import json import os from typing import List, Dict, Any, Optional, Generator from dataclasses import dataclass from enum import Enum import hashlib import logging from datetime import datetime import chromadb from chromadb.config import Settings from sentence_transformers import SentenceTransformer from huggingface_hub import hf_hub_download # Import the base LLM handler from llm_handler import CybersecurityLLM logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # ================================================ # Security Knowledge Definitions # ================================================ class SecurityTopic(Enum): PHISHING = "phishing" PASSWORDS = "passwords" MALWARE = "malware" SOCIAL_ENGINEERING = "social_engineering" DATA_PROTECTION = "data_protection" NETWORK_SECURITY = "network_security" INCIDENT_RESPONSE = "incident_response" PHYSICAL_SECURITY = "physical_security" MOBILE_SECURITY = "mobile_security" CLOUD_SECURITY = "cloud_security" COMPLIANCE = "compliance" EMAIL_SECURITY = "email_security" RANSOMWARE = "ransomware" ZERO_TRUST = "zero_trust" SUPPLY_CHAIN = "supply_chain" @dataclass class SecurityKnowledge: topic: SecurityTopic title: str content: str keywords: List[str] severity: str # low, medium, high, critical last_updated: str = "" def to_dict(self) -> Dict[str, Any]: return { "topic": self.topic.value, "title": self.title, "content": self.content, "keywords": json.dumps(self.keywords), # Serialize list to JSON string "severity": self.severity, "last_updated": self.last_updated or datetime.now().isoformat() } # ================================================ # Main Knowledge Base Class # ================================================ class CybersecurityKnowledgeBase: def __init__(self, persist_directory: str = "./knowledge_db", embedding_model: str = "all-MiniLM-L6-v2"): """ Initialize knowledge base with vector database Args: persist_directory: Directory to persist ChromaDB embedding_model: Sentence transformer model for embeddings """ logger.info(f"Initializing knowledge base at {persist_directory}") # Create directory if it doesn't exist os.makedirs(persist_directory, exist_ok=True) # Initialize ChromaDB with persistence self.client = chromadb.PersistentClient( path=persist_directory, settings=Settings( anonymized_telemetry=False, allow_reset=True ) ) # Create or get collection try: self.collection = self.client.get_collection("cybersecurity_knowledge") logger.info(f"Loaded existing collection with {self.collection.count()} documents") except: self.collection = self.client.create_collection( name="cybersecurity_knowledge", metadata={"description": "Cybersecurity best practices and knowledge"} ) logger.info("Created new knowledge collection") # Initialize embedder logger.info(f"Loading embedding model: {embedding_model}") self.embedder = SentenceTransformer(embedding_model) # Load core knowledge if collection is empty if self.collection.count() == 0: logger.info("Loading core cybersecurity knowledge...") self._load_core_knowledge() # Track statistics self.stats = { "total_documents": self.collection.count(), "queries_processed": 0, "last_updated": datetime.now().isoformat() } def _load_core_knowledge(self): """Load comprehensive cybersecurity knowledge""" knowledge_items = [ # Phishing and Email Security SecurityKnowledge( topic=SecurityTopic.PHISHING, title="Comprehensive Phishing Detection Guide", content=""" IDENTIFYING PHISHING EMAILS - Complete Guide: Red Flags to Watch For: • Generic greetings: "Dear Customer" instead of your actual name • Urgency tactics: "Act now or your account will be closed!" • Grammar/spelling errors: Professional companies proofread their emails • Mismatched sender: Display name doesn't match email address • Suspicious links: Hover to see if URL matches claimed sender • Unexpected attachments: Especially .zip, .exe, .scr, .vbs files • Requests for sensitive info: Legitimate companies don't ask for passwords via email • Too good to be true: "You've won $1 million!" • Emotional manipulation: Fear, greed, curiosity, sympathy How to Verify Suspicious Emails: 1. Check sender's email address carefully (not just display name) 2. Hover over links WITHOUT clicking to preview destination 3. Look for HTTPS and correct domain in links 4. Contact company directly through official channels (not email links) 5. Check for personalization - legitimate emails often include account numbers 6. Verify with IT security team when in doubt What to Do If You Receive Phishing: 1. Don't click links or download attachments 2. Don't reply or provide any information 3. Report to IT security immediately 4. Forward to anti-phishing team if available 5. Delete the email after reporting 6. Warn colleagues if it's widespread If You Clicked a Phishing Link: 1. Disconnect from network immediately 2. Change passwords from a different device 3. Report to IT security IMMEDIATELY 4. Run antivirus scan 5. Monitor accounts for suspicious activity 6. Enable MFA on all accounts if not already done """, keywords=["phishing", "email", "scam", "suspicious", "link", "attachment", "spear phishing", "whaling", "BEC"], severity="critical" ), # Password Security SecurityKnowledge( topic=SecurityTopic.PASSWORDS, title="Password Security Best Practices", content=""" CREATING STRONG PASSWORDS: Requirements for Strong Passwords: • Minimum 12-16 characters (longer is better) • Mix of uppercase and lowercase letters • Include numbers and special characters (!@#$%^&*) • Avoid dictionary words and personal information • Unique for every account - never reuse passwords • Consider passphrases: 'Coffee@7Makes$Me!Happy2024' • Avoid patterns: Password1, Password2, etc. • Don't use keyboard patterns: qwerty, 123456 Password Management Best Practices: • Use a reputable password manager (Bitwarden, 1Password, LastPass) • Enable two-factor authentication (2FA) everywhere possible • Use authenticator apps over SMS when possible • Never share passwords via email, chat, or phone • Change passwords immediately if breach suspected • Don't write passwords on sticky notes • Use different passwords for work and personal accounts • Consider using hardware keys for critical accounts Multi-Factor Authentication (MFA): • Something you know (password) • Something you have (phone, token) • Something you are (biometric) Password Manager Benefits: • Generate random, unique passwords • Securely store all passwords • Auto-fill credentials safely • Sync across devices • Alert you to breaches • Share passwords securely when needed Common Password Mistakes: • Using personal information (birthdate, pet names) • Reusing passwords across sites • Sharing passwords with others • Using simple substitutions (P@ssw0rd) • Not updating default passwords • Ignoring breach notifications """, keywords=["password", "authentication", "2FA", "MFA", "login", "credentials", "passphrase", "password manager"], severity="critical" ), # Malware Prevention SecurityKnowledge( topic=SecurityTopic.MALWARE, title="Malware Prevention and Response", content=""" MALWARE PREVENTION STRATEGIES: Prevention Best Practices: • Keep OS and all software updated with latest patches • Use reputable antivirus with real-time protection • Enable Windows Defender or equivalent • Download software only from official sources • Verify digital signatures on downloads • Scan USB drives before opening files • Disable macros in Office documents from unknown sources • Use application sandboxing when possible • Regular backups following 3-2-1 rule • Keep UAC (User Account Control) enabled Types of Malware: • Viruses: Self-replicating, attaches to files • Worms: Self-spreading through networks • Trojans: Disguised as legitimate software • Ransomware: Encrypts files for ransom • Spyware: Steals information secretly • Adware: Displays unwanted advertisements • Rootkits: Hides presence from system • Keyloggers: Records keystrokes • Cryptominers: Uses resources to mine cryptocurrency Warning Signs of Infection: • Computer running unusually slow • Frequent crashes or blue screens • Programs starting automatically • Browser homepage changed • New toolbars or extensions • Excessive pop-ups • Files encrypted with ransom note • Unusual network activity • Disabled security software • Missing or modified files If Infected - Immediate Steps: 1. Disconnect from all networks (WiFi, Ethernet) 2. Enter Safe Mode if possible 3. Run full antivirus scan 4. Use additional malware removal tools (Malwarebytes) 5. Check for system restore points 6. Contact IT security team immediately 7. Change all passwords from clean device 8. Monitor financial accounts 9. Consider complete system reinstall for severe infections """, keywords=["malware", "virus", "ransomware", "trojan", "antivirus", "infection", "worm", "spyware"], severity="critical" ), # Social Engineering SecurityKnowledge( topic=SecurityTopic.SOCIAL_ENGINEERING, title="Social Engineering Defense Strategies", content=""" DEFENDING AGAINST SOCIAL ENGINEERING: Common Social Engineering Tactics: • Pretexting: Creating fake scenarios to steal information • Baiting: Offering something enticing (USB drives, downloads) • Quid pro quo: Offering service for information • Tailgating: Following into secure areas • Vishing: Voice phishing via phone • Smishing: SMS/text message phishing • Watering hole: Compromising frequently visited websites • Dumpster diving: Searching trash for information • Shoulder surfing: Looking over shoulder for passwords Red Flags to Recognize: • Unsolicited contact asking for information • Urgency without verification • Requests to bypass normal procedures • Appeals to authority without proof • Offers that seem too good to be true • Requests for passwords or sensitive data • Emotional manipulation (fear, greed, sympathy) • Name dropping without context • Resistance to verification Defense Strategies: • Always verify identity before sharing information • Use callback numbers from official sources • Be suspicious of unsolicited contacts • Never give passwords over phone/email • Question unusual requests, even from "colleagues" • Report suspicious behavior immediately • Trust but verify - confirm through separate channel • Be aware of information you share publicly • Secure physical documents and screens • Educate family about work-related scams Verification Techniques: • Call back on known number • Check employee directory • Verify with manager • Ask for employee ID • Request email confirmation • Check digital signatures • Verify through IT security """, keywords=["social engineering", "pretexting", "vishing", "smishing", "manipulation", "tailgating", "phishing"], severity="high" ), # Network Security SecurityKnowledge( topic=SecurityTopic.NETWORK_SECURITY, title="Network and WiFi Security Guide", content=""" NETWORK SECURITY BEST PRACTICES: Home WiFi Security: • Change default router admin credentials immediately • Use WPA3 encryption (WPA2 minimum) • Create strong WiFi password (20+ characters) • Change default network name (SSID) • Disable WPS (WiFi Protected Setup) • Keep router firmware updated monthly • Use guest network for visitors and IoT devices • Disable remote management unless necessary • Turn off SSID broadcast if practical • Use MAC address filtering for added security • Position router centrally to minimize external signal • Regular reboot router (monthly) Public WiFi Safety: • Avoid accessing sensitive accounts • Always use VPN for all connections • Verify network name with venue staff • Turn off automatic WiFi connection • Forget network after use • Never accept certificate warnings • Disable file sharing • Use cellular data for sensitive tasks • Keep firewall enabled • Use HTTPS websites only VPN Best Practices: • Use company-approved VPN only • Connect before accessing any resources • Keep VPN client updated • Report connection issues immediately • Don't use free/public VPN services • Verify VPN is active before working Network Hygiene: • Regular network scans for unknown devices • Monitor bandwidth usage • Check for unauthorized access points • Secure all network equipment physically • Document network configuration • Regular security audits """, keywords=["wifi", "network", "router", "VPN", "encryption", "WPA3", "public wifi", "wireless"], severity="high" ), # Incident Response SecurityKnowledge( topic=SecurityTopic.INCIDENT_RESPONSE, title="Security Incident Response Procedures", content=""" SECURITY INCIDENT RESPONSE GUIDE: IMMEDIATE RESPONSE STEPS: 1. STOP - Don't try to fix it yourself 2. DISCONNECT - Unplug network cable or disable WiFi 3. DOCUMENT - Write down: - What happened - When it occurred - What you were doing - Error messages - Unusual behavior observed 4. REPORT - Contact IT security immediately 5. PRESERVE - Don't delete anything, take screenshots 6. WAIT - For IT security instructions Types of Incidents Requiring Immediate Reporting: • Clicked suspicious link or attachment • Entered credentials on suspicious site • Lost device with company data • Suspicious computer behavior • Unauthorized access attempts • Data breach or leak discovered • Ransomware infection • Physical security breach • Stolen credentials • Suspicious phone calls asking for info Information to Provide: • Your name and contact information • Time and date of incident • Affected systems/accounts • Description of what happened • Actions taken so far • Any error messages (exact wording) • Screenshots if possible • Anyone else who might be affected DO NOT: • Try to fix it yourself • Delete or modify evidence • Inform unauthorized people • Post about it on social media • Continue using affected systems • Pay ransoms Contact Information: IT Security Hotline: [Organization specific] Email: security@[organization] After hours: [Emergency contact] """, keywords=["incident", "breach", "response", "report", "emergency", "compromise", "security incident"], severity="critical" ), # Data Protection SecurityKnowledge( topic=SecurityTopic.DATA_PROTECTION, title="Data Protection and Privacy Guide", content=""" DATA PROTECTION BEST PRACTICES: Data Classification: • Public: Can be freely shared • Internal: Within organization only • Confidential: Specific authorized individuals • Restricted: Highest sensitivity, strict controls Handling Sensitive Data: • Encrypt files before sharing externally • Use approved file sharing platforms only • Never use personal email for work data • Implement clean desk policy • Lock computer when stepping away (Win+L or Cmd+Ctrl+Q) • Use privacy screens in public spaces • Shred physical documents with sensitive info • Secure disposal of electronic media • Don't discuss sensitive info in public • Be aware of smart speakers/devices Encryption Best Practices: • Use full disk encryption (BitLocker, FileVault) • Encrypt removable media • Use encrypted communication channels • Encrypt email with sensitive data • Password protect sensitive documents • Use enterprise encryption tools • Store encryption keys securely Data Backup Practices: • Follow 3-2-1 rule: - 3 copies of important data - 2 different storage media - 1 offsite backup • Test restore procedures regularly • Encrypt backup drives • Store backups securely • Automate where possible • Document what's backed up • Verify backup integrity Privacy Considerations: • Minimize data collection • Only share need-to-know basis • Regular data audits • Respect retention policies • Secure data destruction • GDPR/CCPA compliance """, keywords=["data", "encryption", "backup", "confidential", "sensitive", "GDPR", "privacy", "classification"], severity="high" ), # Mobile Security SecurityKnowledge( topic=SecurityTopic.MOBILE_SECURITY, title="Mobile Device Security Guidelines", content=""" MOBILE DEVICE SECURITY: Device Security Settings: • Enable screen lock (PIN, password, biometric) • Set auto-lock to 1-2 minutes • Keep OS and apps updated automatically • Download apps only from official stores • Review app permissions carefully • Enable remote wipe capability • Use Find My Device features • Encrypt device storage • Disable Bluetooth when not needed • Turn off WiFi auto-connect • Disable Siri/Assistant on lock screen BYOD (Bring Your Own Device) Security: • Separate work and personal data • Use MDM if required by company • Install company security apps • Follow company mobile policy • Report lost/stolen immediately • Don't jailbreak/root devices • Use company VPN for work • Regular security updates Mobile Threats: • Malicious apps • Unsecured WiFi • SMiShing (SMS phishing) • Bluetooth attacks • Physical theft • Shoulder surfing • Juice jacking (USB charging) • SIM swapping Safe Mobile Practices: • Avoid public WiFi for sensitive tasks • Use VPN when on public networks • Don't click links in text messages • Be cautious with QR codes • Use official app stores only • Keep personal info private • Regular app permission audits • Backup device regularly • Use mobile antivirus • Avoid charging at public USB ports """, keywords=["mobile", "smartphone", "tablet", "BYOD", "iOS", "Android", "app security", "MDM"], severity="medium" ), # Ransomware Specific SecurityKnowledge( topic=SecurityTopic.RANSOMWARE, title="Ransomware Prevention and Response", content=""" RANSOMWARE PROTECTION GUIDE: Prevention Strategies: • Regular automated backups (tested restores) • Keep all software patched and updated • Email filtering and sandboxing • Disable macros by default • User training on phishing • Network segmentation • Principle of least privilege • Application whitelisting • Endpoint detection and response (EDR) If Ransomware Strikes: 1. Immediately disconnect from network 2. Power off if actively encrypting 3. Report to IT security immediately 4. Do NOT pay ransom 5. Preserve evidence for investigation 6. Check for decryption tools 7. Restore from clean backups 8. Rebuild affected systems 9. Investigate root cause 10. Implement lessons learned Warning Signs: • Files with strange extensions • Cannot open documents • Ransom notes in folders • Slow computer performance • Renamed files • Wallpaper changed to ransom message Recovery Planning: • Maintain offline backups • Test restore procedures • Document critical systems • Incident response plan • Communication plan • Legal/law enforcement contacts """, keywords=["ransomware", "encryption", "ransom", "backup", "recovery", "bitcoin", "crypto"], severity="critical" ), # Cloud Security SecurityKnowledge( topic=SecurityTopic.CLOUD_SECURITY, title="Cloud Services Security", content=""" CLOUD SECURITY BEST PRACTICES: Account Security: • Use strong, unique passwords • Enable MFA on all cloud accounts • Regular access reviews • Monitor for unusual activity • Use SSO where available • Secure API keys and tokens Data Protection in Cloud: • Understand shared responsibility model • Encrypt data at rest and in transit • Use cloud provider encryption • Control data residency • Regular security audits • Implement DLP policies Safe Cloud Usage: • Only use approved cloud services • Read terms of service • Understand data ownership • Configure privacy settings • Regular permission reviews • Monitor shared links • Set expiration on shares • Audit access logs Common Cloud Risks: • Misconfigured storage buckets • Excessive permissions • Shadow IT • Account takeover • Data leakage • Compliance violations • Insider threats • API vulnerabilities """, keywords=["cloud", "SaaS", "AWS", "Azure", "Google Cloud", "OneDrive", "Dropbox", "Office 365"], severity="high" ) ] # Add all knowledge items to vector database batch_size = 10 for i in range(0, len(knowledge_items), batch_size): batch = knowledge_items[i:i + batch_size] embeddings = [] documents = [] metadatas = [] ids = [] for item in batch: # Generate embedding embedding = self.embedder.encode(item.content).tolist() embeddings.append(embedding) # Prepare document documents.append(item.content) # Prepare metadata metadatas.append(item.to_dict()) # Generate unique ID doc_id = hashlib.md5( f"{item.topic.value}_{item.title}_{len(item.content)}".encode() ).hexdigest() ids.append(doc_id) # Add batch to collection self.collection.add( embeddings=embeddings, documents=documents, metadatas=metadatas, ids=ids ) logger.info(f"Added batch {i // batch_size + 1} of knowledge items") logger.info(f"Successfully loaded {len(knowledge_items)} knowledge items") def search(self, query: str, k: int = 3, filter_topic: Optional[str] = None, min_severity: Optional[str] = None) -> List[Dict[str, Any]]: """ Search for relevant security information Args: query: User's question k: Number of results to return filter_topic: Optional topic filter min_severity: Minimum severity level filter Returns: List of relevant documents with metadata """ self.stats["queries_processed"] += 1 # Generate query embedding query_embedding = self.embedder.encode(query).tolist() # Build filter where_filter = {} if filter_topic: where_filter["topic"] = filter_topic if min_severity: severity_levels = ["low", "medium", "high", "critical"] min_index = severity_levels.index(min_severity) valid_severities = severity_levels[min_index:] where_filter["severity"] = {"$in": valid_severities} # Search with or without filter if where_filter: results = self.collection.query( query_embeddings=[query_embedding], n_results=k, where=where_filter ) else: results = self.collection.query( query_embeddings=[query_embedding], n_results=k ) # Format results formatted_results = [] if results['documents'] and results['documents'][0]: for doc, metadata, distance in zip( results['documents'][0], results['metadatas'][0], results['distances'][0] ): formatted_results.append({ 'content': doc, 'topic': metadata.get('topic', 'unknown'), 'title': metadata.get('title', 'Untitled'), 'severity': metadata.get('severity', 'medium'), 'keywords': json.loads(metadata.get('keywords', '[]')), # Deserialize JSON string back to list 'relevance_score': 1 - (distance / 2) # Convert distance to similarity }) return formatted_results def add_custom_knowledge(self, content: str, topic: str, title: str, keywords: List[str], severity: str = "medium") -> bool: """ Add custom security knowledge to the database Args: content: Knowledge content topic: Topic category title: Title of the knowledge keywords: Related keywords severity: Severity level Returns: Success status """ try: # Generate embedding embedding = self.embedder.encode(content).tolist() # Generate unique ID doc_id = hashlib.md5( f"{topic}_{title}_{len(content)}_{datetime.now().isoformat()}".encode() ).hexdigest() # Add to collection self.collection.add( embeddings=[embedding], documents=[content], metadatas=[{ "topic": topic, "title": title, "keywords": json.dumps(keywords), # Serialize list to JSON string "severity": severity, "last_updated": datetime.now().isoformat(), "custom": True }], ids=[doc_id] ) self.stats["total_documents"] = self.collection.count() logger.info(f"Added custom knowledge: {title}") return True except Exception as e: logger.error(f"Failed to add custom knowledge: {e}") return False def get_statistics(self) -> Dict[str, Any]: """Get knowledge base statistics""" self.stats["total_documents"] = self.collection.count() self.stats["last_accessed"] = datetime.now().isoformat() # Get topic distribution all_metadata = self.collection.get()['metadatas'] topic_counts = {} severity_counts = {} for metadata in all_metadata: topic = metadata.get('topic', 'unknown') severity = metadata.get('severity', 'unknown') topic_counts[topic] = topic_counts.get(topic, 0) + 1 severity_counts[severity] = severity_counts.get(severity, 0) + 1 self.stats["topic_distribution"] = topic_counts self.stats["severity_distribution"] = severity_counts return self.stats def export_knowledge(self, output_file: str = "knowledge_export.json") -> bool: """Export all knowledge to JSON file""" try: all_data = self.collection.get() export_data = { "exported_at": datetime.now().isoformat(), "total_documents": len(all_data['ids']), "documents": [] } for doc, metadata, doc_id in zip( all_data['documents'], all_data['metadatas'], all_data['ids'] ): export_data["documents"].append({ "id": doc_id, "content": doc, "metadata": metadata }) with open(output_file, 'w') as f: json.dump(export_data, f, indent=2) logger.info(f"Exported knowledge to {output_file}") return True except Exception as e: logger.error(f"Failed to export knowledge: {e}") return False # ================================================ # RAG-Enhanced LLM Class # ================================================ class RAGCybersecurityLLM(CybersecurityLLM): def __init__(self, repo_id: str = "daskalos-apps/phi4-cybersec-Q4_K_M", filename: str = "phi4-mini-instruct-Q4_K_M.gguf", local_dir: str = "./models", knowledge_dir: str = "./knowledge_db", force_download: bool = False): """ Initialize LLM with RAG capabilities Args: repo_id: Hugging Face repository ID filename: Model filename local_dir: Local cache directory knowledge_dir: Knowledge base directory force_download: Force model re-download """ # Initialize base LLM super().__init__(repo_id, filename, local_dir, force_download) # Initialize knowledge base logger.info("Initializing RAG knowledge base...") self.knowledge_base = CybersecurityKnowledgeBase(persist_directory=knowledge_dir) # Enhanced system prompt for RAG self.rag_prompt_template = """<|system|> {system} You have access to a comprehensive cybersecurity knowledge base. Use the provided context to give accurate, detailed answers. If the context doesn't contain relevant information, use your general knowledge but indicate when you're doing so. <|end|> <|user|> Context from knowledge base: {context} User Question: {user} <|end|> <|assistant|>""" def generate_with_rag(self, prompt: str, max_tokens: int = 512, use_rag: bool = True, k_documents: int = 3, min_relevance: float = 0.5) -> Dict[str, Any]: """ Generate response with RAG enhancement Args: prompt: User's question max_tokens: Maximum response length use_rag: Whether to use RAG k_documents: Number of documents to retrieve min_relevance: Minimum relevance threshold Returns: Response with metadata and sources """ context = None sources = [] if use_rag: # Search knowledge base logger.info(f"Searching knowledge base for: {prompt[:50]}...") relevant_docs = self.knowledge_base.search(prompt, k=k_documents) # Filter by relevance relevant_docs = [ doc for doc in relevant_docs if doc.get('relevance_score', 0) >= min_relevance ] if relevant_docs: # Build context from relevant documents context_parts = [] for i, doc in enumerate(relevant_docs, 1): context_parts.append( f"[Source {i}: {doc['title']} - Severity: {doc['severity']}]\n" f"{doc['content'][:1000]}..." # Limit context length ) sources.append({ "title": doc['title'], "topic": doc['topic'], "severity": doc['severity'], "relevance": doc['relevance_score'] }) context = "\n\n".join(context_parts) logger.info(f"Found {len(relevant_docs)} relevant documents") else: logger.info("No highly relevant documents found") # Generate response if context and use_rag: # Use RAG prompt template full_prompt = self.rag_prompt_template.format( system=self.system_prompt, context=context, user=prompt ) else: # Use standard prompt template full_prompt = self.format_prompt(prompt) try: response = self.llm( full_prompt, max_tokens=max_tokens, temperature=0.7, top_p=0.95, top_k=40, repeat_penalty=1.1, stop=self.stop_tokens, echo=False ) text = response['choices'][0]['text'].strip() return { "response": text, "tokens_used": response['usage']['total_tokens'], "model": self.model_info['repo_id'], "sources": sources, "rag_used": use_rag and bool(context) } except Exception as e: logger.error(f"Generation error: {e}") return { "response": "I apologize, but I encountered an error. Please try rephrasing your question.", "error": str(e), "sources": [], "rag_used": False } def generate_stream_with_rag(self, prompt: str, max_tokens: int = 512, use_rag: bool = True, k_documents: int = 3) -> Generator: """Stream response with RAG enhancement""" # Get context if using RAG context = None if use_rag: relevant_docs = self.knowledge_base.search(prompt, k=k_documents) if relevant_docs: context_parts = [f"{doc['title']}: {doc['content'][:500]}" for doc in relevant_docs] context = "\n\n".join(context_parts) # Generate prompt if context: full_prompt = self.rag_prompt_template.format( system=self.system_prompt, context=context, user=prompt ) else: full_prompt = self.format_prompt(prompt) # Stream response stream = self.llm( full_prompt, max_tokens=max_tokens, temperature=0.7, top_p=0.95, top_k=40, repeat_penalty=1.1, stop=self.stop_tokens, echo=False, stream=True ) for output in stream: token = output['choices'][0].get('text', '') if token: yield token def add_knowledge(self, content: str, topic: str, title: str, keywords: List[str]) -> bool: """Add new knowledge to the RAG system""" return self.knowledge_base.add_custom_knowledge( content=content, topic=topic, title=title, keywords=keywords ) def get_knowledge_stats(self) -> Dict[str, Any]: """Get knowledge base statistics""" return self.knowledge_base.get_statistics()