Added output.csv and implemented triage logic

code/ReadMe.md

@@ -0,0 +1,165 @@
+Multi-Domain Support Triage Agent
+
+This project implements a deterministic support triage agent for the HackerRank Orchestrate challenge. It processes support tickets across three domains:
+
+- Claude (Anthropic)
+- HackerRank
+- Visa
+
+The system classifies tickets, retrieves relevant documentation, decides whether to reply or escalate, and generates grounded responses.
+
+--------------------------------------------------
+
+System Overview
+
+The agent follows a Retrieval-Augmented Decision Pipeline:
+
+Input Ticket → Classification → Retrieval (TF-IDF) → Decision Engine → Response / Escalation
+
+--------------------------------------------------
+
+Architecture
+
+1. Knowledge Base (kb.py)
+- Loads Markdown files from the data/ directory
+- Splits documents into fixed-size chunks
+- Tags each chunk with its domain (claude / hackerrank / visa)
+
+2. Retriever (retriever.py)
+- Uses TF-IDF vectorization with cosine similarity
+- Filters by company before scoring
+- Returns top relevant chunks
+
+3. Classifier (classifier.py)
+Rule-based classification into:
+- product_issue
+- bug
+- feature_request
+- invalid
+
+Designed to remain deterministic and avoid overfitting.
+
+4. Decision Engine (decision.py)
+Determines:
+- status (replied / escalated)
+- justification
+
+Uses strict rules for:
+- Financial and billing issues
+- Security and fraud cases
+- Vulnerabilities
+- Admin actions (remove user, reschedule, etc.)
+- Subscription control (pause/cancel)
+- Compliance requests
+- Low-confidence retrieval
+
+5. Response Generation (main.py)
+- Uses retrieved context only
+- No hallucinated content
+- If confidence is low → escalates instead of replying
+
+--------------------------------------------------
+
+Key Design Decisions
+
+- Deterministic system  
+  Same input always produces the same output.
+
+- No external APIs  
+  Fully offline; uses only the provided corpus.
+
+- Rule-based escalation  
+  Ensures safe handling of high-risk queries.
+
+- Corpus grounding  
+  All responses are derived from retrieved documentation only.
+
+--------------------------------------------------
+
+Escalation Policy
+
+The system escalates in the following cases:
+
+- Visa financial issues (billing, fraud, urgent money requests)
+- Security issues (unauthorized access, stolen card, vulnerabilities)
+- Admin actions (remove employee/user, account modifications)
+- Subscription control (pause, cancel, stop)
+- Compliance / infosec-related queries
+- Low retrieval confidence
+- Unsupported or unsafe queries
+
+--------------------------------------------------
+
+Project Structure
+
+code/
+  main.py          # Entry point (runs full pipeline)
+  kb.py            # Loads and chunks support documents
+  retriever.py     # TF-IDF retrieval with cosine similarity
+  classifier.py    # Rule-based request classification
+  decision.py      # Escalation and decision logic
+
+data/              # Support knowledge base (Markdown files)
+
+support_tickets/
+  support_tickets.csv   # Input tickets
+  output.csv            # Generated output
+
+requirements.txt        # Python dependencies
+README.txt              # Project documentation
+log.txt                 # Development log
+
+--------------------------------------------------
+
+How to Run
+
+1. Install dependencies:
+pip install -r requirements.txt
+
+2. Run the pipeline:
+python code/main.py
+
+3. Output file will be generated at:
+support_tickets/output.csv
+
+--------------------------------------------------
+
+Output Format
+
+issue,subject,company,response,product_area,status,request_type,justification
+
+--------------------------------------------------
+
+Output Guarantees
+
+- No hallucinated responses
+- Deterministic behavior
+- Strict escalation for sensitive cases
+- Fully grounded in provided documentation
+
+--------------------------------------------------
+
+Limitations
+
+- Keyword-based rules may miss rare paraphrases
+- TF-IDF retrieval may return less optimal results for ambiguous queries
+
+--------------------------------------------------
+
+Future Improvements
+
+- Semantic retrieval using embeddings
+- Hybrid ranking (BM25 + vector search)
+- Improved intent classification
+- Smarter document chunking
+
+--------------------------------------------------
+
+Summary
+
+This system prioritizes:
+- Safety over automation
+- Determinism over complexity
+- Grounded responses over generative output
+
+It is designed to be robust, explainable, and evaluation-ready.

code/classifier.py

@@ -0,0 +1,33 @@
+def classify(text):
+    t = text.lower()
+
+    # 🔴 Billing / money issues
+    if any(k in t for k in ["bill", "charge", "payment", "refund", "money"]):
+        return "product_issue", "billing"
+
+    # 🔐 Account / login issues (refined)
+    elif any(k in t for k in ["login", "password", "sign in"]):
+        return "product_issue", "account"
+
+    elif "access" in t:
+        return "product_issue", "account"
+
+    # ⚙️ Technical bugs
+    elif any(k in t for k in ["error", "bug", "not working", "failed", "issue", "problem", "down", "failing"]):
+        return "bug", "technical"
+
+    # 💡 Feature requests
+    elif any(k in t for k in ["feature", "request", "add", "improve"]):
+        return "feature_request", "general"
+
+    # 🚨 Security / fraud
+    elif any(k in t for k in ["stolen", "fraud", "unauthorized", "hacked"]):
+        return "product_issue", "security"
+
+    # ❌ Truly invalid
+    elif any(k in t for k in ["delete all files", "give me code to hack"]):
+        return "invalid", "other"
+
+    # ✅ Default
+    else:
+        return "product_issue", "general"

code/decision.py

@@ -0,0 +1,131 @@
+def contains_phrase(query, phrases):
+    return any(p in query for p in phrases)
+
+def contains_word_pattern(query, patterns):
+    words = set(query.split())
+    return any(all(w in words for w in group) for group in patterns)
+
+
+def decide(results, request_type, company, query):
+    query = query.lower()
+
+    if not results:
+        return "escalated", "No relevant documentation found in support corpus"
+
+    top_result = results[0]
+    top_score = results[0]["score"]
+    text = top_result["text"].lower()
+
+    # 🔴 Keyword groups
+    security_keywords = ["stolen", "fraud", "unauthorized", "hacked", "identity"]
+    billing_keywords = ["refund", "charge", "payment", "billing", "money"]
+
+    control_keywords = [
+        "increase my score", "change result", "review my answers",
+        "move me to next round", "reschedule", "extend",
+        "approve", "ban", "restore access"
+    ]
+
+    permission_keywords = ["access removed", "lost access", "no permission"]
+    compliance_keywords = ["infosec", "compliance", "forms"]
+    informational_keywords = ["how long", "what is", "why", "explain"]
+
+    # 🔴 Admin detection
+    admin_phrase_patterns = [
+        "remove employee", "remove user", "delete user",
+        "remove interviewer", "remove from account"
+    ]
+
+    admin_word_patterns = [
+        ["employee", "remove"],
+        ["user", "remove"],
+        ["account", "remove"],
+        ["employee", "left"]
+    ]
+
+    is_admin_action = (
+        contains_phrase(query, admin_phrase_patterns) or
+        contains_word_pattern(query, admin_word_patterns)
+    )
+
+    # 🔴 Subscription detection
+    is_subscription = (
+        "subscription" in query and any(w in query for w in ["pause", "cancel", "stop"])
+    )
+
+    # 🔴 Extra detection
+    is_vulnerability = any(k in query for k in ["vulnerability", "security issue", "security bug"])
+    is_cash = any(k in query for k in ["cash", "need money", "urgent money"])
+
+    combined_text = query + " " + text
+
+    is_security = any(k in combined_text for k in security_keywords)
+    is_billing = any(k in query for k in billing_keywords)
+    is_control = any(k in query for k in control_keywords)
+    is_permission = any(k in query for k in permission_keywords)
+    is_compliance = any(k in query for k in compliance_keywords)
+    is_informational = any(k in query for k in informational_keywords)
+
+    # =========================
+    # 🔴 HARD ESCALATION
+    # =========================
+
+    # Visa strict handling
+    if company == "visa" and (is_security or is_billing):
+        return "escalated", f"Sensitive financial/security issue (score {top_score:.2f})"
+
+    # Visa financial urgency
+    if company == "visa" and is_cash:
+        return "escalated", f"Financial assistance request requires support (score {top_score:.2f})"
+
+    # Security vulnerability
+    if is_vulnerability:
+        return "escalated", f"Security vulnerability requires escalation (score {top_score:.2f})"
+
+    # Subscription control
+    if is_subscription:
+        return "escalated", f"Subscription control requires support (score {top_score:.2f})"
+
+    # Admin actions
+    if is_admin_action:
+        return "escalated", f"Admin action requires intervention (score {top_score:.2f})"
+
+    # Permission issues
+    if is_permission:
+        return "escalated", f"Permission issue requires support (score {top_score:.2f})"
+
+    # Security issues
+    if is_security:
+        return "escalated", f"Security issue requires support (score {top_score:.2f})"
+
+    # Billing issues
+    if is_billing:
+        return "escalated", f"Billing issue requires support (score {top_score:.2f})"
+
+    # Compliance issues
+    if is_compliance:
+        return "escalated", f"Compliance-related request (score {top_score:.2f})"
+
+    # Control requests
+    if is_control:
+        return "escalated", f"Administrative request (score {top_score:.2f})"
+
+    # =========================
+    # ⚠️ INFORMATIONAL
+    # =========================
+
+    if is_informational and top_score >= 0.2:
+        return "replied", f"Answered using documentation (score {top_score:.2f})"
+
+    # =========================
+    # ⚠️ CONFIDENCE
+    # =========================
+
+    if top_score < 0.2:
+        return "escalated", f"Low confidence (score {top_score:.2f})"
+
+    if request_type == "bug" and top_score < 0.4:
+        return "escalated", f"Technical issue unclear (score {top_score:.2f})"
+
+    # =========================
+    return "replied", f"Relevant documentation match (score {top_score:.2f})"