, query_embedding, n_results=3): query_vec = np.array(query_embedding).flatten() if not self.embeddings: return [] matrix = np.stack(self.embeddings) dot = np.dot(matrix, query_vec) norms = np.linalg.norm(matrix, axis=1) * np.linalg.norm(query_vec) sims = dot / (norms + 1e-9) top = np.argsort(sims)[::-1][:n_results] return [self.documents[i] for i in top] store = SimpleVectorStore() # ---------- Text Extraction (PDF, DOCX, TXT, CSV) ---------- def extract_text_from_file(file_path): fname = file_path.lower() text = "" if fname.endswith('.pdf'): reader = PdfReader(file_path) for page in reader.pages: t = page.extract_text() if t: text += t + "\n" elif fname.endswith('.docx'): doc = Document(file_path) for para in doc.paragraphs: text += para.text + "\n" elif fname.endswith('.txt'): with open(file_path, 'r', encoding='utf-8') as f: text = f.read() elif fname.endswith('.csv'): with open(file_path, 'r', encoding='utf-8') as f: reader = csv.reader(f) for row in reader: text += ", ".join(row) + "\n" else: with open(file_path, 'rb') as f: raw = f.read() try: text = raw.decode('utf-8') except: text = raw.decode('latin-1', errors='ignore') return text def split_text(text, chunk_size=500, overlap=50): chunks = [] for i in range(0, len(text), chunk_size - overlap): chunk = text[i:i+chunk_size] if len(chunk) > 100: chunks.append(chunk) return chunks def ingest_document(file_path): text = extract_text_from_file(file_path) chunks = split_text(text) if not chunks: return 0 emb = embedder.encode(chunks).tolist() store.add(emb, chunks, [f"doc_{i}" for i in range(len(chunks))]) return len(chunks) def retrieve(query, top_k=3): q_emb = embedder.encode([query]).tolist() return store.query(q_emb[0], n_results=top_k) # ---------- LLM Functions ---------- def setup_gemini(): return OpenAI(base_url="https://openrouter.ai/api/v1", api_key=OPENROUTER_KEY) def gemini_answer(client, context, question): prompt = f"""You are an AI assistant. Answer STRICTLY from the context. If answer not in context, say 'Not found in documents'. Context: {context} Question: {question} Answer with source citations.""" models = [ "google/gemma-4-31b-it:free", "google/gemini-2.0-flash-exp:free", "google/gemma-2-2b-it:free", "google/gemini-1.5-flash:free" ] for m in models: try: resp = client.chat.completions.create( model=m, messages=[{"role":"user","content":prompt}], temperature=0.1) return resp.choices[0].message.content except Exception as e: if "429" in str(e) or "rate" in str(e).lower(): time.sleep(5) return "All models busy. Try later." def setup_openrouter(): return OpenAI(base_url="https://openrouter.ai/api/v1", api_key=OPENROUTER_KEY) def openrouter_answer(client, context, question): prompt = f"""You are an AI assistant. Answer STRICTLY from the context. If answer not in context, say 'Not found in documents'. Context: {context} Question: {question}""" resp = client.chat.completions.create( model="openrouter/owl-alpha", messages=[{"role":"user","content":prompt}], temperature=0.1) return resp.choices[0].message.content def setup_local_model(): model_name = "gpt2" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForCausalLM.from_pretrained(model_name) tokenizer.pad_token = tokenizer.eos_token if torch.cuda.is_available(): model = model.to("cuda") return model, tokenizer def local_answer(model, tokenizer, context, question): prompt = f"Context: {context}\nQuestion: {question}\nAnswer:" inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512) if torch.cuda.is_available(): inputs = {k: v.to("cuda") for k, v in inputs.items()} outputs = model.generate( **inputs, max_new_tokens=100, do_sample=False, pad_token_id=tokenizer.eos_token_id) full = tokenizer.decode(outputs[0], skip_special_tokens=True) return full.split("Answer:")[-1].strip() # ---------- Routes ---------- @app.route('/') def index(): return send_from_directory('.', 'index.html') @app.route('/upload', methods=['POST']) def upload(): try: file = request.files['file'] file_path = f"/tmp/{file.filename}" file.save(file_path) num = ingest_document(file_path) return jsonify({"message": f"✅ {num} chunks processed." if num else "❌ No text could be extracted."}) except Exception as e: return jsonify({"message": f"❌ Error: {str(e)}"}) @app.route('/chat', methods=['POST']) def chat(): data = request.get_json() query = data.get('query','') choice = data.get('model_choice','local') try: chunks = retrieve(query, top_k=3) ctx = "\n\n".join(chunks) if choice == "gemini": ans = gemini_answer(setup_gemini(), ctx, query) elif choice == "openrouter": ans = openrouter_answer(setup_openrouter(), ctx, query) else: m, t = setup_local_model() ans = local_answer(m, t, ctx, query) return jsonify({"answer": ans, "sources": chunks}) except Exception as e: return jsonify({"answer": f"Error: {str(e)}", "sources": []}) if __name__ == "__main__": app.run(host="0.0.0.0", port=7860)