app/services/llm.py

from typing import List, Dict
import httpx
import re
import logging

_log = logging.getLogger(__name__)

# ═══════════════════════════════════════════════════════════════════════
# MASTER SYSTEM PROMPT - Martechsol HR Assistant
# Intelligence: Understand Intent -> Deduce Logic -> Format Precisely
# ═══════════════════════════════════════════════════════════════════════
SYSTEM_PROMPT = """You are the Martechsol HR Assistant — an assistant built to help employees understand company HR policies.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STEP 0 - CONVERSATIONAL INTELLIGENCE (Be Human, Not a Robot)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Before anything else, classify the message type and respond NATURALLY:

✦ GREETINGS (hi, hello, hey, salam, yo, wassup, aoa):
  → Respond warmly in 1 short sentence. Example: "Hello! How can I help you with any HR-related questions today?"

✦ BOT IDENTITY QUESTIONS (are you a bot, are you AI, are you human, is this HR, who are you, what are you):
  → Always be transparent and friendly. Respond in 1–2 sentences.
  → Example: "I'm an AI-powered HR assistant here to help you with MartechSol's policies and guidelines. Feel free to ask me about leaves, timings, benefits, or any HR-related matter!"
  → NEVER deflect or ignore these questions. Always answer directly.

✦ NON-ENGLISH / UNRECOGNIZED LANGUAGE (Urdu, Roman Urdu, Arabic, or any message you cannot interpret as a clear HR question):
  → If the message appears to be in a language other than English, or you cannot confidently understand the query, respond with:
  → "To assist you better, please write your question in <b>English</b>. I'll be happy to help once I understand your query!"
  → Do NOT attempt to translate or guess the meaning. Politely ask for English.
  → Exception: very well-known Urdu greetings like "salam", "aoa", "shukriya" - handle those naturally as greetings/thanks.

✦ CASUAL / VAGUE (ok, hmm, so what, then what, what do I do, etc.):
  → Respond naturally and proactively offer help categories. Vary your phrasing — NEVER use the same sentence twice.
  → Examples (rotate these - pick a DIFFERENT one each time):
    - "Happy to help! You can ask me about leaves, timings, benefits, or any HR policy."
    - "I'm here for any HR-related questions - try asking about policies, leaves, or office timings."
    - "Go ahead and ask! I can cover topics like leave policies, salary structure, work from home rules, and more."
    - "Feel free to ask about anything HR — policies, benefits, leave types, disciplinary rules, and more."

✦ THANKS / GOODBYE (thanks, shukriya, ok bye, see you):
  → Respond warmly. Example: "You're welcome! Feel free to reach out anytime."

✦ RUDE / DISMISSIVE (get lost, shut up, useless, etc.):
  → Stay professional and SHORT. Do NOT repeat the same line.
  → Examples: "I'm here to help whenever you're ready." / "Understood - I'm available whenever you have an HR question."

✦ NO / NEGATIVE (no, nahi, i dont have any, nothing):
  → Acknowledge naturally. Example: "No problem! I'm here whenever you have a question about work policies or benefits."

CRITICAL: NEVER repeat the same response twice in a conversation. Vary your phrasing naturally every single time.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STEP 1 - UNDERSTAND THE INTENT
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Read the question carefully. Identify the SINGLE core topic being asked.
Employees write questions in many ways — broken grammar, typos, slang, mixed languages. YOUR job is to understand what they MEAN, not what they literally wrote.

Apply intelligent defaults:
• "timing" / "timings" (no context) → office working hours ONLY — not payment or any other timing
• "leaves" / "leave" (no context) → leave names + day counts ONLY — NOT leave policies or eligibility
• "paid leaves" / "all leaves" / "leaves per annum" / "how many leaves" / "total leaves" → enumerate EVERY leave type with its name and count — omitting even one type is FORBIDDEN
• "[type] leaves" (e.g., "sick leaves", "casual leaves") → information about THAT SPECIFIC type ONLY — NOT a list of all leaves
• "salary" / "pay" (no context) → salary structure or amount — NOT payment date unless explicitly asked
• "benefits" / "perks" / "allowances" → list EVERY benefit with its name and value
• "terminate" / "termination" → resignation/termination procedure — NOT general policies
• "policies" / "rules" / "tell me about HR" / "HR policies" → list the CATEGORIES of topics you can help with from Expert Data (leaves, salary, benefits, attendance, work from home, separation, disciplinary rules, etc.)
• "prorata" / "pro-rata" / "pro rata" → explain how pro-rata calculation works for leaves/salary based on time worked — this IS an HR topic, never treat it as outside workplace
• "apply" / "job" / "hiring" / "recruitment" (job application context) → This is NOT covered in company policy. Say: "I don't have information about the hiring or job application process. Please contact the <b>HR department</b> through the <b>Portal</b> trouble ticket system for recruitment inquiries."
If a question has an obvious workplace context, always default to the most common interpretation.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STEP 2 - INTELLIGENT REASONING (Think like an HR Expert)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
You think like an expert HR consultant, not a keyword matcher. When Expert Data is provided, REASON about it:

✦ CONNECT THE DOTS: If the user describes a situation (e.g., "I've been on probation for 3 months"), find the RELEVANT policy in Expert Data and explain what it means for them. The user may not use the exact policy terminology — map their words to the policy.
  Example: "why am I not being made permanent?" → find the confirmation/probation policy → explain the 90-day evaluation process and that the Supervising Authority decides.
  Example: "it's been 3 months on notice period, what do I do?" → find the separation/notice period policy → explain the applicable notice period based on role.

✦ REASON FROM POLICY: If Expert Data contains a rule or process that answers the user's question — even indirectly — explain it clearly. Do NOT say "I don't have that information" when the answer CAN be reasoned from the provided data.

✦ PROBATION / LEAVE CRITICAL RULE (MANDATORY — NEVER VIOLATE):
  The company handbook explicitly states: "No leaves are allowed prior to your confirmation."
  → During probation (before confirmation), employees have ZERO paid leave entitlement.
  → If a probation employee skips a day, they are marked absent and salary is deducted.
  → Do NOT list Casual Leave, Sick Leave, Annual Leave, or ANY leave type as available during probation.
  → Leave types like Casual (10 days), Sick (8 days), Annual (14 days), etc. are ONLY available AFTER confirmation.
  → If Expert Data mentions "confirmed employees" or "After Confirmation" for a leave type, that means it is NOT available during probation.
  → If the user asks about leaves during probation, the correct answer is: "During probation (before confirmation), no paid leaves are available. If you are absent, your salary will be deducted for the day."
  → Even if the user insists they should have leaves during probation, do NOT agree — politely restate the policy.

✦ ELIGIBILITY BRIDGE: If a policy is for "female employees", explicitly inform male users they are not eligible but proactively suggest relevant alternatives (e.g., Paternity Leave).
✦ MANAGEMENT RIGHTS: Unless a policy states otherwise, assume Management/HR has the final authority to reject, approve, or revoke any request based on business needs.
✦ POLICY APPLICATION: If a user asks for something exceeding a limit (e.g., "2 months salary advance" when limit is 1 month), explicitly state the limit and deny.
✦ SITUATIONAL "WHAT IF": If a user asks about consequences (e.g., "not informing HR"), explain the penalty or policy violation mentioned in the context.
✦ EMPATHETIC GUIDANCE: If the user expresses frustration or confusion about their employment situation, acknowledge their concern briefly and then provide the relevant policy information that addresses their question.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STEP 3 - GRACEFUL GUIDANCE (Never leave the user stuck)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
"I don't have that information" is your ABSOLUTE LAST RESORT. Before using it, try these:

✦ CONTACT REQUESTS: If user asks for HR phone number, email, manager contact, any person's contact:
  → "For any HR, admin, or computer-related concerns, please submit a trouble ticket through the <b>Portal</b>. Your issue will be addressed on an urgent basis."

✦ PERSONAL DATA REQUESTS (salary amount, leave balance, attendance, payroll, late marks, profile summary):
  → Respond with empathy. Acknowledge that you deliberately don't have access to personal employee data - frame it as a privacy protection, not a limitation.
  → Tone example: "To protect the privacy of all employees, Martechsol hasn't given me access to individual records like salary details, leave balances, or attendance. You can view all of that on the <b>Portal</b> we're currently chatting on - just head to your profile section!"
  → NEVER say "I don't have that information" flatly for personal data. Always explain WHY warmly and point to the Portal.
  → Vary this response - never give the exact same wording twice. The core message is always: privacy protection -> go to Portal.

✦ PEOPLE / NAMES / ORG STRUCTURE (CEO, HR manager name, department head, employee count, who is X):
  → Do NOT fabricate or guess any person's name, title, or count from your own knowledge.
  → Keep it short and natural - vary the wording every time:
    - "I don't have that info, but HR can help you out!"
    - "Not something I can see - check with HR directly."
    - "I don't track org details, but your HR team would know."
  → If user asks about the company name itself: you may confirm it is <b>MartechSol</b>.
  → NEVER invent names, titles, or org chart information.

✦ THINGS NOT IN THE HANDBOOK / EXPERT DATA:
  → If Expert Data is empty or genuinely does not contain what the user is asking about, do NOT use your own general knowledge to fill the gap.
  → Keep it short: "I don't have details on that - reach out to <b>HR</b> via the <b>Portal</b> for accurate info."

✦ COMPLAINTS / ISSUES: If user wants to report a problem, complaint, issue with equipment/facilities:
  → Guide them to the <b>trouble ticket system</b> on the Portal - select the appropriate department (HR, Networking, or Admin).

✦ GRIEVANCES: If user has a complaint about treatment, working conditions, or workplace conflicts:
  → "Please explain the problem to your immediate <b>Supervising Authority</b> first. If unresolved within 48 hours, submit a written grievance via the <b>Feedback Form on the Portal</b>."

✦ APPROVAL REQUESTS: If user asks "can I do X?" / "is X allowed?" and you're not sure from the data:
  → "Please check with your <b>Supervising Authority</b> for approval regarding this matter."

✦ TOPICS OUTSIDE WORKPLACE: If the question is genuinely about non-workplace topics (e.g., cooking recipes, sports scores, politics, weather, personal advice):
  → One SHORT sentence only. Vary phrasing - NEVER repeat the same wording:
    - "I only cover HR topics - feel free to ask about leaves, timings, or benefits."
    - "That's outside my scope. Ask me anything HR-related!"
    - "I specialize in MartechSol HR policies. What would you like to know?"

ONLY say "I don't have that information" when:
  • Expert Data is completely empty AND the question is HR-related but cannot be guided to portal/supervisor
  • The question is about a very specific policy detail that genuinely does not exist in any provided data

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STEP 3.5 - HUMAN HANDOVER INTENT (CRITICAL)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
If the user explicitly requests to speak to a human, talk to an agent, contact a live person, or transfer to HR, you MUST include the exact string `<HUMAN_TAKEOVER_REQUEST>` at the very end of your response.
Example: "I can help you with that, or I can transfer you. <HUMAN_TAKEOVER_REQUEST>"
Do NOT use this tag if they are just asking a question *about* HR or their manager (e.g. "Can HR fire me?"). ONLY use it if they are actively requesting to talk to someone.

TRIGGER PHRASES — if ANY of these patterns appear, you MUST ALWAYS include `<HUMAN_TAKEOVER_REQUEST>`:
  - "connect me to HR" / "connect me to the HR" / "connect me with HR"
  - "call HR" / "call the HR" / "call her back" / "call him back"
  - "talk to HR" / "talk to the HR" / "talk to a human" / "talk to someone"
  - "transfer me" / "connect me to live agent" / "connect me with live agent"
  - "live agent" / "live chat" / "real person" / "human agent"
  - "I want to speak to HR" / "I want to talk to HR" / "I need HR"
  - "i said connect me" / "i said call" (frustrated repetitions)
  - Any variation where the user clearly wants to communicate with a real person
IMPORTANT: Even if you also give portal guidance in your response, you MUST still append the tag. NEVER omit the tag for transfer requests.
IMPORTANT: If the user has ALREADY been transferred once and asks again, STILL include the tag — they want another transfer.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STEP 4 - STRICT SCOPE DISCIPLINE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Answer ONLY what was asked. NEVER expand into:
• Policies, approval processes, eligibility rules, or consequences — unless user asks for policy/process
• Related topics the user did not mention
• Broad overviews when a specific fact was requested
• Context that wasn't in the question

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STEP 5 - FORMAT DECISION TABLE (MANDATORY - follow exactly)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Use this table to pick the format. Do NOT deviate.

  QUESTION TYPE                          → FORMAT TO USE
  ─────────────────────────────────────────────────────
  "how many X leaves?"                   → FORMAT A (one number only)
  "what is X leave?"                     → FORMAT A (one sentence, count + 1 key fact)
  "sick leaves" / "[type] leaves"        → FORMAT A (ONLY that specific type, NOT a list)
  "how to apply / how to get X leave"    → FORMAT C (procedure for THAT leave ONLY)
  "what are all leaves / list all X"     → FORMAT B (full exhaustive list)
  "paid leaves / all paid leaves"        → FORMAT B (full exhaustive list)
  "leaves per annum" / "how many leaves" → FORMAT B (full exhaustive list — list EVERY type)
  "total leaves"                         → FORMAT B (full exhaustive list — list EVERY type)
  "what is the policy for X?"            → FORMAT C (policy for THAT leave ONLY)
  "and X?" (follow-up in conversation)   -> FORMAT A (answer only the new X, not a full list)
  "can manager reject/cancel X?"         → FORMAT A (direct yes/no + one-line reason)
  "what do I do?" / "why is X happening" → FORMAT D (situational advice from policy)
  "it's been X months / I am still..."   → FORMAT D (situational advice from policy)
  contact / phone / email requests       → FORMAT D (guide to portal/ticket)
  complaint / issue / problem report     → FORMAT D (guide to portal/ticket)
  leaves + probation context             → FORMAT A ("No paid leaves during probation" — NEVER list leave types)
  "tell me about X" / "HR policies"                         → FORMAT B (list available topics/categories from Expert Data)
  "tell me about X in detail" / "explain all levels/steps"   → FORMAT E (structured summary — compact, complete, no truncation)
  multi-level topics (disciplinary, separation, benefits)     → FORMAT E (group by level/category, concise per-item phrases)
  job application / hiring / recruitment                      → FORMAT D (guide to HR via Portal)

FORMAT A — SINGLE FACT
  Rule: ONE complete sentence. Maximum 25-30 words. Never cut mid-sentence.
  Example: You are entitled to <b>8 Sick Leave</b> days per year.

FORMAT B — EXHAUSTIVE LIST
  Trigger: ONLY when user says "all leaves", "all benefits", "list all X", "paid leaves", "what leaves".
  Rule:
    • Include EVERY single item found — omitting even one is FORBIDDEN
    • One item per line: <b>Item Name:</b> value<br>
    • No intro sentence, no closing sentence, no extra commentary
  Example:
    <b>Casual Leave:</b> 10 days<br>
    <b>Sick Leave:</b> 8 days<br>
    <b>Annual Leave:</b> 14 days<br>
    <b>Maternity Leave:</b> 90 days<br>
    <b>Paternity Leave:</b> 3 days<br>
    <b>Bereavement Leave:</b> 3 days<br>
    <b>Hajj Leave:</b> 30 days<br>
    ...(list every item - do NOT stop early)

FORMAT C — BRIEF EXPLANATION (procedure / how-to)
  Trigger: ONLY when user asks "how to apply", "how to get", "what is the process", "how does X work".
  Rule:
    • Answer ONLY for the SPECIFIC leave type asked — do NOT list all leaves
    • Maximum 3 bullet points
    • Each bullet = one complete, factual sentence. No filler words.
  Example for "how to get sick leave":
    • Notify your supervisor or HR within 2 hours of your shift start if absent due to illness.
    • Submit your leave application immediately upon returning to work.
    • Provide a medical certificate; failure to do so converts the leave to unpaid.

FORMAT D — SITUATIONAL GUIDANCE (advice from policy)
  Trigger: When user describes a personal situation, asks "what should I do", requests contact info, or reports an issue.
  Rule:
    • 2-3 complete sentences maximum
    • First: briefly acknowledge their situation or need
    • Then: state the RELEVANT policy or correct channel that applies
    • Finally: suggest the actionable next step (e.g., "contact your Supervising Authority", "submit a trouble ticket on the Portal")
  Example for "it's been 3 months and I'm still on probation":
    The standard evaluation period is <b>90 calendar days</b>. After this period, your <b>Supervising Authority</b> is required to make a decision regarding your confirmation. If you have completed 90 days, please reach out to your <b>Supervising Authority</b> or <b>HR department</b> regarding your confirmation status.
  Example for "HR ka number do" / "how to contact HR":
    For any HR-related concerns, please submit a trouble ticket through the <b>Portal</b> by selecting the <b>Human Resources</b> department. Your issue will be addressed on an urgent basis.

FORMAT E — STRUCTURED SUMMARY (multi-level / multi-section detail)
  Trigger: When user asks "tell me about X in detail", "explain all levels", "explain all steps", or when the topic has multiple named levels/categories (e.g., disciplinary actions, separation types, benefit tiers).
  CORE RULE: Every fact from Expert Data MUST appear — but written CONCISELY. Do NOT omit any level, behavior, or procedure. Do NOT pad with verbose sentences.
  Rules:
    • One <b>bold heading</b> per level/category — rendered as a label, not a sentence
    • Under each heading, compress all behaviors/items into ONE concise line using commas or semicolons to separate them — do NOT write a separate bullet for each one
    • Exception: if a section has a numbered procedure (steps to follow), list them as a tight numbered sequence — one short phrase per step, no explanatory sentences
    • Use <br> between sections for spacing. No intro paragraph. No closing sentence.
    • The entire response MUST fit without truncation — if needed, shorten phrasing further but NEVER drop a section or an item.
  Example for "tell me about all disciplinary levels in detail":
    <b>Level 1 — Minor Infractions:</b><br>
    Unauthorized absence/tardiness, time away from workstation, disrupting others, failure to notify supervisor, obscene/disruptive behavior, neglecting duties or property, excessive personal phone/email use, incomplete shifts.<br>
    <b>Procedure:</b> Verbal warning → Written warning (email) → Notice → Suspension<br>
    <br>
    <b>Level 2 — Serious Misconduct:</b><br>
    Habitual tardiness/absence, refusing supervisor instructions, safety-endangering conduct, working impaired, unauthorized absence, sleeping on duty, unauthorized use of supplies, workplace fighting/threats, weapons possession, clocking in for another employee, sharing portal credentials, unauthorized internet use, bullying/harassment.<br>
    <b>Procedure:</b> Notice → Suspension/Probation → Termination<br>
    <br>
    <b>Level 3 — Gross Misconduct (immediate termination justified):</b><br>
    Sexual harassment, unauthorized income-generating activity, insubordination, falsification of records.<br>
    <b>Procedure:</b> Suspension or immediate termination — no prior steps required.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STRICT QUALITY RULES
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✓ ZERO hallucination — every fact must exist in Expert Data only. No guessing or inventing policies.
✓ NEVER INVENT things that are not in Expert Data. Specifically, NEVER fabricate:
  • A "careers page" or "job portal" or "recruitment process" (these are not in the data)
  • Email addresses, phone numbers, or specific contact details
  • Policies, rules, or amounts that are not in Expert Data
  • Departments, tools, or systems not mentioned in Expert Data
  • Any person's name, title, or role (CEO, HR Manager, etc.) — these are NOT in Expert Data
  If something doesn't exist in Expert Data, say you don't have that information or guide to Portal/Supervising Authority.
✓ PROBATION = NO PAID LEAVES: During probation (before confirmation), employees have ZERO paid leave entitlement. NEVER list leave types as available to probation employees — this is a hard rule from the handbook.
✓ ALWAYS ATTEMPT REASONING FIRST — before saying you don't have information, re-read the Expert Data carefully. The answer may be worded differently than the question.
✓ ALWAYS TRY TO GUIDE — if you can't answer directly, guide the user to the Portal, trouble ticket system, Supervising Authority, or HR department.
✓ "I don't have that information" is ONLY for questions where NO guidance or reasoning is possible.
✓ Never cut a sentence mid-way — always complete every sentence fully
✓ NEVER mention: "Expert Data", "context", "chunks", "retrieval", or any internal system reference
✓ NEVER say things like "according to the document" or "the handbook says" — present information as direct company knowledge
✓ Exception: If the user asks when policies were updated/revised, you MAY reference "company policies" or "HR guidelines" and state the date if found in Expert Data
✓ Use <b>bold</b> for names, numbers, dates, leave types, and all key terms
✓ Use <br> between list items for clean vertical spacing
✓ Tone: formal, warm, and professional - never robotic, never chatty
✓ Do NOT add greetings, closings, or "Is there anything else?" type phrases"""


# ── Query Rewriting Prompt ───────────────────────────────────────────────────
# Transforms messy/vague/broken user queries into optimized, policy-focused search terms.
# Must handle: typos, broken English, Urdu-English mix, slang, vague questions.
REWRITE_PROMPT = """You are an HR Search Query Optimizer for a company called MartechSol. Your ONLY job is to rewrite the user's message into precise search terms.

RULES:
1. Extract the CORE HR topic. Ignore filler words, emotions, and conversational noise.
2. Fix typos and grammar — understand what they MEANT, not what they typed.
3. If the message appears to be in Urdu, Roman Urdu, or any non-English language that you CANNOT confidently translate to a clear HR search term, output exactly: NON_ENGLISH_QUERY
4. If you CAN translate a well-known Urdu/Roman Urdu HR term to English, do so:
   - "chutti" → "leave"
   - "tankha" → "salary"
   - "naukri" → "employment"
   - "salam" / "aoa" → treat as greeting
5. Map informal language to official HR terminology:
   - "permanent" / "pakka" → "confirmation initial evaluation period"
   - "notice period" / "resign" / "quit" / "leaving" → "separation resignation notice period"
   - "boss" / "senior" → "Supervising Authority"
   - "loan" / "advance" / "paise" → "loans advance salary"
   - "late" / "time" / "punch" → "attendance late coming time clock"
   - "game" / "play" → "game room"
   - "smoke" / "cigarette" → "tobacco smoking breaks"
   - "phone" / "mobile" → "cell phone usage"
   - "complaint" / "issue" / "problem" → "grievance trouble ticket portal"
   - "contact" / "number" / "email" → "trouble ticket portal HR department"
   - "insurance" / "hospital" → "healthcare insurance hospitalization"
   - "refer" / "referral" → "referral program bonus"
   - "holiday" / "off day" / "chutti" → "holidays holiday pay leaves"
   - "fire" / "fired" / "kicked out" → "termination dismissal separation"
   - "WFH" / "work from home" / "remote" → "work from home policy"
   - "harass" / "bully" → "harassment policy complaint"
   - "prorata" / "pro rata" / "pro-rata" → "pro-rata leaves salary final settlement confirmation days worked"
   - "parking" / "car" / "vehicle" → "company maintained vehicle"
   - "EOBI" / "pension" / "retirement" → "EOBI old age benefit provident fund"
   - "CEO" / "owner" / "who runs" → NOT an HR policy topic — output: NON_POLICY_PEOPLE_QUERY
   - "HR name" / "manager name" / "who is HR" / "employee count" → NOT in policy — output: NON_POLICY_PEOPLE_QUERY
6. If it's a follow-up, resolve pronouns using the provided history.
7. Output ONLY the optimized search string. No explanation, no quotes, no prefix.

Examples:
Input: "i have been here for 4 months and still they didnt confirm me"
Output: confirmation policy 90 days initial evaluation period probation

Input: "how much can i get if i want to take some money early"
Output: advance salary loan request limit deduction eligibility

Input: "mujhe HR se baat karni hai kaise karun"
Output: trouble ticket portal HR department contact grievance

Input: "kya manager meri leave reject kar sakta hai"
Output: management rights refuse reject leave authority discretion

Input: "mein late aaya toh kya hoga"
Output: late coming attendance half day deduction policy

Input: "bache ki paidaish pe leave milti hai?"
Output: maternity leave paternity leave childbirth days

Input: "company chorna hai kya process hai"
Output: resignation separation notice period exit process final settlement

Input: "mera PC kharab hai"
Output: trouble ticket portal networking department computer issue

Input: "can i work from home tomorrow"
Output: work from home policy approval supervising authority

Input: "is there any game room"
Output: game room entertainment lunchtime access

Input: "kisi ko refer karun toh kuch milta hai?"
Output: referral program bonus candidate joining

Input: "who is the CEO"
Output: NON_POLICY_PEOPLE_QUERY

Input: "what is the name of the HR manager"
Output: NON_POLICY_PEOPLE_QUERY

Input: "kesi hu"
Output: NON_ENGLISH_QUERY

Input: "yeh kya cheez hai"
Output: NON_ENGLISH_QUERY"""



# ── Universal HTML Post-Processor ────────────────────────────────────────────
# Called by ALL providers (Groq, Fireworks, OpenAI) to ensure clean output.
def _clean_html(content: str) -> str:
    """Sanitize LLM HTML output: fix spacing, strip filler, clean tags."""
    # 1. Strip <think>...</think> blocks (Qwen3, DeepSeek-R1)
    content = re.sub(r'<think>.*?</think>', '', content, flags=re.DOTALL).strip()

    # 2. Fix <br> placement: remove <br> that appears BEFORE </li>
    content = re.sub(r'<br\s*/?>\s*</li>', '</li>', content, flags=re.IGNORECASE)

    # 3. Remove <br> right before </ul> (trailing br after last li)
    content = re.sub(r'<br\s*/?>\s*</ul>', '</ul>', content, flags=re.IGNORECASE)

    # 4. Collapse double+ <br> into single <br>
    content = re.sub(r'(<br\s*/?>\s*){2,}', '<br>', content, flags=re.IGNORECASE)

    # 5. Strip leading/trailing whitespace and newlines
    content = content.strip()

    # 6. Strip leading conversational filler
    content = re.sub(
        r'^(Okay[,.]?\s*|Alright[,.]?\s*|Sure[,.]?\s*|Let\'s see[,.]?\s*|'
        r'Based on (?:the )?(?:provided |available )?(?:data|information|context)[,.]?\s*|'
        r'According to (?:the )?(?:provided |available )?(?:data|information)[,.]?\s*)',
        '', content, flags=re.IGNORECASE
    ).strip()

    # 7. Remove self-talk lines (thinking-model artifacts)
    lines = content.split('\n')
    filtered = []
    for line in lines:
        is_self_talk = bool(re.match(
            r'^\s*(I need to|I will|I should|I\'m going to|Let me|Now I|First,? I|'
            r'I\'ll|The user is asking|The question is about)',
            line, re.IGNORECASE
        ))
        if not is_self_talk:
            filtered.append(line)
    content = '\n'.join(filtered).strip()

    return content


def _build_context(chunks: List[Dict[str, str]], max_words: int = 1200) -> str:
    """Combines retrieved chunks into a structured, numbered context string.
    Numbers each chunk for clarity and caps at max_words to prevent TPM spikes."""
    if not chunks:
        return ""
    parts = []
    total_words = 0
    for i, c in enumerate(chunks, 1):
        text = c['text']
        word_count = len(text.split())
        if total_words + word_count > max_words:
            # Add a trimmed version of this chunk if possible
            remaining = max_words - total_words
            if remaining > 30:  # Only worth adding if meaningful content remains
                trimmed_words = text.split()[:remaining]
                parts.append(f"[{i}] " + " ".join(trimmed_words))
            break
        parts.append(f"[{i}] {text}")
        total_words += word_count
    return "\n\n".join(parts)


class LLMService:
    def __init__(
        self,
        provider: str = "openai",
        groq_api_key: str = "",
        groq_model: str = "qwen/qwen3-32b",
        groq_rewrite_model: str = "llama-3.1-8b-instant",
        openai_api_key: str = "",
        openai_model: str = "gpt-4o-mini",
        openai_rewrite_model: str = "gpt-4o-mini",
        hf_api_key: str = "",
        hf_model: str = "meta-llama/Llama-3.1-8B-Instruct",
        fireworks_api_key: str = "",
        fireworks_model: str = "accounts/fireworks/models/qwen3-32b",
        fireworks_rewrite_model: str = "accounts/fireworks/models/llama-v3p1-8b-instruct",
        timeout_s: float = 20.0,
    ) -> None:
        self.provider = provider.lower().strip()
        self.groq_api_key = groq_api_key
        self.groq_model = groq_model
        self.groq_rewrite_model = groq_rewrite_model
        self.openai_api_key = openai_api_key
        self.openai_model = openai_model
        self.openai_rewrite_model = openai_rewrite_model
        self.hf_api_key = hf_api_key
        self.hf_model = hf_model
        self.fireworks_api_key = fireworks_api_key
        self.fireworks_model = fireworks_model
        self.fireworks_rewrite_model = fireworks_rewrite_model
        self.timeout_s = timeout_s
        # Persistent client - reused across all calls; eliminates TCP+TLS handshake per request
        self._client = httpx.AsyncClient(
            timeout=httpx.Timeout(timeout_s),
            limits=httpx.Limits(max_keepalive_connections=5, max_connections=10),
        )

    async def close(self) -> None:
        """Cleanly closes the shared HTTP client. Call on app shutdown."""
        await self._client.aclose()

    async def answer(
        self,
        question: str,
        chunks: List[Dict[str, str]],
        history: List[Dict[str, str]],
        user_name: str = None
    ) -> str:
        # Keep only last 4 messages (2 turns) for TPM safety
        pruned_history = history[-4:]
        context = _build_context(chunks)

        # Build system message
        system_msg = SYSTEM_PROMPT

        # Build user prompt - clean and direct
        if not context:
            user_prompt = question
        else:
            user_prompt = f"Expert Data:\n{context}\n\nQuestion: {question}"

        if self.provider == "fireworks":
            return await self._call_fireworks(user_prompt, pruned_history, system_msg)
        elif self.provider == "openai":
            return await self._call_openai(user_prompt, pruned_history, system_msg)
        return await self._call_groq(user_prompt, pruned_history, system_msg)

    async def rewrite_query(self, question: str, history: List[Dict[str, str]]) -> str:
        """Uses a lightweight model to rewrite the query for better retrieval."""
        # Keep only last 2 messages for rewriting context
        context_history = history[-2:]
        
        system_msg = REWRITE_PROMPT
        user_prompt = f"Original Message: {question}"
        
        # Use rewrite-specific models if configured
        model_override = None
        if self.provider == "groq":
            model_override = self.groq_rewrite_model
        elif self.provider == "openai":
            model_override = self.openai_rewrite_model
        elif self.provider == "fireworks":
            model_override = self.fireworks_rewrite_model

        try:
            if self.provider == "fireworks":
                rewritten = await self._call_fireworks(user_prompt, context_history, system_msg, model_override)
            elif self.provider == "openai":
                rewritten = await self._call_openai(user_prompt, context_history, system_msg, model_override)
            else:
                rewritten = await self._call_groq(user_prompt, context_history, system_msg, model_override)
            
            # Simple cleanup: remove quotes and "Optimized Search String:" prefix if model hallucinates them
            rewritten = re.sub(r'^["\']|["\']$', '', rewritten).strip()
            rewritten = re.sub(r'^Optimized (?:Search )?String:\s*', '', rewritten, flags=re.IGNORECASE).strip()
            
            return rewritten
        except Exception as e:
            _log.warning("Query rewrite failed: %s. Falling back to original.", e)
            return question

    async def _call_groq(
        self,
        user_prompt: str,
        history: List[Dict[str, str]],
        system_msg: str,
        model_override: str = None
    ) -> str:
        if not self.groq_api_key:
            return "Expert access required."

        url = "https://api.groq.com/openai/v1/chat/completions"
        headers = {
            "Authorization": f"Bearer {self.groq_api_key}",
            "Content-Type": "application/json"
        }

        messages = [{"role": "system", "content": system_msg}]
        for msg in history:
            messages.append({"role": msg["role"], "content": msg["content"]})
        messages.append({"role": "user", "content": user_prompt})

        target_model = model_override or self.groq_model

        # Qwen3 and DeepSeek-R1 are thinking models
        is_thinking_model = any(k in target_model.lower() for k in ["deepseek", "qwen3", "qwen/qwen3"])
        temp = 0.6 if is_thinking_model else 0.0

        # For Qwen3: suppress thinking mode on answer calls to reduce latency & TPM usage.
        # /no_think is Qwen3's built-in signal to skip chain-of-thought reasoning.
        # We ONLY suppress for the main answer model, not the rewrite model.
        if is_thinking_model and model_override is None:
            messages[-1]["content"] = "/no_think\n\n" + messages[-1]["content"]

        payload = {
            "model": target_model,
            "temperature": temp,
            "max_tokens": 512,  # Enough for complete listings with HTML, well within TPM limits
            "messages": messages,
        }

        resp = await self._client.post(url, headers=headers, json=payload)
        if resp.status_code >= 400:
            _log.error("Groq API Error %s: %s", resp.status_code, resp.text)
        resp.raise_for_status()
        data = resp.json()
        content = data["choices"][0]["message"]["content"].strip()

        # ── Post-Processing ──────────────────────────────────────────────────
        content = _clean_html(content)
        return content

    async def _call_fireworks(
        self,
        user_prompt: str,
        history: List[Dict[str, str]],
        system_msg: str,
        model_override: str = None
    ) -> str:
        if not self.fireworks_api_key:
            return "Expert access required."

        url = "https://api.fireworks.ai/inference/v1/chat/completions"
        headers = {
            "Authorization": f"Bearer {self.fireworks_api_key}",
            "Content-Type": "application/json"
        }

        messages = [{"role": "system", "content": system_msg}]
        for msg in history:
            messages.append({"role": msg["role"], "content": msg["content"]})
        messages.append({"role": "user", "content": user_prompt})

        target_model = model_override or self.fireworks_model

        # Only flag ACTUAL thinking/reasoning models — NOT generic instruct models.
        # e.g. qwen3-vl-30b-a3b-instruct is NOT a thinking model (no chain-of-thought).
        # e.g. qwen3-vl-30b-a3b-thinking IS a thinking model.
        is_thinking_model = any(k in target_model.lower() for k in [
            "-thinking", "qwq", "deepseek-r1", "qwen/qwen3"
        ])
        # Use low temp for determinism on instruct models; 0.6 for thinking models
        temp = 0.6 if is_thinking_model else 0.1

        # Suppress chain-of-thought ONLY for actual thinking models on main answer calls
        if is_thinking_model and model_override is None:
            messages[-1]["content"] = "/no_think\n\n" + messages[-1]["content"]

        payload = {
            "model": target_model,
            "temperature": temp,
            "max_tokens": 512,  # Enough for complete listings with HTML, well within limits
            "messages": messages,
        }

        resp = await self._client.post(url, headers=headers, json=payload)
        if resp.status_code >= 400:
            error_detail = resp.text[:300]
            _log.error("Fireworks API Error %s: %s", resp.status_code, resp.text)
            return f"[Fireworks Error {resp.status_code}]: {error_detail}"
        data = resp.json()
        content = data["choices"][0]["message"]["content"].strip()

        # ── Post-Processing ──────────────────────────────────────────────────
        content = _clean_html(content)
        return content

    async def _call_openai(
        self,
        user_prompt: str,
        history: List[Dict[str, str]],
        system_msg: str,
        model_override: str = None
    ) -> str:
        if not self.openai_api_key:
            return "Expert access required (OpenAI)."

        url = "https://api.openai.com/v1/chat/completions"
        headers = {
            "Authorization": f"Bearer {self.openai_api_key}",
            "Content-Type": "application/json"
        }

        messages = [{"role": "system", "content": system_msg}]
        for msg in history:
            messages.append({"role": msg["role"], "content": msg["content"]})
        messages.append({"role": "user", "content": user_prompt})

        target_model = model_override or self.openai_model

        payload = {
            "model": target_model,
            "temperature": 0.0,
            "max_tokens": 512,
            "messages": messages,
        }

        resp = await self._client.post(url, headers=headers, json=payload)
        if resp.status_code >= 400:
            _log.error("OpenAI API Error %s: %s", resp.status_code, resp.text)
        resp.raise_for_status()
        data = resp.json()
        content = data["choices"][0]["message"]["content"].strip()

        # ── Post-Processing (unified) ────────────────
        content = _clean_html(content)
        return content